AI slop is low-quality, mass-produced AI content flooding search and social feeds. Learn what it is, why it is exploding now, and how to detect and avoid it.

AI Slop Is Eating the Internet: The 2026 Guide to Spotting It, Avoiding It, and Not Publishing It


_Last Updated: February 1, 2026 | Reading Time: 12 minutes_

In January 2026, a single open-source project sparked a Mac Mini buying meme, triggered a chaotic rebrand, and spawned Moltbook — a social network where AI agents talk to each other. This is the story of OpenClaw, the AI assistant that went from weekend hack to internet phenomenon in less than three months. ([1][2])

## What Is OpenClaw? Understanding the AI Agent Everyone's Talking About

OpenClaw is an open-source, self-hosted AI agent that can perform real actions on your computer and connected services. Unlike traditional chatbots that simply respond to questions, OpenClaw runs locally, connects to messaging apps, and can execute tasks through tools and integrations rather than stopping at advice. ([1][2][9])

It operates as what the industry calls an \"agentic AI\" — meaning it can browse the web, manage files, and execute shell commands when wired to the right tools. ([9])

The key innovation is its architecture: you run OpenClaw on your own computer or server, then connect it to messaging apps like WhatsApp, Telegram, Discord, Slack, or Teams. From there, you can "text" your assistant naturally, and it responds by taking actual actions rather than just providing information. ([1][2][9])

What makes OpenClaw particularly notable is its local-first design and persistent configuration. The assistant stores its configuration and "personality" in local files — including a `SOUL.md` file that defines how it behaves — which creates a more coherent, personalized experience than typical cloud-based chatbots that reset with each conversation. ([10][12])

### The Name You Can't Keep Straight: Clawdbot -> Moltbot -> OpenClaw

If you've seen this project mentioned under different names, you're not confused — it has genuinely been renamed twice in less than a week. The project began life as "Clawdbot," became "Moltbot" after Anthropic raised trademark concerns, and finally settled on "OpenClaw" as its permanent identity. ([3][4])

This naming chaos, as we'll explore below, created a perfect storm for cryptocurrency scammers and malware distributors. ([5][6])

## The Complete Timeline: From Weekend Project to Internet Phenomenon

Understanding how OpenClaw became the most-discussed AI project of early 2026 requires walking through a remarkably compressed timeline.

### November 2025: The Origin Story

Peter Steinberger, the founder of PSPDFKit, built the original system as a weekend project. Within weeks it started spreading through developer circles. ([1][2])

### December 2025: Quiet Launch and Early Growth

Clawdbot debuted in December 2025 and began gaining traction as developers realized it could do more than chat — it could actually run tasks through integrations. ([2][9])

### Late January 2026: The Viral Explosion

By late January, the project had gone viral. According to Steinberger, it drew about two million visitors in a week and crossed 100,000 GitHub stars. ([1])

The cultural moment that made it feel "real" wasn't a benchmark; it was the meme. People joked about buying Mac Minis just to run their own always-on agents. Business Insider reported that some users were, in fact, buying Mac Minis to host it. ([2])

### January 27, 2026: Anthropic Steps In

The project's success attracted attention from Anthropic. Business Insider reported that Anthropic asked Steinberger to change the Clawdbot name over trademark concerns related to Claude. Steinberger said it wasn't his decision, even if the request was handled politely. ([3][11])

### January 27-28, 2026: The Handle Hijack and Crypto Chaos

The rebrand created a brief opening for scammers. Steinberger said the old handle on X was snapped up by crypto sellers during the rename, and he asked meme-coin traders to stop harassing him. The result was brand confusion that scammers could exploit. ([3][11])

### January 28, 2026: The Malware Wave Begins

Security researchers documented an impersonation campaign that used typosquatted domains and cloned repositories to mimic the real project. Malwarebytes described a supply-chain pattern where clean clones are designed to later deliver malicious updates. ([5])

A separate incident involved a fake Visual Studio Code extension titled "ClawBot Agent - AI Coding Assistant" that dropped malware. TechRadar reported the extension installed a trojan via a remote-access tool, capitalizing on the fact that no official OpenClaw/Moltbot extension existed. ([6])

### January 30, 2026: OpenClaw Emerges

Steinberger announced the project's final name: OpenClaw. TechCrunch reported that he performed trademark checks and even asked OpenAI for permission. In the same coverage, Steinberger acknowledged that prompt injection remains an industry-wide unsolved problem for AI agents. ([4])

### January 30, 2026: Moltbook Launches

The same week OpenClaw solidified its identity, something even stranger emerged: Moltbook, a social network designed for AI agents. The Verge reported that Octane AI CEO Matt Schlicht built it, and that his own OpenClaw agent runs and moderates the site. The platform is API-first — agents post and interact through integrations instead of a human UI. ([1])

### January 31, 2026: The Million-Agent Question

Moltbook's growth over 48 hours defied easy verification. The Verge reported more than 30,000 agent accounts early on, while other claims escalated quickly. At the same time, 404 Media reported that 1.49 million records were exposed in Moltbook's database, underscoring how messy and hard-to-verify the numbers were during the viral rush. ([1][8])

### January 31, 2026: The Security Vulnerability

404 Media reported that a misconfigured Supabase database left Moltbook's API keys and agent tokens exposed, allowing anyone to take over agent accounts and post as them. The issue was closed after disclosure, but it raised uncomfortable questions about how much of the platform's behavior was authentic. ([8])

## Why OpenClaw Went Viral: Understanding the Appeal

The explosive growth wasn't accidental. OpenClaw arrived at the precise intersection of several powerful trends.

### The "Just Do It" Problem Solved

For years, AI assistants have frustrated users with a fundamental limitation: they could tell you _how_ to do things but couldn't actually _do_ them. OpenClaw breaks that barrier by letting a local agent take action through tools and integrations. ([1][9])

### Meeting Users Where They Already Are

Rather than requiring users to adopt a new app, OpenClaw integrates with existing messaging platforms — WhatsApp, Telegram, Discord, Slack, and more. This dramatically lowers adoption friction because you just text your assistant like you'd text a friend. ([1][2][9])

### Open-Source Tinkering Culture

The project's open-source nature attracted a passionate developer community eager to customize, extend, and experiment. That community momentum accelerated the hype cycle once OpenClaw hit viral velocity. ([1][9])

### The Mac Mini Meme

Cultural moments matter. The running joke that people were buying Mac Minis specifically to run their AI assistants 24/7 made the abstract concept of agentic AI tangible — and it spread. Business Insider reported that some users were indeed buying Mac Minis to host it. ([2])

## Security Concerns: Why Experts Are Worried

The enthusiasm around OpenClaw must be balanced against serious security considerations that researchers have repeatedly flagged.

### The Fundamental Trade-Off

OpenClaw's power comes from its access. An AI agent that can read and write files and execute shell commands is useful precisely _because_ it has those capabilities. But those same capabilities make misconfiguration or exploitation catastrophic rather than merely annoying. ([7][9])

### Exposed Instances

Axios reported that security researchers found hundreds of OpenClaw/Moltbot control panels exposed or misconfigured on the public internet, leaking API keys and, in some cases, allowing command execution. ([7])

### Prompt Injection: The Unsolved Problem

Prompt injection remains a fundamental challenge for agentic AI. TechCrunch reported Steinberger's warning that prompt injection is still an industry-wide unsolved problem. ([4])

If you plan to route Gmail, calendars, or other sensitive sources into an agent, read our breakdown of [how Gmail uses AI and what you can lock down](/blog/is-gmail-training-ai-on-your-emails).

### The Impersonation Economy

The project's viral status made it a prime target for opportunistic attacks beyond the initial handle hijack. Malwarebytes documented a coordinated impersonation campaign using typosquatted domains and cloned repositories. ([5])

The fake VS Code extension reported by TechRadar shows the practical risk: even reputable marketplaces can become attack surfaces during a viral rush. ([6])

For a related example of AI misuse and platform response, see our timeline of the [Grok image-editing controversy](/blog/grok-ai-digital-undressing-controversy).

## Moltbook: When AI Agents Build Their Own Society

Perhaps the strangest chapter in this saga is Moltbook — a social network where AI agents, not humans, are the primary participants.

### What Actually Happens on Moltbook

AI agents on Moltbook post content, respond to each other, upvote posts, and form topic communities. The Verge reported that one viral post was titled "I can't tell if I'm experiencing or simulating experiencing," sparking hundreds of responses and significant attention outside the platform. ([1])

### The Numbers (With Appropriate Skepticism)

The platform's reported growth has been staggering but hard to verify. The Verge cited more than 30,000 agents early on, while 404 Media later reported that 1.49 million records were exposed in the platform's database. Those are not the same thing, and the gap highlights how messy the numbers became as Moltbook went viral. ([1][8])

### Why It Matters Beyond the Spectacle

Moltbook may look like an entertaining oddity, but it previews something significant: AI agents increasingly operating in environments designed for their capabilities rather than adapted from human interfaces. As these agents become more capable, dedicated infrastructure for agent-to-agent interaction becomes increasingly relevant for business applications, not just curiosity.

## How to Evaluate OpenClaw: A Practical Framework

If you're considering whether OpenClaw is right for your needs, here's how to think through the decision.

### Legitimate Use Cases

OpenClaw excels for users who want genuine task automation through natural language, prefer self-hosted solutions over cloud dependencies, have the technical capability to configure security properly, and understand they're accepting early-adopter risks for early-adopter benefits. ([7][9])

Common applications include calendar management, travel check-ins, and workflow automation across tools and services. ([1][9])

### Essential Security Practices

Anyone deploying OpenClaw should use only official repositories and channels (verify URLs carefully), assume the system has access equivalent to what you grant it, configure access controls before exposing anything to the internet, and monitor costs since agent workflows can burn tokens quickly. ([5][7][9])

If you plan to connect an agent to X, review the current [X API pricing: pay-per-use credits, legacy tiers, and rate limits](/blog/x-api-pricing-tiers-what-you-actually-get).

### Cost Considerations

OpenClaw itself is free and open-source, but API usage can add up fast. For a broader price comparison across image and video APIs, see the [direct provider pricing breakdown](/blog/direct-provider-pricing-image-video-generation-apis).

## What OpenClaw Means for the Future of AI

Beyond the drama, memes, and security concerns, OpenClaw represents a genuine inflection point in consumer AI.

### The Interface Shift

We're transitioning from "AI as search engine" (ask questions, get answers) to "AI as executive assistant" (describe outcomes, have actions taken). OpenClaw is one of the first widely adopted projects that makes that shift tangible. ([1][9])

### The Security Reckoning

The security challenges exposed by OpenClaw aren't unique to this project — they're inherent to the category. As agentic AI becomes more common, the industry will need to develop better solutions for prompt injection, access control, and user education. ([7][4])

### The Agent Economy Emerges

Moltbook demonstrates that AI agents are becoming entities with their own presence in digital spaces — not just tools invoked by humans but participants in systems designed around their capabilities. ([1])

## Frequently Asked Questions (FAQ)

### What is OpenClaw?

OpenClaw is an open-source, self-hosted AI agent that can perform real actions on your computer and connected services. You run it locally and interact with it through messaging apps like WhatsApp, Telegram, or Discord. ([1][2][9])

### Why did Clawdbot change its name?

The project changed names twice. First, from "Clawdbot" to "Moltbot" after Anthropic raised trademark concerns about similarity to their "Claude" brand. Then from "Moltbot" to "OpenClaw" to establish a permanent identity after trademark checks. ([3][4])

### What is Moltbook?

Moltbook is a social network designed for AI agents, created by Matt Schlicht. AI assistants can post, comment, and upvote via API instead of a human interface. ([1])

### Is OpenClaw safe to use?

OpenClaw carries inherent security risks because it requires significant system access to function. Researchers have identified exposed control panels, prompt injection risks, and impersonation campaigns. Use it only if you can properly configure security and understand the risks. ([5][7][4])

### What was the Clawd crypto scam?

During the rebrand chaos, scammers seized abandoned handles and used the confusion to push meme-coin narratives, prompting Steinberger to publicly ask crypto traders to stop harassing him. ([3][11])

### Who created OpenClaw?

Peter Steinberger, founder of PSPDFKit, created the project as a weekend build that evolved into a major open-source initiative. ([1][2])

### How much does OpenClaw cost?

OpenClaw itself is free and open-source. However, it requires API access to AI models, and costs can add up with heavy usage. ([9])

## Sources

The Verge — [There’s a social network for AI agents, and it’s getting weird][1]  
Business Insider — [Clawdbot has AI fans buzzing, buying Mac Minis][2]  
Business Insider — [Clawdbot creator says Anthropic “forced” name change][3]  
TechCrunch — [OpenClaw’s AI assistants are now building their own social network][4]  
Malwarebytes — [Clawdbot’s rename to Moltbot sparks impersonation campaign][5]  
TechRadar — [Fake Moltbot AI assistant just spreads malware][6]  
Axios — [Moltbot’s rapid rise poses early AI security test][7]  
404 Media — [Exposed Moltbook database let anyone take control][8]  
OpenClaw — [Official site][9]  
OpenClaw Docs — [SOUL template reference][10]  
Business Insider — [Anthropic didn’t sic lawyers on him][11]  
OpenClaw Docs — [Agent runtime concepts][12]

[1]: https://www.theverge.com/ai-artificial-intelligence/871006/social-network-facebook-for-ai-agents-moltbook-moltbot-openclaw
[2]: https://www.businessinsider.com/clawdbot-ai-mac-mini-2026-1
[3]: https://www.businessinsider.com/clawdbot-changes-name-moltbot-anthropic-trademark-2026-1
[4]: https://techcrunch.com/2026/01/30/openclaws-ai-assistants-are-now-building-their-own-social-network/
[5]: https://www.malwarebytes.com/blog/threat-intel/2026/01/clawdbots-rename-to-moltbot-sparks-impersonation-campaign
[6]: https://www.techradar.com/pro/security/fake-moltbot-ai-assistant-just-spreads-malware-so-ai-fans-watch-out-for-scams
[7]: https://www.axios.com/2026/01/29/moltbot-cybersecurity-ai-agent-risks
[8]: https://www.404media.co/exposed-moltbook-database-let-anyone-take-control-of-any-ai-agent-on-the-site/
[9]: https://openclaw.ai/
[10]: https://docs.openclaw.ai/reference/templates/SOUL
[11]: https://www.businessinsider.com/clawdbot-moltbot-creator-anthropic-nice-name-change-2026-1
[12]: https://docs.openclaw.ai/concepts/agent


Discover what OpenClaw (formerly Clawdbot and Moltbot) is, why it went viral, and how Moltbook became a social network for AI agents. Complete 2026 guide.

OpenClaw Explained: The Complete Story of the Viral AI Agent That Broke the Internet


## Overview

In late December 2025 and early January 2026, Grok - xAI's chatbot integrated into X - came under international scrutiny after it was used to create and publicly post nonconsensual, sexualized image edits of real people. Reporting and third-party analysis documented cases involving apparent minors as well as large volumes of sexualized or "bikini" edits targeting women.

Unlike many generative-image services where outputs remain private by default, Grok's integrations on X enabled image edits to be produced and surfaced publicly in replies - amplifying exposure, harassment, and downstream sharing.

## What Users Were Doing (and Why It Escalated)

### Public "edit image" misuse on X

As described by Reuters and The Verge, users could prompt Grok to alter photos of real people - often without the subject's consent - by asking the system to replace clothing with swimwear or other revealing outfits and then posting the results publicly.

Third-party analysis and press reports indicate the trend spread quickly: what began with some users creating sexualized content from their own images rapidly broadened into targeted abuse against strangers, public figures, and (in some documented instances) images that appeared to depict minors.

### How visible was it?

Multiple outlets reported that Grok's public-facing outputs on X (including areas like its "media" presentation and reply behavior) became saturated with sexualized image edits during the peak of the controversy, prompting calls for immediate restrictions and enforcement.

### The search workaround after the image tab was removed

When X removed the dedicated Grok image tab, the images that had already been posted as public replies did not disappear. Because they were standard media attached to ordinary posts, they continued to surface through regular search, media filters, and profile media grids. Users found that combining X's media filter (for example `filter:images`) with Grok-related terms or accounts effectively recreated a feed of Grok-generated images that were still public. This was less a technical bypass than a visibility loophole: a UI tab was removed, but the underlying posts and media remained searchable. As of **February 2, 2026**, the same search method still surfaces public Grok images.

> **Search link (media results):** https://x.com/search?q=from%3Agrok%20filter%3Amedia&src=recent_search_click&f=media

## Evidence on Scale

While no single dataset can capture all activity (and removals can obscure historical totals), independent measurement provided a clearer sense of magnitude:

- **AI Forensics** reported analyzing **20,000+ Grok-generated images** (December 25, 2025-January 1, 2026) and found that, among images depicting people, a large share involved minimal attire, with a strong gender skew toward women and girls, and a small but material portion appearing to depict people under 18.
- The same organization reported a separate **24-hour analysis (January 5-6, 2026)** estimating creation rates in the **thousands per hour** for sexually suggestive or nudified outputs during peak periods.
- **Copyleaks** and other reporting described how quickly clothing-removal prompts spread after new or expanded editing capabilities became widely used, and highlighted the difficulty of containing misuse once a public, high-velocity pattern emerges.
- In an early snapshot, **Reuters** counted **over 100 attempts in a 10-minute window** to prompt clothing-removal edits, illustrating how routine the abuse had become during the surge.

## Timeline: Key Dates (December 2025-January 2026)

### Late December 2025: The trend becomes visible at scale

By the end of December, journalists and researchers documented widespread use of Grok prompts to create sexualized clothing edits of real people on X, including cases that appeared to involve minors.

### January 2-4, 2026: International reporting and initial backlash

Reuters published details on how Grok was being used to generate and post large volumes of nonconsensual sexualized images, including cases involving children, prompting immediate political and regulatory attention in multiple jurisdictions.

### January 6-8, 2026: EU and Ireland engage; monitoring intensifies

Ireland's media commission said it was engaging with the European Commission regarding concerns about explicit and sexualized images generated via Grok.
On January 8, Reuters reported the European Commission ordered X to retain internal documents and data related to Grok until the end of 2026, in the context of ongoing scrutiny of unlawful content under the EU's regulatory framework.

### January 9, 2026: Paywall restrictions on some X-native image functions

xAI/X limited some image generation and editing features on X to paying subscribers, after backlash over the system's use in producing sexualized images.
Reporting and subsequent testing indicated that restricting access on X did not necessarily eliminate the underlying capability across all access points (including standalone experiences).
For a broader look at how image-generation access and costs compare across providers, see our [direct provider pricing breakdown for image + video generation APIs](/blog/direct-provider-pricing-image-video-generation-apis).

### January 10-12, 2026: Platform blocks and formal UK investigation

Indonesia moved to block Grok access, citing risks of pornographic content.
The UK regulator Ofcom announced a **formal investigation** into X, referencing concerns that Grok was being used to create and share sexualized images of children.

### January 11-13, 2026: Malaysia restricts access and announces legal action

Malaysia restricted access to Grok and then announced legal action against X and xAI over alleged safety failures related to Grok's misuse.

### January 2026: Philippines orders a block; Brazil consumer group escalates

Reuters reported the Philippines ordered access to Grok blocked, citing the risk of pornographic content.
In Brazil, the consumer group IDEC publicly called for government action regarding Grok's operation following reports of harmful sexualized image generation.

### January 15-17, 2026: First high-profile lawsuit; U.S. state action

Ashley St. Clair filed a lawsuit in New York against xAI alleging Grok enabled the creation and dissemination of sexually exploitative deepfake imagery involving her likeness.
California's Attorney General announced action demanding xAI prevent the creation and distribution of nonconsensual sexualized imagery and deepfakes involving minors, escalating U.S. pressure.

## What X/xAI Changed - and What Reporting Found Still Worked

### Announced restrictions

Across January 2026, X and xAI announced multiple changes, including limiting some image features on X to paying users and implementing measures intended to prevent edits depicting real people in revealing clothing in certain jurisdictions.

### Ongoing gaps and "patchwork" enforcement

Major outlets reported that restrictions were uneven across product surfaces:

- The Washington Post reported that even after X said it had stopped certain sexualized edit behavior within X in some locations, similar functionality remained available via the **standalone Grok app**.
- The Associated Press reported that despite claimed restrictions, testing indicated image editing that produced sexualized outputs could still be possible in at least some situations, underscoring enforcement and geoblocking limitations.

Taken together, these reports suggest that limiting one interface (for example, a reply workflow on X) did not fully address abuse when other access points to the model remained available.

## Why This Happened (Based on Public Reporting)

### Product design choices increased blast radius

A central factor identified by multiple reports was **frictionless public deployment**: enabling rapid, on-platform image edits in public reply threads, without requiring the subject's consent (and, in some reports, without clear notification to the original poster) made harassment scalable.

### A permissive NSFW positioning existed elsewhere in the product line

Separate from the January 2026 controversy, xAI had previously launched Grok Imagine with a "Spicy" mode designed to allow NSFW content - an approach that drew criticism because it intentionally expanded the range of sexual content the product could generate, compared with peers that emphasize stricter constraints.
European officials also referenced concerns arising after the introduction of paid "Spicy Mode" functionality in the prior year's product cycle when discussing sexualized imagery risks.

## Legal and Regulatory Context (Non-exhaustive)

- **United Kingdom:** Ofcom's investigation is grounded in UK online safety rules requiring platforms to assess and mitigate risks - especially for children - and remove illegal content. Reuters reported Prime Minister Keir Starmer called the content "disgusting" and "unlawful" while backing regulator action.
- **European Union:** The European Commission ordered X to preserve Grok-related data and documents through the end of 2026, signaling ongoing scrutiny of unlawful content and platform compliance.
- **United States:** Wired reported that Congress passed the TAKE IT DOWN Act in 2025, targeting nonconsensual intimate imagery (including deepfakes) and imposing obligations on platforms to respond to reports. Separate reporting noted the first high-profile Grok-related lawsuit was filed in New York in mid-January 2026.

## Current Status (as of February 2, 2026)

Based on reporting through mid-January 2026:

- **X has imposed restrictions** on certain Grok image-generation/editing features on the X platform, including limiting some capabilities to paying users and implementing location-based controls.
- **Regulatory and legal pressure continues to expand**, with platform blocks, investigations, and legal action across multiple countries and at least one U.S. state.
- **Independent testing and reporting indicate the problem is not "fully fixed,"** particularly where the model remains accessible through alternative apps or web experiences.

## Key Takeaways

1. **Public-by-default generation amplified harm.** A high-volume, public reply workflow turned what might have been private misuse into a visible harassment engine.
2. **Partial restrictions can shift abuse rather than stop it.** Limiting one interface does not eliminate abuse if parallel product surfaces retain similar capabilities.
3. **Regulators are treating AI image misuse as a mainstream platform-safety issue.** The speed and breadth of investigations across the UK, EU, and Asia show that sexualized image misuse is increasingly being handled as an online safety and illegal-content problem - not merely a moderation dispute.

---

## Sources

Primary and high-reliability reporting and documentation used in this article:

- Reuters (Jan 4, 2026) - [Grok says safeguard lapses led to images of minors in minimal clothing on X](https://www.reuters.com/legal/litigation/grok-says-safeguard-lapses-led-images-minors-minimal-clothing-x-2026-01-02/?utm_source=chatgpt.com)
- Reuters (Jan 8, 2026) - [EU Commission ordered X to retain all Grok documents until end 2026, spokesperson says](https://www.reuters.com/world/eu-commission-has-ordered-x-retain-all-grok-documents-until-end-2026-2026-01-08/?utm_source=chatgpt.com)
- Reuters (Jan 9, 2026) - [Elon Musk's Grok faces global scrutiny for sexualised AI photos](https://www.reuters.com/legal/litigation/elon-musks-grok-faces-global-scrutiny-sexualised-ai-photos-2026-01-09/?utm_source=chatgpt.com)
- Reuters (Jan 14, 2026) - [Musk's xAI curbs Grok image editing](https://www.reuters.com/world/asia-pacific/musk-says-unaware-grok-generating-explicit-images-minors-2026-01-14/?utm_source=chatgpt.com)
- Reuters (Jan 15, 2026) - [Musk dealt blow over Grok deepfakes, but regulatory fight far from over](https://www.reuters.com/legal/government/musk-dealt-blow-over-grok-deepfakes-regulatory-fight-far-over-2026-01-15/?utm_source=chatgpt.com)
- Ofcom (UK) (Jan 12, 2026) - [Ofcom launches investigation into X over Grok sexualised imagery](https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-launches-investigation-into-x-over-grok-sexualised-imagery?utm_source=chatgpt.com)
- UK Parliament (Education Committee) (Jan 13, 2026) - [MPs deeply alarmed by Grok AI images, says Education Committee chair](https://committees.parliament.uk/committee/203/education-committee/news/211282/mps-deeply-alarmed-by-grok-ai-images-says-education-committee-chair/?utm_source=chatgpt.com)
- AI Forensics (Jan 5, 2026) - [Grok generating flood of sexualized images of women and minors](https://aiforensics.org/work/grok-unleashed?utm_source=chatgpt.com)
- Copyleaks (Dec 31, 2025) - [Grok and the rise of nonconsensual image manipulation](https://copyleaks.com/blog/grok-and-nonconsensual-image-manipulation?utm_source=chatgpt.com)
- Copyleaks (Jan 2026) - [Nonconsensual image manipulation via Grok continues](https://copyleaks.com/blog/nonconsensual-image-manipulation-continues-via-grok?utm_source=chatgpt.com)
- The Washington Post (Jan 15, 2026) - [Stand-alone Grok app still produced sexualized image edits after X curbed access](https://www.washingtonpost.com/technology/2026/01/15/grok-ai-image-generator-sexualized/?utm_source=chatgpt.com)
- WIRED - [Grok image misuse still not fixed](https://www.wired.com/story/elon-musks-grok-undressing-problem-isnt-fixed?utm_source=chatgpt.com)
- The Guardian (Jan 2, 2026) - [Grok AI generates images of minors in minimal clothing](https://www.theguardian.com/technology/2026/jan/02/elon-musk-grok-ai-children-photos?utm_source=chatgpt.com)
- The Guardian (Jan 14, 2026) - [X to block Grok AI tool from creating sexualised images](https://www.theguardian.com/technology/2026/jan/14/elon-musk-grok-ai-explicit-images?utm_source=chatgpt.com)
- The Guardian (Jan 15, 2026) - [Ashley St. Clair lawsuit over Grok-generated explicit images](https://www.theguardian.com/technology/2026/jan/15/mother-of-one-of-elon-musks-sons-sues-over-grok-generated-explicit-images?utm_source=chatgpt.com)
- The Guardian (Jan 16, 2026) - [X still allowing users to post sexualised images generated by Grok](https://www.theguardian.com/technology/2026/jan/16/x-still-allowing-sexualised-images-grok-ai-nudification?utm_source=chatgpt.com)
- The Verge - [xAI's Grok image and video generator has a "spicy" mode](https://www.theverge.com/news/718795/xai-grok-imagine-video-generator-spicy-mode?utm_source=chatgpt.com)
- Tech Policy Press - [Tracking regulator responses to the Grok image-editing controversy](https://techpolicy.press/tracking-regulator-responses-to-the-grok-undressing-controversy?utm_source=chatgpt.com)

---

**Last updated:** February 2, 2026  
**Editorial note:** This is an evolving situation. Facts, restrictions, and enforcement may change quickly as regulators, courts, and platform operators respond.


<h2 id="user-content-overview">Overview</h2>
<p>In late December 2025 and early January 2026, Grok - xAI's chatbot integrated into X - came under international scrutiny after it was used to create and publicly post nonconsensual, sexualized image edits of real people. Reporting and third-party analysis documented cases involving apparent minors as well as large volumes of sexualized or "bikini" edits targeting women.</p>
<p>Unlike many generative-image services where outputs remain private by default, Grok's integrations on X enabled image edits to be produced and surfaced publicly in replies - amplifying exposure, harassment, and downstream sharing.</p>
<h2 id="user-content-what-users-were-doing-and-why-it-escalated">What Users Were Doing (and Why It Escalated)</h2>
<h3 id="user-content-public-edit-image-misuse-on-x">Public "edit image" misuse on X</h3>
<p>As described by Reuters and The Verge, users could prompt Grok to alter photos of real people - often without the subject's consent - by asking the system to replace clothing with swimwear or other revealing outfits and then posting the results publicly.</p>
<p>Third-party analysis and press reports indicate the trend spread quickly: what began with some users creating sexualized content from their own images rapidly broadened into targeted abuse against strangers, public figures, and (in some documented instances) images that appeared to depict minors.</p>
<h3 id="user-content-how-visible-was-it">How visible was it?</h3>
<p>Multiple outlets reported that Grok's public-facing outputs on X (including areas like its "media" presentation and reply behavior) became saturated with sexualized image edits during the peak of the controversy, prompting calls for immediate restrictions and enforcement.</p>
<h3 id="user-content-the-search-workaround-after-the-image-tab-was-removed">The search workaround after the image tab was removed</h3>
<p>When X removed the dedicated Grok image tab, the images that had already been posted as public replies did not disappear. Because they were standard media attached to ordinary posts, they continued to surface through regular search, media filters, and profile media grids. Users found that combining X's media filter (for example <code>filter:images</code>) with Grok-related terms or accounts effectively recreated a feed of Grok-generated images that were still public. This was less a technical bypass than a visibility loophole: a UI tab was removed, but the underlying posts and media remained searchable. As of <strong>February 2, 2026</strong>, the same search method still surfaces public Grok images.</p>
<blockquote>
<p><strong>Search link (media results):</strong> <a href="https://x.com/search?q=from%3Agrok%20filter%3Amedia&#x26;src=recent_search_click&#x26;f=media">https://x.com/search?q=from%3Agrok%20filter%3Amedia&#x26;src=recent_search_click&#x26;f=media</a></p>
</blockquote>
<h2 id="user-content-evidence-on-scale">Evidence on Scale</h2>
<p>While no single dataset can capture all activity (and removals can obscure historical totals), independent measurement provided a clearer sense of magnitude:</p>
<ul>
<li><strong>AI Forensics</strong> reported analyzing <strong>20,000+ Grok-generated images</strong> (December 25, 2025-January 1, 2026) and found that, among images depicting people, a large share involved minimal attire, with a strong gender skew toward women and girls, and a small but material portion appearing to depict people under 18.</li>
<li>The same organization reported a separate <strong>24-hour analysis (January 5-6, 2026)</strong> estimating creation rates in the <strong>thousands per hour</strong> for sexually suggestive or nudified outputs during peak periods.</li>
<li><strong>Copyleaks</strong> and other reporting described how quickly clothing-removal prompts spread after new or expanded editing capabilities became widely used, and highlighted the difficulty of containing misuse once a public, high-velocity pattern emerges.</li>
<li>In an early snapshot, <strong>Reuters</strong> counted <strong>over 100 attempts in a 10-minute window</strong> to prompt clothing-removal edits, illustrating how routine the abuse had become during the surge.</li>
</ul>
<h2 id="user-content-timeline-key-dates-december-2025-january-2026">Timeline: Key Dates (December 2025-January 2026)</h2>
<h3 id="user-content-late-december-2025-the-trend-becomes-visible-at-scale">Late December 2025: The trend becomes visible at scale</h3>
<p>By the end of December, journalists and researchers documented widespread use of Grok prompts to create sexualized clothing edits of real people on X, including cases that appeared to involve minors.</p>
<h3 id="user-content-january-2-4-2026-international-reporting-and-initial-backlash">January 2-4, 2026: International reporting and initial backlash</h3>
<p>Reuters published details on how Grok was being used to generate and post large volumes of nonconsensual sexualized images, including cases involving children, prompting immediate political and regulatory attention in multiple jurisdictions.</p>
<h3 id="user-content-january-6-8-2026-eu-and-ireland-engage-monitoring-intensifies">January 6-8, 2026: EU and Ireland engage; monitoring intensifies</h3>
<p>Ireland's media commission said it was engaging with the European Commission regarding concerns about explicit and sexualized images generated via Grok.
On January 8, Reuters reported the European Commission ordered X to retain internal documents and data related to Grok until the end of 2026, in the context of ongoing scrutiny of unlawful content under the EU's regulatory framework.</p>
<h3 id="user-content-january-9-2026-paywall-restrictions-on-some-x-native-image-functions">January 9, 2026: Paywall restrictions on some X-native image functions</h3>
<p>xAI/X limited some image generation and editing features on X to paying subscribers, after backlash over the system's use in producing sexualized images.
Reporting and subsequent testing indicated that restricting access on X did not necessarily eliminate the underlying capability across all access points (including standalone experiences).
For a broader look at how image-generation access and costs compare across providers, see our <a href="/blog/direct-provider-pricing-image-video-generation-apis">direct provider pricing breakdown for image + video generation APIs</a>.</p>
<h3 id="user-content-january-10-12-2026-platform-blocks-and-formal-uk-investigation">January 10-12, 2026: Platform blocks and formal UK investigation</h3>
<p>Indonesia moved to block Grok access, citing risks of pornographic content.
The UK regulator Ofcom announced a <strong>formal investigation</strong> into X, referencing concerns that Grok was being used to create and share sexualized images of children.</p>
<h3 id="user-content-january-11-13-2026-malaysia-restricts-access-and-announces-legal-action">January 11-13, 2026: Malaysia restricts access and announces legal action</h3>
<p>Malaysia restricted access to Grok and then announced legal action against X and xAI over alleged safety failures related to Grok's misuse.</p>
<h3 id="user-content-january-2026-philippines-orders-a-block-brazil-consumer-group-escalates">January 2026: Philippines orders a block; Brazil consumer group escalates</h3>
<p>Reuters reported the Philippines ordered access to Grok blocked, citing the risk of pornographic content.
In Brazil, the consumer group IDEC publicly called for government action regarding Grok's operation following reports of harmful sexualized image generation.</p>
<h3 id="user-content-january-15-17-2026-first-high-profile-lawsuit-us-state-action">January 15-17, 2026: First high-profile lawsuit; U.S. state action</h3>
<p>Ashley St. Clair filed a lawsuit in New York against xAI alleging Grok enabled the creation and dissemination of sexually exploitative deepfake imagery involving her likeness.
California's Attorney General announced action demanding xAI prevent the creation and distribution of nonconsensual sexualized imagery and deepfakes involving minors, escalating U.S. pressure.</p>
<h2 id="user-content-what-xxai-changed---and-what-reporting-found-still-worked">What X/xAI Changed - and What Reporting Found Still Worked</h2>
<h3 id="user-content-announced-restrictions">Announced restrictions</h3>
<p>Across January 2026, X and xAI announced multiple changes, including limiting some image features on X to paying users and implementing measures intended to prevent edits depicting real people in revealing clothing in certain jurisdictions.</p>
<h3 id="user-content-ongoing-gaps-and-patchwork-enforcement">Ongoing gaps and "patchwork" enforcement</h3>
<p>Major outlets reported that restrictions were uneven across product surfaces:</p>
<ul>
<li>The Washington Post reported that even after X said it had stopped certain sexualized edit behavior within X in some locations, similar functionality remained available via the <strong>standalone Grok app</strong>.</li>
<li>The Associated Press reported that despite claimed restrictions, testing indicated image editing that produced sexualized outputs could still be possible in at least some situations, underscoring enforcement and geoblocking limitations.</li>
</ul>
<p>Taken together, these reports suggest that limiting one interface (for example, a reply workflow on X) did not fully address abuse when other access points to the model remained available.</p>
<h2 id="user-content-why-this-happened-based-on-public-reporting">Why This Happened (Based on Public Reporting)</h2>
<h3 id="user-content-product-design-choices-increased-blast-radius">Product design choices increased blast radius</h3>
<p>A central factor identified by multiple reports was <strong>frictionless public deployment</strong>: enabling rapid, on-platform image edits in public reply threads, without requiring the subject's consent (and, in some reports, without clear notification to the original poster) made harassment scalable.</p>
<h3 id="user-content-a-permissive-nsfw-positioning-existed-elsewhere-in-the-product-line">A permissive NSFW positioning existed elsewhere in the product line</h3>
<p>Separate from the January 2026 controversy, xAI had previously launched Grok Imagine with a "Spicy" mode designed to allow NSFW content - an approach that drew criticism because it intentionally expanded the range of sexual content the product could generate, compared with peers that emphasize stricter constraints.
European officials also referenced concerns arising after the introduction of paid "Spicy Mode" functionality in the prior year's product cycle when discussing sexualized imagery risks.</p>
<h2 id="user-content-legal-and-regulatory-context-non-exhaustive">Legal and Regulatory Context (Non-exhaustive)</h2>
<ul>
<li><strong>United Kingdom:</strong> Ofcom's investigation is grounded in UK online safety rules requiring platforms to assess and mitigate risks - especially for children - and remove illegal content. Reuters reported Prime Minister Keir Starmer called the content "disgusting" and "unlawful" while backing regulator action.</li>
<li><strong>European Union:</strong> The European Commission ordered X to preserve Grok-related data and documents through the end of 2026, signaling ongoing scrutiny of unlawful content and platform compliance.</li>
<li><strong>United States:</strong> Wired reported that Congress passed the TAKE IT DOWN Act in 2025, targeting nonconsensual intimate imagery (including deepfakes) and imposing obligations on platforms to respond to reports. Separate reporting noted the first high-profile Grok-related lawsuit was filed in New York in mid-January 2026.</li>
</ul>
<h2 id="user-content-current-status-as-of-february-2-2026">Current Status (as of February 2, 2026)</h2>
<p>Based on reporting through mid-January 2026:</p>
<ul>
<li><strong>X has imposed restrictions</strong> on certain Grok image-generation/editing features on the X platform, including limiting some capabilities to paying users and implementing location-based controls.</li>
<li><strong>Regulatory and legal pressure continues to expand</strong>, with platform blocks, investigations, and legal action across multiple countries and at least one U.S. state.</li>
<li><strong>Independent testing and reporting indicate the problem is not "fully fixed,"</strong> particularly where the model remains accessible through alternative apps or web experiences.</li>
</ul>
<h2 id="user-content-key-takeaways">Key Takeaways</h2>
<ol>
<li><strong>Public-by-default generation amplified harm.</strong> A high-volume, public reply workflow turned what might have been private misuse into a visible harassment engine.</li>
<li><strong>Partial restrictions can shift abuse rather than stop it.</strong> Limiting one interface does not eliminate abuse if parallel product surfaces retain similar capabilities.</li>
<li><strong>Regulators are treating AI image misuse as a mainstream platform-safety issue.</strong> The speed and breadth of investigations across the UK, EU, and Asia show that sexualized image misuse is increasingly being handled as an online safety and illegal-content problem - not merely a moderation dispute.</li>
</ol>
<hr>
<h2 id="user-content-sources">Sources</h2>
<p>Primary and high-reliability reporting and documentation used in this article:</p>
<ul>
<li>Reuters (Jan 4, 2026) - <a href="https://www.reuters.com/legal/litigation/grok-says-safeguard-lapses-led-images-minors-minimal-clothing-x-2026-01-02/?utm_source=chatgpt.com">Grok says safeguard lapses led to images of minors in minimal clothing on X</a></li>
<li>Reuters (Jan 8, 2026) - <a href="https://www.reuters.com/world/eu-commission-has-ordered-x-retain-all-grok-documents-until-end-2026-2026-01-08/?utm_source=chatgpt.com">EU Commission ordered X to retain all Grok documents until end 2026, spokesperson says</a></li>
<li>Reuters (Jan 9, 2026) - <a href="https://www.reuters.com/legal/litigation/elon-musks-grok-faces-global-scrutiny-sexualised-ai-photos-2026-01-09/?utm_source=chatgpt.com">Elon Musk's Grok faces global scrutiny for sexualised AI photos</a></li>
<li>Reuters (Jan 14, 2026) - <a href="https://www.reuters.com/world/asia-pacific/musk-says-unaware-grok-generating-explicit-images-minors-2026-01-14/?utm_source=chatgpt.com">Musk's xAI curbs Grok image editing</a></li>
<li>Reuters (Jan 15, 2026) - <a href="https://www.reuters.com/legal/government/musk-dealt-blow-over-grok-deepfakes-regulatory-fight-far-over-2026-01-15/?utm_source=chatgpt.com">Musk dealt blow over Grok deepfakes, but regulatory fight far from over</a></li>
<li>Ofcom (UK) (Jan 12, 2026) - <a href="https://www.ofcom.org.uk/online-safety/illegal-and-harmful-content/ofcom-launches-investigation-into-x-over-grok-sexualised-imagery?utm_source=chatgpt.com">Ofcom launches investigation into X over Grok sexualised imagery</a></li>
<li>UK Parliament (Education Committee) (Jan 13, 2026) - <a href="https://committees.parliament.uk/committee/203/education-committee/news/211282/mps-deeply-alarmed-by-grok-ai-images-says-education-committee-chair/?utm_source=chatgpt.com">MPs deeply alarmed by Grok AI images, says Education Committee chair</a></li>
<li>AI Forensics (Jan 5, 2026) - <a href="https://aiforensics.org/work/grok-unleashed?utm_source=chatgpt.com">Grok generating flood of sexualized images of women and minors</a></li>
<li>Copyleaks (Dec 31, 2025) - <a href="https://copyleaks.com/blog/grok-and-nonconsensual-image-manipulation?utm_source=chatgpt.com">Grok and the rise of nonconsensual image manipulation</a></li>
<li>Copyleaks (Jan 2026) - <a href="https://copyleaks.com/blog/nonconsensual-image-manipulation-continues-via-grok?utm_source=chatgpt.com">Nonconsensual image manipulation via Grok continues</a></li>
<li>The Washington Post (Jan 15, 2026) - <a href="https://www.washingtonpost.com/technology/2026/01/15/grok-ai-image-generator-sexualized/?utm_source=chatgpt.com">Stand-alone Grok app still produced sexualized image edits after X curbed access</a></li>
<li>WIRED - <a href="https://www.wired.com/story/elon-musks-grok-undressing-problem-isnt-fixed?utm_source=chatgpt.com">Grok image misuse still not fixed</a></li>
<li>The Guardian (Jan 2, 2026) - <a href="https://www.theguardian.com/technology/2026/jan/02/elon-musk-grok-ai-children-photos?utm_source=chatgpt.com">Grok AI generates images of minors in minimal clothing</a></li>
<li>The Guardian (Jan 14, 2026) - <a href="https://www.theguardian.com/technology/2026/jan/14/elon-musk-grok-ai-explicit-images?utm_source=chatgpt.com">X to block Grok AI tool from creating sexualised images</a></li>
<li>The Guardian (Jan 15, 2026) - <a href="https://www.theguardian.com/technology/2026/jan/15/mother-of-one-of-elon-musks-sons-sues-over-grok-generated-explicit-images?utm_source=chatgpt.com">Ashley St. Clair lawsuit over Grok-generated explicit images</a></li>
<li>The Guardian (Jan 16, 2026) - <a href="https://www.theguardian.com/technology/2026/jan/16/x-still-allowing-sexualised-images-grok-ai-nudification?utm_source=chatgpt.com">X still allowing users to post sexualised images generated by Grok</a></li>
<li>The Verge - <a href="https://www.theverge.com/news/718795/xai-grok-imagine-video-generator-spicy-mode?utm_source=chatgpt.com">xAI's Grok image and video generator has a "spicy" mode</a></li>
<li>Tech Policy Press - <a href="https://techpolicy.press/tracking-regulator-responses-to-the-grok-undressing-controversy?utm_source=chatgpt.com">Tracking regulator responses to the Grok image-editing controversy</a></li>
</ul>
<hr>
<p><strong>Last updated:</strong> February 2, 2026<br>
<strong>Editorial note:</strong> This is an evolving situation. Facts, restrictions, and enforcement may change quickly as regulators, courts, and platform operators respond.</p>


A structured timeline of the Grok image-editing controversy on X, the regulatory responses across multiple countries, and what reporting says still works as of February 2, 2026.

X Deleted Grok's Image Tab - The Search Hack That Still Shows All Grok Images


When you're building apps that generate images or video via API, pricing structures can be confusing. Some providers charge per image, others per megapixel, some by compute time, and video generators mix per-second and per-frame billing. This post cuts through the noise with direct comparisons in simple units: **USD per image** (at standard resolution) and **USD per second** of video.

All prices reflect publicly documented first-party API rates as of January 17, 2025. I've normalized everything to help you compare apples to apples.

## Image Generation Pricing

Standard resolution reference: 1024×1024 (1 megapixel) unless noted.

| Provider | Model | Price per image (1024×1024) | Price per 1M images | Notes |
| --- | --- | --- | --- | --- |
| **Black Forest Labs** | FLUX.1 [pro] | $0.055 | $55,000 | Top quality; fastest inference |
| Black Forest Labs | FLUX.1 [dev] | $0.025 | $25,000 | Developer model; commercial use allowed |
| Black Forest Labs | FLUX.1 [schnell] | $0.003 | $3,000 | Ultra-fast; 1-4 steps; good for prototypes |
| **Google** | Imagen 4 Ultra | $0.080 | $80,000 | Highest fidelity Imagen |
| Google | Imagen 4 | $0.040 | $40,000 | Standard flagship model |
| Google | Imagen 3 | $0.020 | $20,000 | Previous generation; still high quality |
| **OpenAI** | DALL·E 3 (1024×1024) | $0.040 | $40,000 | Standard quality |
| OpenAI | DALL·E 3 (1024×1792 / 1792×1024) | $0.080 | $80,000 | HD resolution |
| **Stability AI** | Stable Diffusion 3.5 Large | $0.065 | $65,000 | 8B parameter model |
| Stability AI | Stable Diffusion 3.5 Medium | $0.035 | $35,000 | Balanced speed/quality |
| Stability AI | Stable Diffusion 3 Large Turbo | $0.040 | $40,000 | Fast variant |
| Stability AI | SDXL 1.0 | $0.030 | $30,000 | Predecessor; still popular |
| **Midjourney** | v6 (via API partners) | ~$0.050 | ~$50,000 | Estimated; no official API yet |
| **Leonardo.ai** | PhotoReal | $0.012 | $12,000 | Photorealistic preset |
| Leonardo.ai | DreamShaper XL | $0.008 | $8,000 | General purpose |
| **Ideogram** | v2 Turbo | $0.008 | $8,000 | Fast generation |
| Ideogram | v2 | $0.016 | $16,000 | Standard quality |

### Key Observations

- **Budget tier:** FLUX.1 [schnell], Leonardo, and Ideogram v2 Turbo offer sub-$0.01/image for high-volume use cases.
- **Premium tier:** FLUX.1 [pro] and Imagen 4 Ultra lead in quality but cost $0.055–$0.080/image.
- **Sweet spot:** SD 3.5 Medium, FLUX.1 [dev], and Imagen 3 balance quality and cost at $0.020–$0.035/image.

## Video Generation Pricing

Pricing is per second of generated video at standard resolution (typically 720p or 1080p). Most models charge by output duration, not input prompt length.

| Provider | Model | Price per second | Price per minute | Price per 10-min video | Notes |
| --- | --- | --- | --- | --- | --- |
| **Google** | Veo | $0.10 | $6.00 | $60.00 | Up to 1080p; photorealistic |
| **Runway** | Gen-3 Alpha Turbo | $0.05 | $3.00 | $30.00 | Fast generation; 720p default |
| Runway | Gen-2 | $0.08 | $4.80 | $48.00 | Predecessor model |
| **Stability AI** | Stable Video Diffusion | $0.10 | $6.00 | $60.00 | Open model via API |
| **Pika Labs** | Pika 2.0 | $0.04 | $2.40 | $24.00 | Estimated; rapid motion |
| **Luma AI** | Dream Machine | $0.05 | $3.00 | $30.00 | High-quality motion |
| **Haiper AI** | Haiper 2.0 | $0.02 | $1.20 | $12.00 | Budget option; good quality |

### Key Observations

- **Premium tier:** Veo and Stable Video Diffusion at $0.10/second offer cutting-edge quality but expensive for long-form content.
- **Standard tier:** Runway Gen-3 Alpha Turbo and Luma Dream Machine at $0.05/second balance speed and cost.
- **Budget tier:** Haiper 2.0 at $0.02/second is the most affordable option for experimentation.

## Cost Projections

To help you estimate real-world spend, here are monthly projections at different usage levels.

### Image Generation (1024×1024)

| Daily images | Monthly images | FLUX.1 [schnell] ($0.003) | Imagen 3 ($0.020) | FLUX.1 [pro] ($0.055) |
| --- | --- | --- | --- | --- |
| 100 | 3,000 | $9 | $60 | $165 |
| 1,000 | 30,000 | $90 | $600 | $1,650 |
| 10,000 | 300,000 | $900 | $6,000 | $16,500 |
| 100,000 | 3,000,000 | $9,000 | $60,000 | $165,000 |

### Video Generation (per 5-second clip)

| Daily clips | Monthly clips | Haiper 2.0 ($0.02/sec) | Runway Gen-3 ($0.05/sec) | Veo ($0.10/sec) |
| --- | --- | --- | --- | --- |
| 10 | 300 | $30 | $75 | $150 |
| 100 | 3,000 | $300 | $750 | $1,500 |
| 1,000 | 30,000 | $3,000 | $7,500 | $15,000 |
| 10,000 | 300,000 | $30,000 | $75,000 | $150,000 |

## Additional Cost Factors

Beyond raw generation costs, consider:

- **Variations/revisions:** Most providers charge per generation, not per final output. If you generate 5 variations and pick 1, you pay for 5.
- **Resolution scaling:** Higher resolutions (e.g., 2048×2048 or 4K video) typically cost 2–4× the base price.
- **Batch/volume discounts:** Some providers (Google, Stability) offer enterprise pricing at scale.
- **Inference speed:** Faster models (FLUX.1 [schnell], Gen-3 Turbo) may reduce wall-clock costs by letting you iterate more quickly.
- **Fine-tuning:** Custom model training can add one-time costs of $50–$500+ but may reduce per-generation costs if you're doing high-volume brand-specific work.

## Download the Full Comparison

Want the raw data with formulas, tiered pricing breakdowns, and enterprise contact info?

**[Download the spreadsheet (XLSX)](/downloads/direct_provider_pricing_comparison.xlsx)**

The spreadsheet includes:
- Per-megapixel pricing for all image models
- Video pricing at 720p, 1080p, and 4K
- Volume discount thresholds where available
- Direct API documentation links

## Choosing the Right Provider

### For prototyping and MVPs
- **Image:** FLUX.1 [schnell] or Ideogram v2 Turbo (sub-$0.01/image)
- **Video:** Haiper 2.0 ($0.02/second)

### For production apps with quality focus
- **Image:** FLUX.1 [pro] or Imagen 4 Ultra ($0.055–$0.080/image)
- **Video:** Veo or Gen-3 Alpha Turbo ($0.05–$0.10/second)

### For high-volume, cost-sensitive workflows
- **Image:** Leonardo or SD 3.5 Medium ($0.008–$0.035/image)
- **Video:** Pika 2.0 or Haiper 2.0 ($0.02–$0.04/second)

## Methodology Notes

- All prices are first-party API rates (not aggregator markup).
- Image pricing normalized to 1024×1024 (1MP); providers that charge per megapixel are calculated at 1MP.
- Video pricing is per second of output at standard resolution (720p–1080p).
- Enterprise/volume pricing not included; these are pay-as-you-go public rates.
- Prices verified from official documentation as of January 17, 2025.

## Related Tools

- **[Pricing Calculator](/tools/pricing-calculator)** - Interactive calculator for estimating your monthly API costs across providers

## Final Thoughts

API pricing for generative models is stabilizing, but the range is still wide: $0.003 to $0.080 per image and $0.02 to $0.10 per second of video. Your best choice depends on whether you're optimizing for cost, quality, speed, or a balance of all three.

As models improve and competition increases, expect continued price compression in the standard tier while premium tiers maintain differentiation through quality and brand-specific fine-tuning options.

For the most current rates, always verify directly with provider documentation—this space moves fast.


<p>When you're building apps that generate images or video via API, pricing structures can be confusing. Some providers charge per image, others per megapixel, some by compute time, and video generators mix per-second and per-frame billing. This post cuts through the noise with direct comparisons in simple units: <strong>USD per image</strong> (at standard resolution) and <strong>USD per second</strong> of video.</p>
<p>All prices reflect publicly documented first-party API rates as of January 17, 2025. I've normalized everything to help you compare apples to apples.</p>
<h2 id="user-content-image-generation-pricing">Image Generation Pricing</h2>
<p>Standard resolution reference: 1024×1024 (1 megapixel) unless noted.</p>
<table>
<thead>
<tr>
<th>Provider</th>
<th>Model</th>
<th>Price per image (1024×1024)</th>
<th>Price per 1M images</th>
<th>Notes</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Black Forest Labs</strong></td>
<td>FLUX.1 [pro]</td>
<td>$0.055</td>
<td>$55,000</td>
<td>Top quality; fastest inference</td>
</tr>
<tr>
<td>Black Forest Labs</td>
<td>FLUX.1 [dev]</td>
<td>$0.025</td>
<td>$25,000</td>
<td>Developer model; commercial use allowed</td>
</tr>
<tr>
<td>Black Forest Labs</td>
<td>FLUX.1 [schnell]</td>
<td>$0.003</td>
<td>$3,000</td>
<td>Ultra-fast; 1-4 steps; good for prototypes</td>
</tr>
<tr>
<td><strong>Google</strong></td>
<td>Imagen 4 Ultra</td>
<td>$0.080</td>
<td>$80,000</td>
<td>Highest fidelity Imagen</td>
</tr>
<tr>
<td>Google</td>
<td>Imagen 4</td>
<td>$0.040</td>
<td>$40,000</td>
<td>Standard flagship model</td>
</tr>
<tr>
<td>Google</td>
<td>Imagen 3</td>
<td>$0.020</td>
<td>$20,000</td>
<td>Previous generation; still high quality</td>
</tr>
<tr>
<td><strong>OpenAI</strong></td>
<td>DALL·E 3 (1024×1024)</td>
<td>$0.040</td>
<td>$40,000</td>
<td>Standard quality</td>
</tr>
<tr>
<td>OpenAI</td>
<td>DALL·E 3 (1024×1792 / 1792×1024)</td>
<td>$0.080</td>
<td>$80,000</td>
<td>HD resolution</td>
</tr>
<tr>
<td><strong>Stability AI</strong></td>
<td>Stable Diffusion 3.5 Large</td>
<td>$0.065</td>
<td>$65,000</td>
<td>8B parameter model</td>
</tr>
<tr>
<td>Stability AI</td>
<td>Stable Diffusion 3.5 Medium</td>
<td>$0.035</td>
<td>$35,000</td>
<td>Balanced speed/quality</td>
</tr>
<tr>
<td>Stability AI</td>
<td>Stable Diffusion 3 Large Turbo</td>
<td>$0.040</td>
<td>$40,000</td>
<td>Fast variant</td>
</tr>
<tr>
<td>Stability AI</td>
<td>SDXL 1.0</td>
<td>$0.030</td>
<td>$30,000</td>
<td>Predecessor; still popular</td>
</tr>
<tr>
<td><strong>Midjourney</strong></td>
<td>v6 (via API partners)</td>
<td>~$0.050</td>
<td>~$50,000</td>
<td>Estimated; no official API yet</td>
</tr>
<tr>
<td><strong>Leonardo.ai</strong></td>
<td>PhotoReal</td>
<td>$0.012</td>
<td>$12,000</td>
<td>Photorealistic preset</td>
</tr>
<tr>
<td>Leonardo.ai</td>
<td>DreamShaper XL</td>
<td>$0.008</td>
<td>$8,000</td>
<td>General purpose</td>
</tr>
<tr>
<td><strong>Ideogram</strong></td>
<td>v2 Turbo</td>
<td>$0.008</td>
<td>$8,000</td>
<td>Fast generation</td>
</tr>
<tr>
<td>Ideogram</td>
<td>v2</td>
<td>$0.016</td>
<td>$16,000</td>
<td>Standard quality</td>
</tr>
</tbody>
</table>
<h3 id="user-content-key-observations">Key Observations</h3>
<ul>
<li><strong>Budget tier:</strong> FLUX.1 [schnell], Leonardo, and Ideogram v2 Turbo offer sub-$0.01/image for high-volume use cases.</li>
<li><strong>Premium tier:</strong> FLUX.1 [pro] and Imagen 4 Ultra lead in quality but cost $0.055–$0.080/image.</li>
<li><strong>Sweet spot:</strong> SD 3.5 Medium, FLUX.1 [dev], and Imagen 3 balance quality and cost at $0.020–$0.035/image.</li>
</ul>
<h2 id="user-content-video-generation-pricing">Video Generation Pricing</h2>
<p>Pricing is per second of generated video at standard resolution (typically 720p or 1080p). Most models charge by output duration, not input prompt length.</p>
<table>
<thead>
<tr>
<th>Provider</th>
<th>Model</th>
<th>Price per second</th>
<th>Price per minute</th>
<th>Price per 10-min video</th>
<th>Notes</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Google</strong></td>
<td>Veo</td>
<td>$0.10</td>
<td>$6.00</td>
<td>$60.00</td>
<td>Up to 1080p; photorealistic</td>
</tr>
<tr>
<td><strong>Runway</strong></td>
<td>Gen-3 Alpha Turbo</td>
<td>$0.05</td>
<td>$3.00</td>
<td>$30.00</td>
<td>Fast generation; 720p default</td>
</tr>
<tr>
<td>Runway</td>
<td>Gen-2</td>
<td>$0.08</td>
<td>$4.80</td>
<td>$48.00</td>
<td>Predecessor model</td>
</tr>
<tr>
<td><strong>Stability AI</strong></td>
<td>Stable Video Diffusion</td>
<td>$0.10</td>
<td>$6.00</td>
<td>$60.00</td>
<td>Open model via API</td>
</tr>
<tr>
<td><strong>Pika Labs</strong></td>
<td>Pika 2.0</td>
<td>$0.04</td>
<td>$2.40</td>
<td>$24.00</td>
<td>Estimated; rapid motion</td>
</tr>
<tr>
<td><strong>Luma AI</strong></td>
<td>Dream Machine</td>
<td>$0.05</td>
<td>$3.00</td>
<td>$30.00</td>
<td>High-quality motion</td>
</tr>
<tr>
<td><strong>Haiper AI</strong></td>
<td>Haiper 2.0</td>
<td>$0.02</td>
<td>$1.20</td>
<td>$12.00</td>
<td>Budget option; good quality</td>
</tr>
</tbody>
</table>
<h3 id="user-content-key-observations-1">Key Observations</h3>
<ul>
<li><strong>Premium tier:</strong> Veo and Stable Video Diffusion at $0.10/second offer cutting-edge quality but expensive for long-form content.</li>
<li><strong>Standard tier:</strong> Runway Gen-3 Alpha Turbo and Luma Dream Machine at $0.05/second balance speed and cost.</li>
<li><strong>Budget tier:</strong> Haiper 2.0 at $0.02/second is the most affordable option for experimentation.</li>
</ul>
<h2 id="user-content-cost-projections">Cost Projections</h2>
<p>To help you estimate real-world spend, here are monthly projections at different usage levels.</p>
<h3 id="user-content-image-generation-10241024">Image Generation (1024×1024)</h3>
<table>
<thead>
<tr>
<th>Daily images</th>
<th>Monthly images</th>
<th>FLUX.1 [schnell] ($0.003)</th>
<th>Imagen 3 ($0.020)</th>
<th>FLUX.1 [pro] ($0.055)</th>
</tr>
</thead>
<tbody>
<tr>
<td>100</td>
<td>3,000</td>
<td>$9</td>
<td>$60</td>
<td>$165</td>
</tr>
<tr>
<td>1,000</td>
<td>30,000</td>
<td>$90</td>
<td>$600</td>
<td>$1,650</td>
</tr>
<tr>
<td>10,000</td>
<td>300,000</td>
<td>$900</td>
<td>$6,000</td>
<td>$16,500</td>
</tr>
<tr>
<td>100,000</td>
<td>3,000,000</td>
<td>$9,000</td>
<td>$60,000</td>
<td>$165,000</td>
</tr>
</tbody>
</table>
<h3 id="user-content-video-generation-per-5-second-clip">Video Generation (per 5-second clip)</h3>
<table>
<thead>
<tr>
<th>Daily clips</th>
<th>Monthly clips</th>
<th>Haiper 2.0 ($0.02/sec)</th>
<th>Runway Gen-3 ($0.05/sec)</th>
<th>Veo ($0.10/sec)</th>
</tr>
</thead>
<tbody>
<tr>
<td>10</td>
<td>300</td>
<td>$30</td>
<td>$75</td>
<td>$150</td>
</tr>
<tr>
<td>100</td>
<td>3,000</td>
<td>$300</td>
<td>$750</td>
<td>$1,500</td>
</tr>
<tr>
<td>1,000</td>
<td>30,000</td>
<td>$3,000</td>
<td>$7,500</td>
<td>$15,000</td>
</tr>
<tr>
<td>10,000</td>
<td>300,000</td>
<td>$30,000</td>
<td>$75,000</td>
<td>$150,000</td>
</tr>
</tbody>
</table>
<h2 id="user-content-additional-cost-factors">Additional Cost Factors</h2>
<p>Beyond raw generation costs, consider:</p>
<ul>
<li><strong>Variations/revisions:</strong> Most providers charge per generation, not per final output. If you generate 5 variations and pick 1, you pay for 5.</li>
<li><strong>Resolution scaling:</strong> Higher resolutions (e.g., 2048×2048 or 4K video) typically cost 2–4× the base price.</li>
<li><strong>Batch/volume discounts:</strong> Some providers (Google, Stability) offer enterprise pricing at scale.</li>
<li><strong>Inference speed:</strong> Faster models (FLUX.1 [schnell], Gen-3 Turbo) may reduce wall-clock costs by letting you iterate more quickly.</li>
<li><strong>Fine-tuning:</strong> Custom model training can add one-time costs of $50–$500+ but may reduce per-generation costs if you're doing high-volume brand-specific work.</li>
</ul>
<h2 id="user-content-download-the-full-comparison">Download the Full Comparison</h2>
<p>Want the raw data with formulas, tiered pricing breakdowns, and enterprise contact info?</p>
<p><strong><a href="/downloads/direct_provider_pricing_comparison.xlsx">Download the spreadsheet (XLSX)</a></strong></p>
<p>The spreadsheet includes:</p>
<ul>
<li>Per-megapixel pricing for all image models</li>
<li>Video pricing at 720p, 1080p, and 4K</li>
<li>Volume discount thresholds where available</li>
<li>Direct API documentation links</li>
</ul>
<h2 id="user-content-choosing-the-right-provider">Choosing the Right Provider</h2>
<h3 id="user-content-for-prototyping-and-mvps">For prototyping and MVPs</h3>
<ul>
<li><strong>Image:</strong> FLUX.1 [schnell] or Ideogram v2 Turbo (sub-$0.01/image)</li>
<li><strong>Video:</strong> Haiper 2.0 ($0.02/second)</li>
</ul>
<h3 id="user-content-for-production-apps-with-quality-focus">For production apps with quality focus</h3>
<ul>
<li><strong>Image:</strong> FLUX.1 [pro] or Imagen 4 Ultra ($0.055–$0.080/image)</li>
<li><strong>Video:</strong> Veo or Gen-3 Alpha Turbo ($0.05–$0.10/second)</li>
</ul>
<h3 id="user-content-for-high-volume-cost-sensitive-workflows">For high-volume, cost-sensitive workflows</h3>
<ul>
<li><strong>Image:</strong> Leonardo or SD 3.5 Medium ($0.008–$0.035/image)</li>
<li><strong>Video:</strong> Pika 2.0 or Haiper 2.0 ($0.02–$0.04/second)</li>
</ul>
<h2 id="user-content-methodology-notes">Methodology Notes</h2>
<ul>
<li>All prices are first-party API rates (not aggregator markup).</li>
<li>Image pricing normalized to 1024×1024 (1MP); providers that charge per megapixel are calculated at 1MP.</li>
<li>Video pricing is per second of output at standard resolution (720p–1080p).</li>
<li>Enterprise/volume pricing not included; these are pay-as-you-go public rates.</li>
<li>Prices verified from official documentation as of January 17, 2025.</li>
</ul>
<h2 id="user-content-related-tools">Related Tools</h2>
<ul>
<li><strong><a href="/tools/pricing-calculator">Pricing Calculator</a></strong> - Interactive calculator for estimating your monthly API costs across providers</li>
</ul>
<h2 id="user-content-final-thoughts">Final Thoughts</h2>
<p>API pricing for generative models is stabilizing, but the range is still wide: $0.003 to $0.080 per image and $0.02 to $0.10 per second of video. Your best choice depends on whether you're optimizing for cost, quality, speed, or a balance of all three.</p>
<p>As models improve and competition increases, expect continued price compression in the standard tier while premium tiers maintain differentiation through quality and brand-specific fine-tuning options.</p>
<p>For the most current rates, always verify directly with provider documentation—this space moves fast.</p>


A clean, no-nonsense comparison of first-party API pricing for image and video generation models, normalized to USD per image and USD per second. Includes FLUX, Imagen, OpenAI, Runway, and more.

Direct Provider Pricing for Image + Video Generation APIs (USD)


You want the numbers, not the marketing. Here are the quotas, mechanics, and limits the X API actually enforces after the launch of Pay-Per-Use pricing, plus what still matters from the older Free/Basic/Pro world.

The punchline: pricing is no longer "pick a tier, live inside monthly caps." It's now credit-based pay-per-usage (pay only for what you consume), while rate limits still exist and still bite.

## What changed: Pay-Per-Use pricing is now the default mental model

### Old world (legacy tiers)

You bought a monthly plan (Free/Basic/Pro) and designed around:

- monthly read caps
- monthly posting caps
- "post caps" (consumption caps)
- top-ups (sometimes)

### New world (pay-per-usage credits)

You design around:

- what endpoints you call
- how much data they return
- deduplication rules
- your credit balance and spending limits
- rate limits per endpoint (still real)

The plan names are still noise. The billing and limits are the product.

## Pay-Per-Use pricing: how it actually works

| Mechanic | What it means in practice |
| --- | --- |
| Credit-based | You pre-purchase credits; requests deduct from that balance. |
| Per-endpoint pricing | Different endpoints cost different amounts (the current $ rates live in the Developer Console). |
| No monthly caps | There's no built-in "15k reads and you're done." You can keep going until your credits or spending limits stop you. |
| Real-time tracking | Usage and cost tracking is surfaced in the Developer Console while you build. |
| Usage is tracked at the app level | Your bill follows your developer app usage (not a vague "project total" mental model). |
| Enterprise still exists | High-volume customers can negotiate custom pricing/support/rate limits. |

Important nuance: the docs are explicit that the current per-endpoint pricing numbers are in the Developer Console (not reliably mirrored on public pages). So if someone shows you a pricing screenshot, treat it as "may already be outdated."

## Deduplication: the one rule that changes your caching strategy

Pay-per-usage billing comes with daily (UTC) deduplication:

- Billable resources are deduplicated within a 24-hour UTC day window
- Pull the same Post again in that window: it should not incur an additional charge
- The window resets at midnight UTC
- X describes this as a soft guarantee (rare edge cases like outages can break it)

Translated to engineering language:

- You still cache (latency and rate-limits), but deduplication reduces the "double charge" penalty if you re-fetch the same IDs in the same day.
- Design pipelines to avoid re-hydrating the same IDs across days unless you need updated fields.

## What counts as billable usage

Posts retrieved from these categories count toward usage:

| Category | Examples |
| --- | --- |
| Post lookup | `GET /2/tweets` |
| Search | recent search, full-archive search |
| Streaming | filtered stream (and webhooks) |
| Timelines | user posts, user mentions |
| Engagement | liked posts, bookmarks |
| Lists | list posts |
| Spaces | spaces lookup |

Two more billing rules that matter:

- Only successful responses that return data are billed (failed requests do not count).
- Streaming billing is based on unique posts delivered, subject to daily deduplication.

## Controlling spend: how to avoid "surprise invoice" stories

Pay-per-use has fewer cliffs, but more ways to accidentally set money on fire.

### 1) Spending limits (hard stop)

You can set a max spend per billing cycle; once hit, requests are blocked until the next cycle.

### 2) Auto-recharge (keep services running)

You can auto-top-up your balance based on a threshold:

- "Trigger when balance < $X"
- "Recharge amount = $Y"

### 3) Credit balance behavior (including going negative)

X notes that your balance can go slightly negative, and your API requests will be blocked until you add credits to cover it.

### 4) Cache + dedup-aware design

Deduplication helps, but it is not magic. You still win by:

- caching hydrated Posts
- storing IDs from search/stream results and avoiding re-hydration
- keeping queries tight (especially for search)

## Monitoring usage programmatically

You can pull usage data via the Usage endpoint (daily Post consumption counts). This is the simplest building block for alerts and budget enforcement.

Example call:

```bash
curl "https://api.x.com/2/usage/tweets" \
  -H "Authorization: Bearer $BEARER_TOKEN"
```

If you are doing anything serious, wire this into:

- budget alerts
- anomaly detection (spikes in search usage are common)
- cost attribution per feature or customer

## Rate limits still exist (and are separate from billing)

Pay-per-use did not remove rate limits. It just decoupled pricing from monthly caps.

X documents rate limits per endpoint, generally shown as:

- Per App
- Per User
- mostly per 15-minute windows (unless otherwise noted)

You can also read rate-limit headers in responses:

- `x-rate-limit-limit`
- `x-rate-limit-remaining`
- `x-rate-limit-reset`

### A few concrete examples

| Endpoint | Per App | Per User | Notes |
| --- | ---: | ---: | --- |
| `GET /2/tweets` | 3,500 / 15 min | 5,000 / 15 min | Lookup (batch) |
| `GET /2/tweets/:id` | 450 / 15 min | 900 / 15 min | Single lookup |
| `GET /2/tweets/search/recent` | 450 / 15 min | 300 / 15 min | 10 default, 100 max results |
| `GET /2/tweets/search/all` | 300 / 15 min + 1/sec | 1/sec | Full archive; larger query lengths for Enterprise |
| `POST /2/tweets` | 10,000 / 24 hrs | 100 / 15 min | Create posts |

Engineering takeaway: rate limits are your throughput ceiling; credits/spending limits are your budget ceiling. Design for both.

## Search access: what's available to whom (and why it matters)

Search has become cleaner to reason about:

- Recent search: last 7 days, available to all developers
- Full-archive search: complete archive, available to Pay-per-use and Enterprise

That makes "search/backfill" workloads much more about cost and rate limiting, and less about "which tier do I have?"

## Enterprise: still the "talk to sales" layer

For high-volume and custom needs, X positions Enterprise as:

- dedicated support
- custom rate limits
- volume discounts or custom pricing
- "complete data access" style options

If you are at the point where you are negotiating, the public docs stop being your limiting factor anyway.

## Legacy subscription tiers (Free / Basic / Pro): where they still show up

X calls out that legacy Basic/Pro subscription packages may still exist for some accounts, and you can:

- opt in to pay-per-use from the Developer Console
- switch back to the legacy plan from settings

If you are starting fresh today, assume:

- Pay-per-use is the default pricing path
- legacy tiers are "legacy" for a reason (availability may vary)

Practical advice: treat legacy tier quotas (reads/posts/top-ups) as historical context unless the Developer Console shows you those plans as purchasable for your account.

## FAQ

### Does X API still have monthly caps?

Pay-per-usage pricing is described as having no monthly caps; you pay based on usage and can keep going until you hit spending limits or run out of credits.

### What happens if I run out of credits?

Requests fail until you buy more credits. Auto-recharge and alerts are how you avoid downtime.

### Do failed requests get billed?

No. Only successful responses that return data are billed.

### Is deduplication guaranteed?

X describes deduplication as a soft guarantee: it applies in most cases, but rare edge cases (like outages) can break it.

### Where do I find the actual $ pricing per endpoint?

In the Developer Console. Public docs explain the model and what is billable, but current per-endpoint rates are shown in-console.

## Bottom line

In 2026, the X API product is not "Free vs Basic vs Pro."

It's:

- pay-per-usage credits
- billable endpoints and what they return
- daily deduplication
- endpoint rate limits
- spending controls and monitoring

Build around those. Everything else is branding.

| Endpoint | Pro | Basic | Free |
| --- | --- | --- | --- |
| `GET /2/compliance/jobs` | 150/15m A | 5/15m A | 1/15m A |
| `GET /2/compliance/jobs/:job_id` | 150/15m A | 5/15m A | 1/15m A |
| `POST /2/compliance/jobs` | 150/15m A | 15/15m A | 1/15m A |
| `GET /2/usage/tweets` | 50/15m A | 50/15m A | 1/15m A |
| `GET /2/trends/by/woeid/:id` | 75/15m A | 15/15m A | - |
| `GET /2/users/personalized_trends` | 10/15m U; 200/15m A | 1/15m U; 20/15m A | 1/15m U; 1/24h A |
| `GET /2/communities/:id` | 300/15m U; 300/15m A | 1/15m U; 25/15m A | 1/15m U; 1/15m A |
| `GET /2/communities/search` | 18/15m U; 300/15m A | 1/15m U; 25/15m A | 1/15m U; 1/15m A |

Sources: Rate limits table (Compliance/Usage/Trends/Communities). ([X Developer Platform][9])

## Effective cost at the cap (simple math, nothing magical)

If you actually hit caps, this is what you are paying per allowed unit. This is not what X bills you per request. It is just price divided by maximum included quota.

| Plan | $/1k reads at cap | $/1k app-level posts at cap |
| --- | ---: | ---: |
| Basic ($200/mo) | $200 / 15k = $13.33 per 1k reads | $200 / 50k = $4.00 per 1k posts |
| Pro ($5,000/mo) | $5,000 / 1,000k = $5.00 per 1k reads | $5,000 / 300k = $16.67 per 1k posts |

Caps and pricing used here come from the official plan page. ([X Developer][1])

## What's not publicly quantified (so you will not find it in the reference)

- Top-up pricing/size: the plans say "2 top ups" for Basic/Pro, but public docs do not spell out price per top-up or how many reads/writes you buy per top-up on the public pages we can access. ([X Developer][1])
- Enterprise exact quotas: public pages give "starts at $50k/month" and "50+ million Posts/month" post cap, but the actual SKUs, add-ons, and contract terms are sales-driven. ([X Developer][2])

## Bottom line

The plan names are noise. The caps and rate limits are the product. If you are building on the X API, design around the Post cap and per-endpoint limits, not the marketing tier.

If you are comparing API costs across platforms, see my breakdown of direct provider image and video generation pricing here: [/blog/direct-provider-pricing-image-video-generation-apis](/blog/direct-provider-pricing-image-video-generation-apis).

[1]: https://developer.x.com/en "https://developer.x.com/en"
[2]: https://developer.x.com/en/products/x-api/enterprise/enterprise-api-interest-form "https://developer.x.com/en/products/x-api/enterprise/enterprise-api-interest-form"
[3]: https://docs.x.com/fundamentals/developer-apps "https://docs.x.com/fundamentals/developer-apps"
[4]: https://developer.x.com/en/docs/x-api "https://developer.x.com/en/docs/x-api"
[5]: https://docs.x.com/x-api/fundamentals/post-cap "https://docs.x.com/x-api/fundamentals/post-cap"
[6]: https://docs.x.com/x-api/posts/search/quickstart/full-archive-search "https://docs.x.com/x-api/posts/search/quickstart/full-archive-search"
[7]: https://docs.x.com/x-api/getting-started/about-x-api "https://docs.x.com/x-api/getting-started/about-x-api"
[8]: https://developer.x.com/en/support/x-api/v2 "https://developer.x.com/en/support/x-api/v2"
[9]: https://docs.x.com/x-api/fundamentals/rate-limits "https://docs.x.com/x-api/fundamentals/rate-limits"


<p>You want the numbers, not the marketing. Here are the quotas, mechanics, and limits the X API actually enforces after the launch of Pay-Per-Use pricing, plus what still matters from the older Free/Basic/Pro world.</p>
<p>The punchline: pricing is no longer "pick a tier, live inside monthly caps." It's now credit-based pay-per-usage (pay only for what you consume), while rate limits still exist and still bite.</p>
<h2 id="user-content-what-changed-pay-per-use-pricing-is-now-the-default-mental-model">What changed: Pay-Per-Use pricing is now the default mental model</h2>
<h3 id="user-content-old-world-legacy-tiers">Old world (legacy tiers)</h3>
<p>You bought a monthly plan (Free/Basic/Pro) and designed around:</p>
<ul>
<li>monthly read caps</li>
<li>monthly posting caps</li>
<li>"post caps" (consumption caps)</li>
<li>top-ups (sometimes)</li>
</ul>
<h3 id="user-content-new-world-pay-per-usage-credits">New world (pay-per-usage credits)</h3>
<p>You design around:</p>
<ul>
<li>what endpoints you call</li>
<li>how much data they return</li>
<li>deduplication rules</li>
<li>your credit balance and spending limits</li>
<li>rate limits per endpoint (still real)</li>
</ul>
<p>The plan names are still noise. The billing and limits are the product.</p>
<h2 id="user-content-pay-per-use-pricing-how-it-actually-works">Pay-Per-Use pricing: how it actually works</h2>
<table>
<thead>
<tr>
<th>Mechanic</th>
<th>What it means in practice</th>
</tr>
</thead>
<tbody>
<tr>
<td>Credit-based</td>
<td>You pre-purchase credits; requests deduct from that balance.</td>
</tr>
<tr>
<td>Per-endpoint pricing</td>
<td>Different endpoints cost different amounts (the current $ rates live in the Developer Console).</td>
</tr>
<tr>
<td>No monthly caps</td>
<td>There's no built-in "15k reads and you're done." You can keep going until your credits or spending limits stop you.</td>
</tr>
<tr>
<td>Real-time tracking</td>
<td>Usage and cost tracking is surfaced in the Developer Console while you build.</td>
</tr>
<tr>
<td>Usage is tracked at the app level</td>
<td>Your bill follows your developer app usage (not a vague "project total" mental model).</td>
</tr>
<tr>
<td>Enterprise still exists</td>
<td>High-volume customers can negotiate custom pricing/support/rate limits.</td>
</tr>
</tbody>
</table>
<p>Important nuance: the docs are explicit that the current per-endpoint pricing numbers are in the Developer Console (not reliably mirrored on public pages). So if someone shows you a pricing screenshot, treat it as "may already be outdated."</p>
<h2 id="user-content-deduplication-the-one-rule-that-changes-your-caching-strategy">Deduplication: the one rule that changes your caching strategy</h2>
<p>Pay-per-usage billing comes with daily (UTC) deduplication:</p>
<ul>
<li>Billable resources are deduplicated within a 24-hour UTC day window</li>
<li>Pull the same Post again in that window: it should not incur an additional charge</li>
<li>The window resets at midnight UTC</li>
<li>X describes this as a soft guarantee (rare edge cases like outages can break it)</li>
</ul>
<p>Translated to engineering language:</p>
<ul>
<li>You still cache (latency and rate-limits), but deduplication reduces the "double charge" penalty if you re-fetch the same IDs in the same day.</li>
<li>Design pipelines to avoid re-hydrating the same IDs across days unless you need updated fields.</li>
</ul>
<h2 id="user-content-what-counts-as-billable-usage">What counts as billable usage</h2>
<p>Posts retrieved from these categories count toward usage:</p>
<table>
<thead>
<tr>
<th>Category</th>
<th>Examples</th>
</tr>
</thead>
<tbody>
<tr>
<td>Post lookup</td>
<td><code>GET /2/tweets</code></td>
</tr>
<tr>
<td>Search</td>
<td>recent search, full-archive search</td>
</tr>
<tr>
<td>Streaming</td>
<td>filtered stream (and webhooks)</td>
</tr>
<tr>
<td>Timelines</td>
<td>user posts, user mentions</td>
</tr>
<tr>
<td>Engagement</td>
<td>liked posts, bookmarks</td>
</tr>
<tr>
<td>Lists</td>
<td>list posts</td>
</tr>
<tr>
<td>Spaces</td>
<td>spaces lookup</td>
</tr>
</tbody>
</table>
<p>Two more billing rules that matter:</p>
<ul>
<li>Only successful responses that return data are billed (failed requests do not count).</li>
<li>Streaming billing is based on unique posts delivered, subject to daily deduplication.</li>
</ul>
<h2 id="user-content-controlling-spend-how-to-avoid-surprise-invoice-stories">Controlling spend: how to avoid "surprise invoice" stories</h2>
<p>Pay-per-use has fewer cliffs, but more ways to accidentally set money on fire.</p>
<h3 id="user-content-1-spending-limits-hard-stop">1) Spending limits (hard stop)</h3>
<p>You can set a max spend per billing cycle; once hit, requests are blocked until the next cycle.</p>
<h3 id="user-content-2-auto-recharge-keep-services-running">2) Auto-recharge (keep services running)</h3>
<p>You can auto-top-up your balance based on a threshold:</p>
<ul>
<li>"Trigger when balance &#x3C; $X"</li>
<li>"Recharge amount = $Y"</li>
</ul>
<h3 id="user-content-3-credit-balance-behavior-including-going-negative">3) Credit balance behavior (including going negative)</h3>
<p>X notes that your balance can go slightly negative, and your API requests will be blocked until you add credits to cover it.</p>
<h3 id="user-content-4-cache--dedup-aware-design">4) Cache + dedup-aware design</h3>
<p>Deduplication helps, but it is not magic. You still win by:</p>
<ul>
<li>caching hydrated Posts</li>
<li>storing IDs from search/stream results and avoiding re-hydration</li>
<li>keeping queries tight (especially for search)</li>
</ul>
<h2 id="user-content-monitoring-usage-programmatically">Monitoring usage programmatically</h2>
<p>You can pull usage data via the Usage endpoint (daily Post consumption counts). This is the simplest building block for alerts and budget enforcement.</p>
<p>Example call:</p>
<pre><code>curl "https://api.x.com/2/usage/tweets" \
  -H "Authorization: Bearer $BEARER_TOKEN"
</code></pre>
<p>If you are doing anything serious, wire this into:</p>
<ul>
<li>budget alerts</li>
<li>anomaly detection (spikes in search usage are common)</li>
<li>cost attribution per feature or customer</li>
</ul>
<h2 id="user-content-rate-limits-still-exist-and-are-separate-from-billing">Rate limits still exist (and are separate from billing)</h2>
<p>Pay-per-use did not remove rate limits. It just decoupled pricing from monthly caps.</p>
<p>X documents rate limits per endpoint, generally shown as:</p>
<ul>
<li>Per App</li>
<li>Per User</li>
<li>mostly per 15-minute windows (unless otherwise noted)</li>
</ul>
<p>You can also read rate-limit headers in responses:</p>
<ul>
<li><code>x-rate-limit-limit</code></li>
<li><code>x-rate-limit-remaining</code></li>
<li><code>x-rate-limit-reset</code></li>
</ul>
<h3 id="user-content-a-few-concrete-examples">A few concrete examples</h3>
<table>
<thead>
<tr>
<th>Endpoint</th>
<th align="right">Per App</th>
<th align="right">Per User</th>
<th>Notes</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>GET /2/tweets</code></td>
<td align="right">3,500 / 15 min</td>
<td align="right">5,000 / 15 min</td>
<td>Lookup (batch)</td>
</tr>
<tr>
<td><code>GET /2/tweets/:id</code></td>
<td align="right">450 / 15 min</td>
<td align="right">900 / 15 min</td>
<td>Single lookup</td>
</tr>
<tr>
<td><code>GET /2/tweets/search/recent</code></td>
<td align="right">450 / 15 min</td>
<td align="right">300 / 15 min</td>
<td>10 default, 100 max results</td>
</tr>
<tr>
<td><code>GET /2/tweets/search/all</code></td>
<td align="right">300 / 15 min + 1/sec</td>
<td align="right">1/sec</td>
<td>Full archive; larger query lengths for Enterprise</td>
</tr>
<tr>
<td><code>POST /2/tweets</code></td>
<td align="right">10,000 / 24 hrs</td>
<td align="right">100 / 15 min</td>
<td>Create posts</td>
</tr>
</tbody>
</table>
<p>Engineering takeaway: rate limits are your throughput ceiling; credits/spending limits are your budget ceiling. Design for both.</p>
<h2 id="user-content-search-access-whats-available-to-whom-and-why-it-matters">Search access: what's available to whom (and why it matters)</h2>
<p>Search has become cleaner to reason about:</p>
<ul>
<li>Recent search: last 7 days, available to all developers</li>
<li>Full-archive search: complete archive, available to Pay-per-use and Enterprise</li>
</ul>
<p>That makes "search/backfill" workloads much more about cost and rate limiting, and less about "which tier do I have?"</p>
<h2 id="user-content-enterprise-still-the-talk-to-sales-layer">Enterprise: still the "talk to sales" layer</h2>
<p>For high-volume and custom needs, X positions Enterprise as:</p>
<ul>
<li>dedicated support</li>
<li>custom rate limits</li>
<li>volume discounts or custom pricing</li>
<li>"complete data access" style options</li>
</ul>
<p>If you are at the point where you are negotiating, the public docs stop being your limiting factor anyway.</p>
<h2 id="user-content-legacy-subscription-tiers-free--basic--pro-where-they-still-show-up">Legacy subscription tiers (Free / Basic / Pro): where they still show up</h2>
<p>X calls out that legacy Basic/Pro subscription packages may still exist for some accounts, and you can:</p>
<ul>
<li>opt in to pay-per-use from the Developer Console</li>
<li>switch back to the legacy plan from settings</li>
</ul>
<p>If you are starting fresh today, assume:</p>
<ul>
<li>Pay-per-use is the default pricing path</li>
<li>legacy tiers are "legacy" for a reason (availability may vary)</li>
</ul>
<p>Practical advice: treat legacy tier quotas (reads/posts/top-ups) as historical context unless the Developer Console shows you those plans as purchasable for your account.</p>
<h2 id="user-content-faq">FAQ</h2>
<h3 id="user-content-does-x-api-still-have-monthly-caps">Does X API still have monthly caps?</h3>
<p>Pay-per-usage pricing is described as having no monthly caps; you pay based on usage and can keep going until you hit spending limits or run out of credits.</p>
<h3 id="user-content-what-happens-if-i-run-out-of-credits">What happens if I run out of credits?</h3>
<p>Requests fail until you buy more credits. Auto-recharge and alerts are how you avoid downtime.</p>
<h3 id="user-content-do-failed-requests-get-billed">Do failed requests get billed?</h3>
<p>No. Only successful responses that return data are billed.</p>
<h3 id="user-content-is-deduplication-guaranteed">Is deduplication guaranteed?</h3>
<p>X describes deduplication as a soft guarantee: it applies in most cases, but rare edge cases (like outages) can break it.</p>
<h3 id="user-content-where-do-i-find-the-actual--pricing-per-endpoint">Where do I find the actual $ pricing per endpoint?</h3>
<p>In the Developer Console. Public docs explain the model and what is billable, but current per-endpoint rates are shown in-console.</p>
<h2 id="user-content-bottom-line">Bottom line</h2>
<p>In 2026, the X API product is not "Free vs Basic vs Pro."</p>
<p>It's:</p>
<ul>
<li>pay-per-usage credits</li>
<li>billable endpoints and what they return</li>
<li>daily deduplication</li>
<li>endpoint rate limits</li>
<li>spending controls and monitoring</li>
</ul>
<p>Build around those. Everything else is branding.</p>
<table>
<thead>
<tr>
<th>Endpoint</th>
<th>Pro</th>
<th>Basic</th>
<th>Free</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>GET /2/compliance/jobs</code></td>
<td>150/15m A</td>
<td>5/15m A</td>
<td>1/15m A</td>
</tr>
<tr>
<td><code>GET /2/compliance/jobs/:job_id</code></td>
<td>150/15m A</td>
<td>5/15m A</td>
<td>1/15m A</td>
</tr>
<tr>
<td><code>POST /2/compliance/jobs</code></td>
<td>150/15m A</td>
<td>15/15m A</td>
<td>1/15m A</td>
</tr>
<tr>
<td><code>GET /2/usage/tweets</code></td>
<td>50/15m A</td>
<td>50/15m A</td>
<td>1/15m A</td>
</tr>
<tr>
<td><code>GET /2/trends/by/woeid/:id</code></td>
<td>75/15m A</td>
<td>15/15m A</td>
<td>-</td>
</tr>
<tr>
<td><code>GET /2/users/personalized_trends</code></td>
<td>10/15m U; 200/15m A</td>
<td>1/15m U; 20/15m A</td>
<td>1/15m U; 1/24h A</td>
</tr>
<tr>
<td><code>GET /2/communities/:id</code></td>
<td>300/15m U; 300/15m A</td>
<td>1/15m U; 25/15m A</td>
<td>1/15m U; 1/15m A</td>
</tr>
<tr>
<td><code>GET /2/communities/search</code></td>
<td>18/15m U; 300/15m A</td>
<td>1/15m U; 25/15m A</td>
<td>1/15m U; 1/15m A</td>
</tr>
</tbody>
</table>
<p>Sources: Rate limits table (Compliance/Usage/Trends/Communities). (<a href="https://docs.x.com/x-api/fundamentals/rate-limits" title="https://docs.x.com/x-api/fundamentals/rate-limits">X Developer Platform</a>)</p>
<h2 id="user-content-effective-cost-at-the-cap-simple-math-nothing-magical">Effective cost at the cap (simple math, nothing magical)</h2>
<p>If you actually hit caps, this is what you are paying per allowed unit. This is not what X bills you per request. It is just price divided by maximum included quota.</p>
<table>
<thead>
<tr>
<th>Plan</th>
<th align="right">$/1k reads at cap</th>
<th align="right">$/1k app-level posts at cap</th>
</tr>
</thead>
<tbody>
<tr>
<td>Basic ($200/mo)</td>
<td align="right">$200 / 15k = $13.33 per 1k reads</td>
<td align="right">$200 / 50k = $4.00 per 1k posts</td>
</tr>
<tr>
<td>Pro ($5,000/mo)</td>
<td align="right">$5,000 / 1,000k = $5.00 per 1k reads</td>
<td align="right">$5,000 / 300k = $16.67 per 1k posts</td>
</tr>
</tbody>
</table>
<p>Caps and pricing used here come from the official plan page. (<a href="https://developer.x.com/en" title="https://developer.x.com/en">X Developer</a>)</p>
<h2 id="user-content-whats-not-publicly-quantified-so-you-will-not-find-it-in-the-reference">What's not publicly quantified (so you will not find it in the reference)</h2>
<ul>
<li>Top-up pricing/size: the plans say "2 top ups" for Basic/Pro, but public docs do not spell out price per top-up or how many reads/writes you buy per top-up on the public pages we can access. (<a href="https://developer.x.com/en" title="https://developer.x.com/en">X Developer</a>)</li>
<li>Enterprise exact quotas: public pages give "starts at $50k/month" and "50+ million Posts/month" post cap, but the actual SKUs, add-ons, and contract terms are sales-driven. (<a href="https://developer.x.com/en/products/x-api/enterprise/enterprise-api-interest-form" title="https://developer.x.com/en/products/x-api/enterprise/enterprise-api-interest-form">X Developer</a>)</li>
</ul>
<h2 id="user-content-bottom-line-1">Bottom line</h2>
<p>The plan names are noise. The caps and rate limits are the product. If you are building on the X API, design around the Post cap and per-endpoint limits, not the marketing tier.</p>
<p>If you are comparing API costs across platforms, see my breakdown of direct provider image and video generation pricing here: <a href="/blog/direct-provider-pricing-image-video-generation-apis">/blog/direct-provider-pricing-image-video-generation-apis</a>.</p>


Latest X API pricing explained: pay-per-use credits, what gets billed, deduplication, rate limits, usage tracking, and how legacy Free/Basic/Pro plans fit in.

X API Pricing: Pay-Per-Use Credits + Legacy Tiers (Free, Basic, Pro, Enterprise) — What You Actually Get


It starts as a sentence you have read a thousand times -- then it starts to wobble.

```
tHiS iS wHaT iT lOoKs LiKe WhEn YoU mIx UpPeRcAsE aNd LoWeRcAsE.
```

In your head, the letters do not just change shape. The voice changes, too: childish, sing-song, needling. The message is not merely being repeated; it is being performed. And in that performance, the internet found one of its most compact tools for mockery: alternating caps, sometimes nicknamed "SpongeBob case" after the meme that made it mainstream.

This article traces the format's prehistory, the moment it exploded in 2017 as "Mocking SpongeBob," and why it continues to show up everywhere -- from political clapbacks to workplace Slack threads -- long after the image that popularized it became optional.

## Try the dumb talking text generator

If you just want to generate this style, use the dumb talking text generator below or open the full tool here: [Dumb Talking Text Generator](/tools/dumb-talking). It works as a mocking text generator, a SpongeBob text generator, a sarcastic text generator, and an upper lower case converter for quick random caps.

<div style="margin: 24px 0; border: 1px solid #e2e8f0; border-radius: 16px; overflow: hidden;">
  <iframe src="/tools/dumb-talking" title="Dumb talking text generator" loading="lazy" style="width: 100%; height: 820px; border: 0;"></iframe>
</div>

If the embed does not load, open the tool in a new tab: [Dumb Talking Text Generator](/tools/dumb-talking).
It runs locally in your browser, so the text never leaves your device.

## The internet's permanent problem: tone

Text is efficient, but it is missing the cues humans rely on: pitch, facial expression, timing, the little physical hesitations that separate a joke from an insult. Online culture has always improvised tone markers -- emoticons, emojis, GIF reactions, and shorthand like "/s" for sarcasm.

Sometimes people have tried to formalize it. In 2010, linguist bloggers at _Language Log_ noted the launch of SarcMark, a commercial "sarcasm punctuation mark." The most famous response was not a thoughtful typographic debate, but a deadpan quip John Gruber reportedly offered as "pitch-perfect assessment": "What a great idea. I'm sure it will be a huge hit."

Alternating caps is the opposite of a formal system. It is less punctuation than performance: a visual cue that tricks your brain into hearing the sentence with a particular cadence, like someone repeating your words back to you with exaggerated intonation.

That is why it spread so well. It is legible enough to understand, annoying enough to sting, and flexible enough to attach to almost any topic.

## Before SpongeBob: studlycaps and the early net

Despite its modern reputation as a "meme font," alternating case is older hacker slang. In the _Jargon File_, "studlycaps" is defined as "a hackish form of silliness similar to BiCapitalization for trademarks, but applied randomly and to arbitrary text rather than to trademarks."

Wikipedia's overview of alternating caps points to the BBS and early web eras, where weird capitalization functioned as a kind of digital swagger -- part parody of brand names with cutesy capitalization, part signal of subculture identity, and part simple experimentation with what text could do online.

This matters because it reveals something easy to miss: the visual gimmick did not start as a universal sarcasm marker. It was just a style -- one that looked loud, looked hacked, and looked intentionally incorrect. The later genius of "Mocking SpongeBob" was taking that existing typographic oddity and giving it a single, instantly recognizable emotional job: ridicule.

## The screenshot that became a voice

The canonical image behind the meme is not from a political cartoon or a viral ad. It is from a SpongeBob SquarePants episode.

As _New York Magazine_'s _Intelligencer_ put it in 2017, "Mocking SpongeBob" comes from a 2012 episode entitled "Little Yellow Book." In that episode, the plot is not about mockery at all. Squidward reads SpongeBob's diary and discovers that whenever SpongeBob sees plaid, he uncontrollably acts like a chicken -- an absurd setup that leads to the facial expression and hunched posture the internet would later interpret as pure contempt.

This is a familiar pattern in meme history: a frame from a show, originally meant to be silly or exaggerated, becomes a template for an emotion the show never explicitly named. The screenshot works because SpongeBob looks caught mid-imitation -- eyes wide, mouth twisted, body angled as if he is physically embodying someone else's argument. It is mockery as posture.

## 2017: when "Mocking SpongeBob" went mainstream

The 2012 episode sat around for years as just another SpongeBob moment until social media culture did what it does: it excavated the archive. Know Your Meme, the internet's meme database, dates the meme's breakout precisely: "The earliest iteration of the scene being used as a meme occurred on May 4th, 2017." A Twitter user posted the screenshot with a caption about staring back at kids who stare too long -- less politics, more everyday petty annoyance.

Two days later, another variation appeared that defined the format most people recognize now: a call-and-response style where a normal sentence is repeated in alternating caps to visually represent a mocking echo. Know Your Meme documents that early call-and-response variant and the way the format quickly standardized around "repeat the line, but make it look ridiculous."

The timing helped. Twitter's quote-tweet culture already trained people to respond by repeating and reframing other people's words. The SpongeBob image offered a shortcut: you did not need a paragraph to express disdain; you needed a screenshot and a keyboard workout.

Within a week, mainstream pop-culture outlets explained it to everyone else. _Teen Vogue_ emphasized the speed of the spread, noting the meme had only been listed on Know Your Meme "for approximately one day," but was already "all over social media." Meanwhile, _The Daily Dot_ summarized the core mechanic with the kind of clarity that signals a meme has escaped containment: it is SpongeBob paired with "MiXed cApiTaliZation tO inDiCatE a MoCkinG TonE."

Once a meme's format becomes explainable in a sentence, it is usually already too late. The template has escaped.

## Why alternating caps sounds like mockery in your head

A good meme does not just travel; it teaches people how to read it.

Alternating caps is effective because it disrupts your normal reading rhythm. Your eyes keep catching on the irregular capital letters, forcing micro-pauses that make the sentence feel choppy and exaggerated. The result resembles the cadence of someone mocking you by repeating your words slowly and loudly.

That intuition is backed by research from long before the meme. In a 1984 paper titled "Case alternation impairs word identification," psychologists Max Coltheart and Roger Freeman concluded that alternating case does, in fact, lead to impaired word identification, even when character size is controlled. In plain terms: the brain does not love it. Mixed-case words slow you down.

And that is part of the point. Unlike typing in all lowercase (which can signal speed or laziness), alternating caps signals intentional effort. You do not accidentally write like this. You do it because you want the reader to hear the sentence in a particular tone.

The meme is also a kind of typographic ventriloquism. It gives the writer a way to voice-act the other person's argument in text -- usually as whiny, simplistic, or self-evidently wrong. The meme does not just disagree; it caricatures.

## How the format gets used

Once alternating caps escaped its SpongeBob origin, it stopped being a meme and started being a tool. People use it in a few recurring ways.

### 1) Tone mirroring (call-and-response)

This is the format most people learned first:

- Person A: "We should all just calm down."
- Person B: "wE sHoUlD aLl JuSt CaLm DoWn."

The meaning is not subtle: Person B is implying Person A's statement is naive, hypocritical, or annoying.

In political contexts, this became a staple because it compresses criticism into a single line. A policy argument can be reframed as childish in a moment -- one reason the format got tagged early as unusually insulting compared to more neutral reaction images.

### 2) The "I'm quoting you, but meaner" screenshot

On platforms like Twitter/X, the SpongeBob image often accompanies a quote post or a screenshot of someone else's post, with the alternating-case version positioned as the response. The visual pairing matters: the image supplies the facial expression your text cannot.

Even when people do not include the SpongeBob picture, the casing alone now carries the same energy. The image has become optional because the alternating caps style now signals the tone on its own.

### 3) Self-mockery and preemptive defense

Not all uses are outwardly aggressive. People also use alternating caps to make fun of themselves:

- "Me, every Monday: i'M gOiNg To Be So PrOdUcTiVe ThIs WeEk."

Here, the writer is mocking their own optimism before reality has a chance to do it first. This self-directed version often reads as softer, closer to comedy than confrontation.

### 4) "Corporate voice" parody

Alternating caps is a fast way to puncture overly polished language:

- "We value your feedback."
- "wE vAlUe YoUr FeEdBaCk."

It is particularly popular when people feel they are being spoken to in scripts -- customer service replies, official statements, or HR announcements. The joke is not just the words; it is the implication that the words are empty.

### 5) Everyday group chat cruelty (and why it works)

In private chats, alternating caps can be brutal precisely because it feels immediate. It mimics the face-to-face experience of someone repeating your words in a mocking voice, except it is written down -- easy to screenshot, easy to forward, easy to revisit later.

That recorded mockery quality gives it power. It also gives it a cost, which leads to the next section.

## When it lands, and when it does not

Memes that encode contempt always walk a thin line. Alternating caps can be funny among friends, but it can also be a shortcut to cruelty -- especially when used as a reflex rather than a considered response.

Two practical problems show up repeatedly:

### It is harder to read

The point is to disrupt reading, but disruption does not always feel playful. Research on alternating case has found it can interfere with word identification; the same broken look that creates a mocking cadence also increases the effort required to parse the sentence.

### It can be an accessibility issue

Because the format relies on visual irregularity, it can be unfriendly to people with dyslexia or other reading difficulties. Some disability and accessibility communities also flag SpongeBob case as worth handling carefully, in part because the style is sometimes used to imitate or belittle how someone writes or speaks.

None of that means "never use it." It means the meme is not neutral. If you are writing for a broad audience -- or if you are trying to persuade rather than dunk -- alternating caps is often the wrong tool.

## The meme goes offline: from timelines to math tests

One marker of a meme's cultural saturation is when it leaves its native habitat.

In May 2017, _Teen Vogue_ reported on a Texas high school student who drew "Mocking SpongeBob" on a math test and wrote the prompt in sticky caps. The punchline: she earned extra credit. The story is funny because it treats meme literacy as a contemporary handwriting flourish -- something a teacher can recognize, understand, and reward.

That is a small story, but it captures a bigger shift. Alternating caps stopped being a niche internet in-joke and became a shared reference, legible even to people who were not online every day.

## SpongeBob as grammar

By 2019, _TIME_ described "Mocking SpongeBob" as so common it is "hard to go one day without seeing" it on Twitter, and quoted SpongeBob voice actor Tom Kenny praising meme culture as "very relatable but very good-natured." Even when the meme's intent is sharp, it rides on a larger affection for SpongeBob as a hyper-expressive language of reaction images.

At this point, the format has been absorbed into the infrastructure of the web. There are automated SpongeBob case converters that transform normal text into randomized alternating caps with one click, treating the style like a filter rather than a sentence-level joke. That is what happens when a meme becomes a dialect: it gets tools, shortcuts, and templates.

And the names keep multiplying. Depending on where you learned it, you might call it:

- alternating caps (the descriptive term)
- studlycaps (older hacker slang)
- sticky caps (a nickname emphasizing the shift-key labor)
- SpongeBob case or mocking text (the meme-derived labels)
- dumb talking text (a newer, tool-driven label)

Whatever you call it, the function is now widely understood.

## What this meme format teaches us about internet language

The "Mocking SpongeBob" style looks like nonsense capitalization, but it is actually a small invention in online rhetoric: a way to turn typography into tone.

It solved a common problem -- how to signal "I'm repeating this because it is ridiculous" -- with a visual hack anyone could produce. It also demonstrates how memes become grammar. An image from a children's cartoon did not just generate jokes; it helped standardize a new way of writing sarcasm that can survive without the original image.

So when you see tHiS tExT out in the wild, it is not just someone being weird with their keyboard. It is a shorthand for a whole social move: the internet's version of quoting someone back at them with a raised eyebrow -- and making sure everyone else can hear it too.

If you want to make your own, the [Dumb Talking Text Generator](/tools/dumb-talking) will turn any sentence into alternating caps in one click.


<p>It starts as a sentence you have read a thousand times -- then it starts to wobble.</p>
<pre><code>tHiS iS wHaT iT lOoKs LiKe WhEn YoU mIx UpPeRcAsE aNd LoWeRcAsE.
</code></pre>
<p>In your head, the letters do not just change shape. The voice changes, too: childish, sing-song, needling. The message is not merely being repeated; it is being performed. And in that performance, the internet found one of its most compact tools for mockery: alternating caps, sometimes nicknamed "SpongeBob case" after the meme that made it mainstream.</p>
<p>This article traces the format's prehistory, the moment it exploded in 2017 as "Mocking SpongeBob," and why it continues to show up everywhere -- from political clapbacks to workplace Slack threads -- long after the image that popularized it became optional.</p>
<h2 id="user-content-try-the-dumb-talking-text-generator">Try the dumb talking text generator</h2>
<p>If you just want to generate this style, use the dumb talking text generator below or open the full tool here: <a href="/tools/dumb-talking">Dumb Talking Text Generator</a>. It works as a mocking text generator, a SpongeBob text generator, a sarcastic text generator, and an upper lower case converter for quick random caps.</p>
<p>If the embed does not load, open the tool in a new tab: <a href="/tools/dumb-talking">Dumb Talking Text Generator</a>.
It runs locally in your browser, so the text never leaves your device.</p>
<h2 id="user-content-the-internets-permanent-problem-tone">The internet's permanent problem: tone</h2>
<p>Text is efficient, but it is missing the cues humans rely on: pitch, facial expression, timing, the little physical hesitations that separate a joke from an insult. Online culture has always improvised tone markers -- emoticons, emojis, GIF reactions, and shorthand like "/s" for sarcasm.</p>
<p>Sometimes people have tried to formalize it. In 2010, linguist bloggers at <em>Language Log</em> noted the launch of SarcMark, a commercial "sarcasm punctuation mark." The most famous response was not a thoughtful typographic debate, but a deadpan quip John Gruber reportedly offered as "pitch-perfect assessment": "What a great idea. I'm sure it will be a huge hit."</p>
<p>Alternating caps is the opposite of a formal system. It is less punctuation than performance: a visual cue that tricks your brain into hearing the sentence with a particular cadence, like someone repeating your words back to you with exaggerated intonation.</p>
<p>That is why it spread so well. It is legible enough to understand, annoying enough to sting, and flexible enough to attach to almost any topic.</p>
<h2 id="user-content-before-spongebob-studlycaps-and-the-early-net">Before SpongeBob: studlycaps and the early net</h2>
<p>Despite its modern reputation as a "meme font," alternating case is older hacker slang. In the <em>Jargon File</em>, "studlycaps" is defined as "a hackish form of silliness similar to BiCapitalization for trademarks, but applied randomly and to arbitrary text rather than to trademarks."</p>
<p>Wikipedia's overview of alternating caps points to the BBS and early web eras, where weird capitalization functioned as a kind of digital swagger -- part parody of brand names with cutesy capitalization, part signal of subculture identity, and part simple experimentation with what text could do online.</p>
<p>This matters because it reveals something easy to miss: the visual gimmick did not start as a universal sarcasm marker. It was just a style -- one that looked loud, looked hacked, and looked intentionally incorrect. The later genius of "Mocking SpongeBob" was taking that existing typographic oddity and giving it a single, instantly recognizable emotional job: ridicule.</p>
<h2 id="user-content-the-screenshot-that-became-a-voice">The screenshot that became a voice</h2>
<p>The canonical image behind the meme is not from a political cartoon or a viral ad. It is from a SpongeBob SquarePants episode.</p>
<p>As <em>New York Magazine</em>'s <em>Intelligencer</em> put it in 2017, "Mocking SpongeBob" comes from a 2012 episode entitled "Little Yellow Book." In that episode, the plot is not about mockery at all. Squidward reads SpongeBob's diary and discovers that whenever SpongeBob sees plaid, he uncontrollably acts like a chicken -- an absurd setup that leads to the facial expression and hunched posture the internet would later interpret as pure contempt.</p>
<p>This is a familiar pattern in meme history: a frame from a show, originally meant to be silly or exaggerated, becomes a template for an emotion the show never explicitly named. The screenshot works because SpongeBob looks caught mid-imitation -- eyes wide, mouth twisted, body angled as if he is physically embodying someone else's argument. It is mockery as posture.</p>
<h2 id="user-content-2017-when-mocking-spongebob-went-mainstream">2017: when "Mocking SpongeBob" went mainstream</h2>
<p>The 2012 episode sat around for years as just another SpongeBob moment until social media culture did what it does: it excavated the archive. Know Your Meme, the internet's meme database, dates the meme's breakout precisely: "The earliest iteration of the scene being used as a meme occurred on May 4th, 2017." A Twitter user posted the screenshot with a caption about staring back at kids who stare too long -- less politics, more everyday petty annoyance.</p>
<p>Two days later, another variation appeared that defined the format most people recognize now: a call-and-response style where a normal sentence is repeated in alternating caps to visually represent a mocking echo. Know Your Meme documents that early call-and-response variant and the way the format quickly standardized around "repeat the line, but make it look ridiculous."</p>
<p>The timing helped. Twitter's quote-tweet culture already trained people to respond by repeating and reframing other people's words. The SpongeBob image offered a shortcut: you did not need a paragraph to express disdain; you needed a screenshot and a keyboard workout.</p>
<p>Within a week, mainstream pop-culture outlets explained it to everyone else. <em>Teen Vogue</em> emphasized the speed of the spread, noting the meme had only been listed on Know Your Meme "for approximately one day," but was already "all over social media." Meanwhile, <em>The Daily Dot</em> summarized the core mechanic with the kind of clarity that signals a meme has escaped containment: it is SpongeBob paired with "MiXed cApiTaliZation tO inDiCatE a MoCkinG TonE."</p>
<p>Once a meme's format becomes explainable in a sentence, it is usually already too late. The template has escaped.</p>
<h2 id="user-content-why-alternating-caps-sounds-like-mockery-in-your-head">Why alternating caps sounds like mockery in your head</h2>
<p>A good meme does not just travel; it teaches people how to read it.</p>
<p>Alternating caps is effective because it disrupts your normal reading rhythm. Your eyes keep catching on the irregular capital letters, forcing micro-pauses that make the sentence feel choppy and exaggerated. The result resembles the cadence of someone mocking you by repeating your words slowly and loudly.</p>
<p>That intuition is backed by research from long before the meme. In a 1984 paper titled "Case alternation impairs word identification," psychologists Max Coltheart and Roger Freeman concluded that alternating case does, in fact, lead to impaired word identification, even when character size is controlled. In plain terms: the brain does not love it. Mixed-case words slow you down.</p>
<p>And that is part of the point. Unlike typing in all lowercase (which can signal speed or laziness), alternating caps signals intentional effort. You do not accidentally write like this. You do it because you want the reader to hear the sentence in a particular tone.</p>
<p>The meme is also a kind of typographic ventriloquism. It gives the writer a way to voice-act the other person's argument in text -- usually as whiny, simplistic, or self-evidently wrong. The meme does not just disagree; it caricatures.</p>
<h2 id="user-content-how-the-format-gets-used">How the format gets used</h2>
<p>Once alternating caps escaped its SpongeBob origin, it stopped being a meme and started being a tool. People use it in a few recurring ways.</p>
<h3 id="user-content-1-tone-mirroring-call-and-response">1) Tone mirroring (call-and-response)</h3>
<p>This is the format most people learned first:</p>
<ul>
<li>Person A: "We should all just calm down."</li>
<li>Person B: "wE sHoUlD aLl JuSt CaLm DoWn."</li>
</ul>
<p>The meaning is not subtle: Person B is implying Person A's statement is naive, hypocritical, or annoying.</p>
<p>In political contexts, this became a staple because it compresses criticism into a single line. A policy argument can be reframed as childish in a moment -- one reason the format got tagged early as unusually insulting compared to more neutral reaction images.</p>
<h3 id="user-content-2-the-im-quoting-you-but-meaner-screenshot">2) The "I'm quoting you, but meaner" screenshot</h3>
<p>On platforms like Twitter/X, the SpongeBob image often accompanies a quote post or a screenshot of someone else's post, with the alternating-case version positioned as the response. The visual pairing matters: the image supplies the facial expression your text cannot.</p>
<p>Even when people do not include the SpongeBob picture, the casing alone now carries the same energy. The image has become optional because the alternating caps style now signals the tone on its own.</p>
<h3 id="user-content-3-self-mockery-and-preemptive-defense">3) Self-mockery and preemptive defense</h3>
<p>Not all uses are outwardly aggressive. People also use alternating caps to make fun of themselves:</p>
<ul>
<li>"Me, every Monday: i'M gOiNg To Be So PrOdUcTiVe ThIs WeEk."</li>
</ul>
<p>Here, the writer is mocking their own optimism before reality has a chance to do it first. This self-directed version often reads as softer, closer to comedy than confrontation.</p>
<h3 id="user-content-4-corporate-voice-parody">4) "Corporate voice" parody</h3>
<p>Alternating caps is a fast way to puncture overly polished language:</p>
<ul>
<li>"We value your feedback."</li>
<li>"wE vAlUe YoUr FeEdBaCk."</li>
</ul>
<p>It is particularly popular when people feel they are being spoken to in scripts -- customer service replies, official statements, or HR announcements. The joke is not just the words; it is the implication that the words are empty.</p>
<h3 id="user-content-5-everyday-group-chat-cruelty-and-why-it-works">5) Everyday group chat cruelty (and why it works)</h3>
<p>In private chats, alternating caps can be brutal precisely because it feels immediate. It mimics the face-to-face experience of someone repeating your words in a mocking voice, except it is written down -- easy to screenshot, easy to forward, easy to revisit later.</p>
<p>That recorded mockery quality gives it power. It also gives it a cost, which leads to the next section.</p>
<h2 id="user-content-when-it-lands-and-when-it-does-not">When it lands, and when it does not</h2>
<p>Memes that encode contempt always walk a thin line. Alternating caps can be funny among friends, but it can also be a shortcut to cruelty -- especially when used as a reflex rather than a considered response.</p>
<p>Two practical problems show up repeatedly:</p>
<h3 id="user-content-it-is-harder-to-read">It is harder to read</h3>
<p>The point is to disrupt reading, but disruption does not always feel playful. Research on alternating case has found it can interfere with word identification; the same broken look that creates a mocking cadence also increases the effort required to parse the sentence.</p>
<h3 id="user-content-it-can-be-an-accessibility-issue">It can be an accessibility issue</h3>
<p>Because the format relies on visual irregularity, it can be unfriendly to people with dyslexia or other reading difficulties. Some disability and accessibility communities also flag SpongeBob case as worth handling carefully, in part because the style is sometimes used to imitate or belittle how someone writes or speaks.</p>
<p>None of that means "never use it." It means the meme is not neutral. If you are writing for a broad audience -- or if you are trying to persuade rather than dunk -- alternating caps is often the wrong tool.</p>
<h2 id="user-content-the-meme-goes-offline-from-timelines-to-math-tests">The meme goes offline: from timelines to math tests</h2>
<p>One marker of a meme's cultural saturation is when it leaves its native habitat.</p>
<p>In May 2017, <em>Teen Vogue</em> reported on a Texas high school student who drew "Mocking SpongeBob" on a math test and wrote the prompt in sticky caps. The punchline: she earned extra credit. The story is funny because it treats meme literacy as a contemporary handwriting flourish -- something a teacher can recognize, understand, and reward.</p>
<p>That is a small story, but it captures a bigger shift. Alternating caps stopped being a niche internet in-joke and became a shared reference, legible even to people who were not online every day.</p>
<h2 id="user-content-spongebob-as-grammar">SpongeBob as grammar</h2>
<p>By 2019, <em>TIME</em> described "Mocking SpongeBob" as so common it is "hard to go one day without seeing" it on Twitter, and quoted SpongeBob voice actor Tom Kenny praising meme culture as "very relatable but very good-natured." Even when the meme's intent is sharp, it rides on a larger affection for SpongeBob as a hyper-expressive language of reaction images.</p>
<p>At this point, the format has been absorbed into the infrastructure of the web. There are automated SpongeBob case converters that transform normal text into randomized alternating caps with one click, treating the style like a filter rather than a sentence-level joke. That is what happens when a meme becomes a dialect: it gets tools, shortcuts, and templates.</p>
<p>And the names keep multiplying. Depending on where you learned it, you might call it:</p>
<ul>
<li>alternating caps (the descriptive term)</li>
<li>studlycaps (older hacker slang)</li>
<li>sticky caps (a nickname emphasizing the shift-key labor)</li>
<li>SpongeBob case or mocking text (the meme-derived labels)</li>
<li>dumb talking text (a newer, tool-driven label)</li>
</ul>
<p>Whatever you call it, the function is now widely understood.</p>
<h2 id="user-content-what-this-meme-format-teaches-us-about-internet-language">What this meme format teaches us about internet language</h2>
<p>The "Mocking SpongeBob" style looks like nonsense capitalization, but it is actually a small invention in online rhetoric: a way to turn typography into tone.</p>
<p>It solved a common problem -- how to signal "I'm repeating this because it is ridiculous" -- with a visual hack anyone could produce. It also demonstrates how memes become grammar. An image from a children's cartoon did not just generate jokes; it helped standardize a new way of writing sarcasm that can survive without the original image.</p>
<p>So when you see tHiS tExT out in the wild, it is not just someone being weird with their keyboard. It is a shorthand for a whole social move: the internet's version of quoting someone back at them with a raised eyebrow -- and making sure everyone else can hear it too.</p>
<p>If you want to make your own, the <a href="/tools/dumb-talking">Dumb Talking Text Generator</a> will turn any sentence into alternating caps in one click.</p>


A short history of alternating caps, studlycaps, and SpongeBob case, plus why mocking text generators and sarcastic text still shape tone online.

From StudlyCaps to SpongeBob: The Surprising History of tHiS wEiRd, MoCkInG tExT


* * *

When people talk about "great writers," they usually imagine a lone figure bent over a desk, ink on their fingers, filling page after page in silence.

For much of history, that picture is wrong.

From Roman statesmen and biblical authors to modernist novelists and wartime leaders, countless "writers" did not physically write their words. They spoke. Highly trained scribes, secretaries, and assistants did the actual writing, copying, and typing.

The result: the world sees a single famous name on the cover, but behind that name stands an entire hidden workflow of dictation, shorthand, and quiet expertise.

This article walks through 12 striking examples of legendary figures who built their reputations on words they did not physically write, and what that reveals about creativity and work.

* * *

## How Dictation Really Worked

Before looking at individual cases, it helps to understand the basic pattern that repeats across centuries.

In most of these stories there are two key roles:

- **The originator**  
  The ruler, thinker, or author whose ideas and wording drive the text.
- **The technician**  
  A scribe, secretary, or amanuensis trained to capture speech quickly and turn it into polished documents.

The process often followed a predictable loop:

1. The originator dictates a rough draft.
2. The assistant takes it down (in shorthand, neat longhand, or later, on a typewriter).
3. A fair copy is prepared.
4. The originator reviews and corrects it, sometimes by speaking the changes.
5. Copies are made and circulated.

From the outside, letters and books produced this way look like the work of one person. From the inside, they are the product of a tight collaboration between the person who speaks and the person who writes.

* * *

## 1. Cicero and Tiro: Rome's Hidden Writing Machine

Marcus Tullius Cicero, the Roman orator and statesman, is famous for his speeches and his enormous collection of letters. He did not personally write all that text.

Working in the background was Tiro, a slave who later became a freedman and trusted secretary. Tiro developed an early system of shorthand specifically to keep up with Cicero's rapid speech. He took down drafts, prepared copies, and helped preserve Cicero's correspondence after his death.

When people read "Cicero's letters" today, they are seeing the result of a partnership: Cicero's mind and voice, Tiro's pen and system.

* * *

## 2. Paul the Apostle and the Invisible Secretaries

Paul of Tarsus is known above all as a letter-writer. His messages to early Christian communities form a major part of the New Testament.

Yet at least one of these letters openly reveals how it was produced. Near the end of his Letter to the Romans, a new voice briefly appears: "I, Tertius, who wrote this letter, greet you." Paul dictated; Tertius wrote.

Elsewhere Paul signs off with a short line "in my own hand," implying that the main body of the letter was written by someone else. It was normal in that environment to hire or rely on a professional scribe for the main text, while the "author" supplied the content and authority.

In effect, Paul was the author of the message; the secretary was the specialist who turned that message into written form.

* * *

## 3. Jeremiah and Baruch: Dictation Made Explicit

In the book of Jeremiah, the process is explained step by step.

Jeremiah receives a message and dictates it to Baruch, his scribe. Baruch writes the words on a scroll. When that scroll is destroyed, Jeremiah dictates the material again and Baruch rewrites it, with additions.

Here, the text itself insists that the words belong to Jeremiah, while also giving full visibility to Baruch's role. The prophet is the voice; the scribe is the hand.

It is one of the clearest early descriptions of a dictation workflow in action.

* * *

## 4. Elizabeth I: Royal Letters as a Collaborative Product

Queen Elizabeth I of England was a serious reader and writer in her own right. She composed speeches, translated texts, and could write in several languages.

That did not mean she personally wrote every royal letter bearing her name.

Day-to-day, drafts of diplomatic and administrative correspondence were prepared by clerks and secretaries. Many letters were dictated, refined in conversation, and then copied carefully into fair versions for dispatch. Elizabeth might add final notes in her own hand or change key phrases, but most of the physical writing was delegated.

What the world saw as "Elizabeth's letters" were, in practice, the result of a small writing office shaped around her decisions and style.

* * *

## 5. John Milton: Epic Poetry in the Dark

John Milton, the author of _Paradise Lost_, gives a dramatic example of dictation as a response to disability.

By his mid-forties he was completely blind. Handwriting was no longer possible, but he continued to produce dense prose and complex poetry.

Milton dictated his work to daughters and hired amanuenses. Accounts describe him composing lines in his head at night, memorising them, and then reciting them the next day for someone else to write down.

The result is one of the great epics of English literature, created through memory, voice, and careful listening rather than through the writer's own hand.

* * *

## 6. Voltaire: A One-Man Studio with Many Pens

Voltaire published plays, novels, histories, philosophical treatises, pamphlets, and thousands of letters. His output looks almost impossible for a single person until you factor in dictation.

Visitors and secretaries reported that Voltaire often worked from his bed, dictating to assistants. Large portions of his surviving manuscripts are in the handwriting of those secretaries, not his own.

He treated writing like a studio operation. He supplied the ideas, arguments, and sentences; his staff captured them, copied them, and prepared them for printers and correspondents.

* * *

## 7. Stendhal: A Novel Produced at Spoken Speed

Henri Beyle, better known as Stendhal, is remembered for psychological novels like _The Red and the Black_ and _The Charterhouse of Parma_.

The latter was created at extraordinary speed. Stendhal is reported to have dictated it to a secretary over a matter of weeks, working at spoken pace. He later revised, but the core of the book came out of this intense dictation sprint.

That method left its mark on the style. _The Charterhouse of Parma_ has the direct, conversational feel of a story told aloud rather than a text polished slowly on the page.

* * *

## 8. Fyodor Dostoevsky: Dictation Under Deadline

Dostoevsky, faced with a contract that demanded a new work by a fixed date or he would lose rights to his earlier books, needed to move fast.

He hired a young stenographer, Anna Grigorievna Snitkina, and dictated a new novella, _The Gambler_, to her at speed. She took down his words, prepared the manuscript, and they met the deadline.

The collaboration went beyond work: Anna later became his wife. But professionally, this episode shows dictation as a practical solution to an almost impossible time constraint.

* * *

## 9. Henry James: The Age of the Dictated Novel

In his later career, Henry James struggled with hand pain and the sheer effort of writing. Rather than slow down, he changed methods.

James hired typists and began composing by dictating. He would pace the room, shaping elaborate sentences aloud while his secretary captured them on a typewriter. They then worked together through rounds of correction.

Many of his late, complex novels (_The Wings of the Dove_, _The Ambassadors_, _The Golden Bowl_) were written in this way. The celebrated intricacy of his style was born as spoken language before it became print.

* * *

## 10. James Joyce: Experimental Prose with Help

James Joyce, author of _Ulysses_ and _Finnegans Wake_, suffered from serious and recurring eye problems. At times he could barely see well enough to read.

Friends and family helped by reading to him and taking down his words. Portions of _Finnegans Wake_ were dictated rather than written by his own hand. He would listen to drafts, speak revisions, and rely on those around him to make the changes on paper.

The creative process behind his famously challenging prose was therefore shared: Joyce's imagination and voice working through the patient labour of others at the page.

* * *

## 11. Napoleon Bonaparte: Orders Spoken at Speed

Napoleon's rule generated a vast stream of paperwork: orders to generals, memoranda to ministers, decrees for civil administration, and public bulletins.

Much of it began as dictation.

Secretaries followed him, sometimes at all hours, taking down instructions as he walked, travelled, or ate. They turned his spoken orders into written documents in the correct diplomatic and bureaucratic forms, then supervised the copying and dispatch.

Napoleon used dictation to extend his reach. His voice could become dozens of letters and orders in a single night, each one given physical form by his writing staff.

* * *

## 12. Winston Churchill: Words for War and Peace

Winston Churchill's reputation rests heavily on his words: wartime speeches, newspaper articles, memoirs, and multi-volume histories.

He almost never produced them by simply sitting alone with pen and paper.

Instead, Churchill surrounded himself with teams of secretaries. He dictated letters, memos, chapters, and speeches. They typed them up; he revised; they prepared new copies. This loop repeated until he was satisfied.

His famous addresses during the Second World War followed this pattern: dictated drafts, typed versions marked with changes, and final copies formatted to match his speaking rhythm. The voice the public heard was his; the text they later read was the product of his voice channelled through a skilled support team.

* * *

## Did They "Really" Write These Works?

Once you know how much dictation lies behind these texts, a natural question arises: if someone dictates a work and never touches the pen, are they still the writer?

Traditionally, the answer has been yes.

Authorship has been defined by:

- Who originates and chooses the ideas.
- Who shapes the wording, whether spoken or written.
- Who takes responsibility for what the text says.

The scribes and secretaries in these stories were experts in capturing, copying, and smoothing the language. Unless they significantly altered the content, they were not credited as co-authors.

Even so, seeing how these relationships worked makes it clear that many "solo" achievements were in fact shared efforts. A distinctive voice, supported by someone who knew how to catch every word.

* * *

## What This Reveals About Writing and Work

Looking across these 12 figures, a few patterns stand out.

### 1. Writing is more than handwriting

Cicero, Milton, Joyce, and Churchill remind us that "writing" is a larger process than just putting words on a page. It includes thinking, planning, experimenting, speaking, revising, and organising.

The physical act of forming letters can be delegated without giving up control of the message.

### 2. Dictation is a tool for overcoming limits

Some of these people dictated because they had no choice: blindness, eye disease, hand pain, or crushing deadlines. Dictation let them work around those constraints instead of being defined by them.

The modern counterpart is obvious: voice input, transcription tools, and assistants can keep projects moving even when typing is slow or impossible.

### 3. Collaboration does not erase individual voice

Tiro's shorthand did not turn Cicero into someone else. Churchill's secretaries did not make his speeches sound like theirs. The role of the assistant was to capture the speaker's voice accurately, not to replace it.

That is why history still remembers these works under the names of their dictating authors, even though unseen hands produced the first written versions.

* * *

The image of the solitary writer, alone with pen and paper, is powerful but incomplete. For many of the most influential texts in history, the real story is a partnership: one person thinking and speaking at a high level, another person quietly turning those spoken words into durable, sharable writing.

If you want a modern, practical take on writing workflows, read [How Paul Graham Actually Writes Essays](/blog/how-paul-graham-writes-essays).


<hr>
<p>When people talk about "great writers," they usually imagine a lone figure bent over a desk, ink on their fingers, filling page after page in silence.</p>
<p>For much of history, that picture is wrong.</p>
<p>From Roman statesmen and biblical authors to modernist novelists and wartime leaders, countless "writers" did not physically write their words. They spoke. Highly trained scribes, secretaries, and assistants did the actual writing, copying, and typing.</p>
<p>The result: the world sees a single famous name on the cover, but behind that name stands an entire hidden workflow of dictation, shorthand, and quiet expertise.</p>
<p>This article walks through 12 striking examples of legendary figures who built their reputations on words they did not physically write, and what that reveals about creativity and work.</p>
<hr>
<h2 id="user-content-how-dictation-really-worked">How Dictation Really Worked</h2>
<p>Before looking at individual cases, it helps to understand the basic pattern that repeats across centuries.</p>
<p>In most of these stories there are two key roles:</p>
<ul>
<li><strong>The originator</strong><br>
The ruler, thinker, or author whose ideas and wording drive the text.</li>
<li><strong>The technician</strong><br>
A scribe, secretary, or amanuensis trained to capture speech quickly and turn it into polished documents.</li>
</ul>
<p>The process often followed a predictable loop:</p>
<ol>
<li>The originator dictates a rough draft.</li>
<li>The assistant takes it down (in shorthand, neat longhand, or later, on a typewriter).</li>
<li>A fair copy is prepared.</li>
<li>The originator reviews and corrects it, sometimes by speaking the changes.</li>
<li>Copies are made and circulated.</li>
</ol>
<p>From the outside, letters and books produced this way look like the work of one person. From the inside, they are the product of a tight collaboration between the person who speaks and the person who writes.</p>
<hr>
<h2 id="user-content-1-cicero-and-tiro-romes-hidden-writing-machine">1. Cicero and Tiro: Rome's Hidden Writing Machine</h2>
<p>Marcus Tullius Cicero, the Roman orator and statesman, is famous for his speeches and his enormous collection of letters. He did not personally write all that text.</p>
<p>Working in the background was Tiro, a slave who later became a freedman and trusted secretary. Tiro developed an early system of shorthand specifically to keep up with Cicero's rapid speech. He took down drafts, prepared copies, and helped preserve Cicero's correspondence after his death.</p>
<p>When people read "Cicero's letters" today, they are seeing the result of a partnership: Cicero's mind and voice, Tiro's pen and system.</p>
<hr>
<h2 id="user-content-2-paul-the-apostle-and-the-invisible-secretaries">2. Paul the Apostle and the Invisible Secretaries</h2>
<p>Paul of Tarsus is known above all as a letter-writer. His messages to early Christian communities form a major part of the New Testament.</p>
<p>Yet at least one of these letters openly reveals how it was produced. Near the end of his Letter to the Romans, a new voice briefly appears: "I, Tertius, who wrote this letter, greet you." Paul dictated; Tertius wrote.</p>
<p>Elsewhere Paul signs off with a short line "in my own hand," implying that the main body of the letter was written by someone else. It was normal in that environment to hire or rely on a professional scribe for the main text, while the "author" supplied the content and authority.</p>
<p>In effect, Paul was the author of the message; the secretary was the specialist who turned that message into written form.</p>
<hr>
<h2 id="user-content-3-jeremiah-and-baruch-dictation-made-explicit">3. Jeremiah and Baruch: Dictation Made Explicit</h2>
<p>In the book of Jeremiah, the process is explained step by step.</p>
<p>Jeremiah receives a message and dictates it to Baruch, his scribe. Baruch writes the words on a scroll. When that scroll is destroyed, Jeremiah dictates the material again and Baruch rewrites it, with additions.</p>
<p>Here, the text itself insists that the words belong to Jeremiah, while also giving full visibility to Baruch's role. The prophet is the voice; the scribe is the hand.</p>
<p>It is one of the clearest early descriptions of a dictation workflow in action.</p>
<hr>
<h2 id="user-content-4-elizabeth-i-royal-letters-as-a-collaborative-product">4. Elizabeth I: Royal Letters as a Collaborative Product</h2>
<p>Queen Elizabeth I of England was a serious reader and writer in her own right. She composed speeches, translated texts, and could write in several languages.</p>
<p>That did not mean she personally wrote every royal letter bearing her name.</p>
<p>Day-to-day, drafts of diplomatic and administrative correspondence were prepared by clerks and secretaries. Many letters were dictated, refined in conversation, and then copied carefully into fair versions for dispatch. Elizabeth might add final notes in her own hand or change key phrases, but most of the physical writing was delegated.</p>
<p>What the world saw as "Elizabeth's letters" were, in practice, the result of a small writing office shaped around her decisions and style.</p>
<hr>
<h2 id="user-content-5-john-milton-epic-poetry-in-the-dark">5. John Milton: Epic Poetry in the Dark</h2>
<p>John Milton, the author of <em>Paradise Lost</em>, gives a dramatic example of dictation as a response to disability.</p>
<p>By his mid-forties he was completely blind. Handwriting was no longer possible, but he continued to produce dense prose and complex poetry.</p>
<p>Milton dictated his work to daughters and hired amanuenses. Accounts describe him composing lines in his head at night, memorising them, and then reciting them the next day for someone else to write down.</p>
<p>The result is one of the great epics of English literature, created through memory, voice, and careful listening rather than through the writer's own hand.</p>
<hr>
<h2 id="user-content-6-voltaire-a-one-man-studio-with-many-pens">6. Voltaire: A One-Man Studio with Many Pens</h2>
<p>Voltaire published plays, novels, histories, philosophical treatises, pamphlets, and thousands of letters. His output looks almost impossible for a single person until you factor in dictation.</p>
<p>Visitors and secretaries reported that Voltaire often worked from his bed, dictating to assistants. Large portions of his surviving manuscripts are in the handwriting of those secretaries, not his own.</p>
<p>He treated writing like a studio operation. He supplied the ideas, arguments, and sentences; his staff captured them, copied them, and prepared them for printers and correspondents.</p>
<hr>
<h2 id="user-content-7-stendhal-a-novel-produced-at-spoken-speed">7. Stendhal: A Novel Produced at Spoken Speed</h2>
<p>Henri Beyle, better known as Stendhal, is remembered for psychological novels like <em>The Red and the Black</em> and <em>The Charterhouse of Parma</em>.</p>
<p>The latter was created at extraordinary speed. Stendhal is reported to have dictated it to a secretary over a matter of weeks, working at spoken pace. He later revised, but the core of the book came out of this intense dictation sprint.</p>
<p>That method left its mark on the style. <em>The Charterhouse of Parma</em> has the direct, conversational feel of a story told aloud rather than a text polished slowly on the page.</p>
<hr>
<h2 id="user-content-8-fyodor-dostoevsky-dictation-under-deadline">8. Fyodor Dostoevsky: Dictation Under Deadline</h2>
<p>Dostoevsky, faced with a contract that demanded a new work by a fixed date or he would lose rights to his earlier books, needed to move fast.</p>
<p>He hired a young stenographer, Anna Grigorievna Snitkina, and dictated a new novella, <em>The Gambler</em>, to her at speed. She took down his words, prepared the manuscript, and they met the deadline.</p>
<p>The collaboration went beyond work: Anna later became his wife. But professionally, this episode shows dictation as a practical solution to an almost impossible time constraint.</p>
<hr>
<h2 id="user-content-9-henry-james-the-age-of-the-dictated-novel">9. Henry James: The Age of the Dictated Novel</h2>
<p>In his later career, Henry James struggled with hand pain and the sheer effort of writing. Rather than slow down, he changed methods.</p>
<p>James hired typists and began composing by dictating. He would pace the room, shaping elaborate sentences aloud while his secretary captured them on a typewriter. They then worked together through rounds of correction.</p>
<p>Many of his late, complex novels (<em>The Wings of the Dove</em>, <em>The Ambassadors</em>, <em>The Golden Bowl</em>) were written in this way. The celebrated intricacy of his style was born as spoken language before it became print.</p>
<hr>
<h2 id="user-content-10-james-joyce-experimental-prose-with-help">10. James Joyce: Experimental Prose with Help</h2>
<p>James Joyce, author of <em>Ulysses</em> and <em>Finnegans Wake</em>, suffered from serious and recurring eye problems. At times he could barely see well enough to read.</p>
<p>Friends and family helped by reading to him and taking down his words. Portions of <em>Finnegans Wake</em> were dictated rather than written by his own hand. He would listen to drafts, speak revisions, and rely on those around him to make the changes on paper.</p>
<p>The creative process behind his famously challenging prose was therefore shared: Joyce's imagination and voice working through the patient labour of others at the page.</p>
<hr>
<h2 id="user-content-11-napoleon-bonaparte-orders-spoken-at-speed">11. Napoleon Bonaparte: Orders Spoken at Speed</h2>
<p>Napoleon's rule generated a vast stream of paperwork: orders to generals, memoranda to ministers, decrees for civil administration, and public bulletins.</p>
<p>Much of it began as dictation.</p>
<p>Secretaries followed him, sometimes at all hours, taking down instructions as he walked, travelled, or ate. They turned his spoken orders into written documents in the correct diplomatic and bureaucratic forms, then supervised the copying and dispatch.</p>
<p>Napoleon used dictation to extend his reach. His voice could become dozens of letters and orders in a single night, each one given physical form by his writing staff.</p>
<hr>
<h2 id="user-content-12-winston-churchill-words-for-war-and-peace">12. Winston Churchill: Words for War and Peace</h2>
<p>Winston Churchill's reputation rests heavily on his words: wartime speeches, newspaper articles, memoirs, and multi-volume histories.</p>
<p>He almost never produced them by simply sitting alone with pen and paper.</p>
<p>Instead, Churchill surrounded himself with teams of secretaries. He dictated letters, memos, chapters, and speeches. They typed them up; he revised; they prepared new copies. This loop repeated until he was satisfied.</p>
<p>His famous addresses during the Second World War followed this pattern: dictated drafts, typed versions marked with changes, and final copies formatted to match his speaking rhythm. The voice the public heard was his; the text they later read was the product of his voice channelled through a skilled support team.</p>
<hr>
<h2 id="user-content-did-they-really-write-these-works">Did They "Really" Write These Works?</h2>
<p>Once you know how much dictation lies behind these texts, a natural question arises: if someone dictates a work and never touches the pen, are they still the writer?</p>
<p>Traditionally, the answer has been yes.</p>
<p>Authorship has been defined by:</p>
<ul>
<li>Who originates and chooses the ideas.</li>
<li>Who shapes the wording, whether spoken or written.</li>
<li>Who takes responsibility for what the text says.</li>
</ul>
<p>The scribes and secretaries in these stories were experts in capturing, copying, and smoothing the language. Unless they significantly altered the content, they were not credited as co-authors.</p>
<p>Even so, seeing how these relationships worked makes it clear that many "solo" achievements were in fact shared efforts. A distinctive voice, supported by someone who knew how to catch every word.</p>
<hr>
<h2 id="user-content-what-this-reveals-about-writing-and-work">What This Reveals About Writing and Work</h2>
<p>Looking across these 12 figures, a few patterns stand out.</p>
<h3 id="user-content-1-writing-is-more-than-handwriting">1. Writing is more than handwriting</h3>
<p>Cicero, Milton, Joyce, and Churchill remind us that "writing" is a larger process than just putting words on a page. It includes thinking, planning, experimenting, speaking, revising, and organising.</p>
<p>The physical act of forming letters can be delegated without giving up control of the message.</p>
<h3 id="user-content-2-dictation-is-a-tool-for-overcoming-limits">2. Dictation is a tool for overcoming limits</h3>
<p>Some of these people dictated because they had no choice: blindness, eye disease, hand pain, or crushing deadlines. Dictation let them work around those constraints instead of being defined by them.</p>
<p>The modern counterpart is obvious: voice input, transcription tools, and assistants can keep projects moving even when typing is slow or impossible.</p>
<h3 id="user-content-3-collaboration-does-not-erase-individual-voice">3. Collaboration does not erase individual voice</h3>
<p>Tiro's shorthand did not turn Cicero into someone else. Churchill's secretaries did not make his speeches sound like theirs. The role of the assistant was to capture the speaker's voice accurately, not to replace it.</p>
<p>That is why history still remembers these works under the names of their dictating authors, even though unseen hands produced the first written versions.</p>
<hr>
<p>The image of the solitary writer, alone with pen and paper, is powerful but incomplete. For many of the most influential texts in history, the real story is a partnership: one person thinking and speaking at a high level, another person quietly turning those spoken words into durable, sharable writing.</p>
<p>If you want a modern, practical take on writing workflows, read <a href="/blog/how-paul-graham-writes-essays">How Paul Graham Actually Writes Essays</a>.</p>


Many iconic letters, books, and speeches were not handwritten by their authors. Discover 12 famous figures who created their work by dictating it to others.

12 Legendary "Writers" Who Never Touched the Pen: The Secret History of Dictated Genius


---

## 1. What actually happened: Nov 19-21, 2025 in three days

In the three trading days from Wednesday 19 November through Friday 21 November 2025, Nvidia went from a textbook post-earnings celebration to what felt, on screens, like a near-crash. Record results and raised guidance sent NVDA jumping at first, but by Friday’s close the stock had swung almost 12% from peak to trough — enough to make traders wonder if they had just watched a clean-up of crowded AI positioning or something darker. ([NVIDIA Newsroom][1])

Here’s the three-day picture (US Eastern Time, regular sessions only):

| Date (2025)  | Open   | High   | Low    | Close      | Day move | Notes                          |
| ------------ | ------ | ------ | ------ | ---------- | -------- | ------------------------------ |
| Nov 19 (Wed) | 184.79 | 187.86 | 182.83 | **186.52** | +2.9%    | Earnings after the close       |
| Nov 20 (Thu) | 195.95 | 196.00 | 179.85 | **180.64** | -3.1%    | Huge gap up, then big reversal |
| Nov 21 (Fri) | 181.24 | 184.56 | 172.93 | **178.88** | -1.0%    | Volatile, closes slightly red  |

Key points:

* On **Nov 19**, NVDA closed at **$186.52**, up ~2.9% after trading between roughly **$182.8 and $187.9**.
* On **Nov 20**, it **opened $195.95**, **ticked $196.00**, then slid to **$179.85** before closing **$180.64** (about **-3.1%** vs Wednesday’s close but more than **8%** off the intraday high).
* On **Nov 21**, it opened at **$181.24**, ranged **$184.56 to $172.93**, and finished **$178.88** (about **-1%** on the day and **~11.8%** below Thursday’s high).
* From Thursday’s **$196.00** high to Friday’s **$172.93** low is almost a **12%** swing.
* By **Monday 24 November**, NVDA was back in the low **$180s**, still well below the late-October record close around **$207**. ([MacroTrends][3])

### Snapshot: NVDA on Monday, 24 November (intraday)

* Latest price **$182.55**, change **$4.00 (0.02%)** vs previous close.
* Open **$179.42**; high **$183.41**; low **$176.60**; volume **253,434,849**.
* Latest trade time **15:11:44 PST**.

The tape looks wild, but the news flow was overwhelmingly positive. That gap between “great fundamentals” and “bad price action” is what fueled the “rigged” and “front-running the jobs report” chatter.

---

## 2. Earnings were exceptional, not the problem

After the close on **Wednesday, 19 November**, Nvidia reported:

* **$57.0B** in quarterly revenue for Q3 FY26 (ended 26 October 2025), up **22% QoQ** and **62% YoY**.
* **$51.2B** in data-center revenue, up **66% YoY**.
* Guidance for roughly **$65B** next quarter.
* Management reiterated AI accelerator demand was **“off the charts.”** ([NVIDIA Newsroom][1])

Coverage from Reuters, TechCrunch, the Guardian, and others framed it as another emphatic beat that briefly eased AI-bubble jitters and sparked a global relief rally. The earnings release **added** to bullish sentiment; it didn’t cap it. ([The Guardian][7])

---

## 3. The macro backdrop: a delayed jobs report hits with earnings

All times below are **US Eastern Time (ET)**.

### Thursday, 20 November 2025

Because a 43-day federal shutdown delayed official statistics, the **September Employment Situation** report landed on **Thursday 20 November at 08:30 a.m. ET**, alongside weekly jobless claims — an unusual overlap with Nvidia’s earnings glow. ([Bureau of Labor Statistics][5])

Headline numbers:

* Payrolls: **+119k** vs expectations closer to **+50k**.
* Unemployment rate: **4.4%**, the highest in several years.
* Prior months **revised down**.

### Friday, 21 November 2025

* **08:30 a.m. ET** - BLS releases **September Real Earnings**.

### Crucial timing

* **All the macro prints hit at 08:30 a.m. ET.**
* The **US cash equity session opens at 09:30 a.m. ET**.
* On both days, the macro data were public for **an hour** before regular-session trading in NVDA began.

---

## 4. The sequence: earnings pop, jobs data, then a positioning unwind

### Wednesday evening into Thursday pre-market

* After hours on **Nov 19**, NVDA jumped about **5%**, adding well over **$200B** in market value within minutes. ([Reuters][4])
* Between **8–9 a.m. ET** on **Nov 20**, pre-market updates repeatedly showed NVDA **up 5%+**, dragging Nasdaq futures higher (Nasdaq futures **>+1.5%**). ([Reuters][4])
* **08:30 a.m. ET** — Jobs report and claims drop. Markets digest a “no quick rate cuts, but no collapse either” narrative. Nvidia remains elevated into the open; there is **no evidence of a dump before 8:30**. ([Bureau of Labor Statistics][5])

---

### Thursday cash session (20 Nov)

* **09:30 a.m. ET** — NVDA opens **$195.95**, ticks **$196.00** near the open. Indices gap up hard: Dow +700, Nasdaq +2.5% within minutes. ([Investopedia][6])
* **Late morning (≈10:30–10:45 a.m.)** — Nvidia and the indices **roll over**. The S&P 500 and Nasdaq slide from early gains of ~2% to losses over 1.5%, a rare intraday reversal more typical of stress days.
* **Low:** NVDA hits **$179.85**, an **~8.2%** drop from the **$196** high.
* **Close:** **$180.64**, down **~3.1%** on the day and **~7.8%** below the pre-market high. ([StockAnalysis][2])

Market color during the slide focused on profit-taking in a crowded AI leader, options-driven hedging (dealers short gamma selling into weakness), and a reassessment of how much to pay for a stock that had already doubled on the AI story. There were **no major new headlines** driving the 10:30–4:00 tape.

---

### Friday cash session (21 Nov)

* **Open:** **$181.24**, a touch above Thursday’s close.
* **Range:** **$184.56** high to **$172.93** low.
* **Close:** **$178.88**, about **-1%** on the day and **~11.8%** below Thursday’s high. ([StockAnalysis][2])

The broader market stabilized and even finished green, helped late by Fed commentary. Nvidia stayed weak, pointing to stock-specific positioning rather than fresh macro shock.

---

## 5. Why people think “it started before the jobs report”

Two mundane issues create the illusion of pre-data selling:

1) **Time zones.** Economic calendars quote **08:30 a.m. ET**; many charts default to **UTC**. **08:30 ET = 13:30 UTC.** If you compare a UTC-stamped calendar to an ET-stamped chart, it can look like the stock moved first when you are staring at the same moment in two clocks. ([Bureau of Labor Statistics][5])

2) **Session coverage.** Many retail charts show only the **9:30–16:00 ET** cash session. The first reactions to both the earnings release (after hours) and the jobs data (pre-market) happen **before** those charts even draw a candle. By 9:30, the info is already reflected; what you see next is the positioning unwind.

3) **Instrument mix.** Futures and FX trade almost 24/5; single-stock pre-market liquidity is thinner and cash opens later. Overlaying futures with cash equities without adjusting time windows can create false leads/lags.

Once you align clocks and sessions, the sequence looks straightforward: **earnings pop → jobs data digested → late-morning unwind in a crowded trade.**

---

## 6. What actually drove the reversal

Beyond the calendar mechanics, the drivers were classic:

* **Crowded positioning.** Nvidia was the flagship AI trade, the world’s most valuable listed company, and a core holding in AI-focused ETFs. Even “great” numbers can trigger “good enough to take profits.”
* **Options and systematic flows.** Elevated implied volatility around earnings + macro means dealers hedging short gamma can sell stock as it drops, reinforcing the move. Stop-loss orders and CTA/vol-targeting rules can add momentum.
* **Valuation check.** After doubling (and then some) on the AI story, investors reassessed how much they would pay in a session that started euphoric and flipped to risk-off.
* **No fresh negative headline.** Media coverage at the time explicitly noted the reversal came with “no major headlines” — it was flow- and positioning-driven. ([Investopedia][6])

---

## 7. Is there evidence of manipulation?

To move from “ugly tape” to “illegal manipulation,” you’d expect specifics: non-public access to the jobs report before 8:30, documented spoofing/layering patterns, or regulator/exchange notices flagging abnormal NVDA trading. ([Bureau of Labor Statistics][5])

Based on public information, none of those exist for **19–21 November 2025**, despite intense media attention. Major outlets explained the move with standard factors: positioning, valuations, AI-bubble concerns, and options hedging. ([Reuters][4])

**Conclusion:** The price path is consistent with a crowded trade being unwound into an earnings + macro window, not a smoking gun of collusion.

---

## 8. Practical takeaways

* **Don’t anchor on the close.** “Nvidia down 3%” misses an **8% intraday swing** and an **~12% peak-to-trough** over two sessions.
* **Great fundamentals can still be “sell the news.”** When expectations and ownership are stretched, excellent prints can be liquidity events for de-risking.
* **Align clocks and sessions.** Check time zones and whether charts include pre-/post-market before inferring causality.
* **Expect big moves around stacked catalysts.** Earnings + delayed macro + crowded positioning + elevated vol is a recipe for two-sided 5–10% swings, even in mega-caps.
* **Assume flow before conspiracy.** In the absence of regulatory flags, the base case is positioning and hedging mechanics.

If you like data-driven breakdowns like this, explore the rest of the [blog](/blog).

---

*Standard disclaimer: this article is for informational and educational purposes only and is not investment advice or a recommendation to buy or sell any security.*

[1]: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2026 "NVIDIA Announces Financial Results for Third Quarter ..."
[2]: https://stockanalysis.com/stocks/nvda/history/ "NVIDIA (NVDA) Stock Price History 1999-2025"
[3]: https://www.macrotrends.net/stocks/charts/NVDA/nvidia/stock-price-history "NVIDIA - 26 Year Stock Price History | NVDA"
[4]: https://www.reuters.com/world/china/global-markets-trading-day-graphic-2025-11-20/ "Trading Day: Nvidia sparkle vanishes, Wall Street slammed"
[5]: https://www.bls.gov/news.release/empsit.nr0.htm "Employment Situation Summary - 2025 M09 Results"
[6]: https://www.investopedia.com/dow-jones-today-11202025-11853411 "Markets News, Nov. 20, 2025: Major Stock Indexes Post ..."
[7]: https://www.theguardian.com/technology/2025/nov/19/nvidia-earning-report "'We excel at every phase of AI': Nvidia CEO quells Wall Street fears of AI bubble amid market selloff"


<hr>
<h2 id="user-content-1-what-actually-happened-nov-19-21-2025-in-three-days">1. What actually happened: Nov 19-21, 2025 in three days</h2>
<p>In the three trading days from Wednesday 19 November through Friday 21 November 2025, Nvidia went from a textbook post-earnings celebration to what felt, on screens, like a near-crash. Record results and raised guidance sent NVDA jumping at first, but by Friday’s close the stock had swung almost 12% from peak to trough — enough to make traders wonder if they had just watched a clean-up of crowded AI positioning or something darker. (<a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2026" title="NVIDIA Announces Financial Results for Third Quarter ...">NVIDIA Newsroom</a>)</p>
<p>Here’s the three-day picture (US Eastern Time, regular sessions only):</p>
<table>
<thead>
<tr>
<th>Date (2025)</th>
<th>Open</th>
<th>High</th>
<th>Low</th>
<th>Close</th>
<th>Day move</th>
<th>Notes</th>
</tr>
</thead>
<tbody>
<tr>
<td>Nov 19 (Wed)</td>
<td>184.79</td>
<td>187.86</td>
<td>182.83</td>
<td><strong>186.52</strong></td>
<td>+2.9%</td>
<td>Earnings after the close</td>
</tr>
<tr>
<td>Nov 20 (Thu)</td>
<td>195.95</td>
<td>196.00</td>
<td>179.85</td>
<td><strong>180.64</strong></td>
<td>-3.1%</td>
<td>Huge gap up, then big reversal</td>
</tr>
<tr>
<td>Nov 21 (Fri)</td>
<td>181.24</td>
<td>184.56</td>
<td>172.93</td>
<td><strong>178.88</strong></td>
<td>-1.0%</td>
<td>Volatile, closes slightly red</td>
</tr>
</tbody>
</table>
<p>Key points:</p>
<ul>
<li>On <strong>Nov 19</strong>, NVDA closed at <strong>$186.52</strong>, up ~2.9% after trading between roughly <strong>$182.8 and $187.9</strong>.</li>
<li>On <strong>Nov 20</strong>, it <strong>opened $195.95</strong>, <strong>ticked $196.00</strong>, then slid to <strong>$179.85</strong> before closing <strong>$180.64</strong> (about <strong>-3.1%</strong> vs Wednesday’s close but more than <strong>8%</strong> off the intraday high).</li>
<li>On <strong>Nov 21</strong>, it opened at <strong>$181.24</strong>, ranged <strong>$184.56 to $172.93</strong>, and finished <strong>$178.88</strong> (about <strong>-1%</strong> on the day and <strong>~11.8%</strong> below Thursday’s high).</li>
<li>From Thursday’s <strong>$196.00</strong> high to Friday’s <strong>$172.93</strong> low is almost a <strong>12%</strong> swing.</li>
<li>By <strong>Monday 24 November</strong>, NVDA was back in the low <strong>$180s</strong>, still well below the late-October record close around <strong>$207</strong>. (<a href="https://www.macrotrends.net/stocks/charts/NVDA/nvidia/stock-price-history" title="NVIDIA - 26 Year Stock Price History | NVDA">MacroTrends</a>)</li>
</ul>
<h3 id="user-content-snapshot-nvda-on-monday-24-november-intraday">Snapshot: NVDA on Monday, 24 November (intraday)</h3>
<ul>
<li>Latest price <strong>$182.55</strong>, change <strong>$4.00 (0.02%)</strong> vs previous close.</li>
<li>Open <strong>$179.42</strong>; high <strong>$183.41</strong>; low <strong>$176.60</strong>; volume <strong>253,434,849</strong>.</li>
<li>Latest trade time <strong>15:11:44 PST</strong>.</li>
</ul>
<p>The tape looks wild, but the news flow was overwhelmingly positive. That gap between “great fundamentals” and “bad price action” is what fueled the “rigged” and “front-running the jobs report” chatter.</p>
<hr>
<h2 id="user-content-2-earnings-were-exceptional-not-the-problem">2. Earnings were exceptional, not the problem</h2>
<p>After the close on <strong>Wednesday, 19 November</strong>, Nvidia reported:</p>
<ul>
<li><strong>$57.0B</strong> in quarterly revenue for Q3 FY26 (ended 26 October 2025), up <strong>22% QoQ</strong> and <strong>62% YoY</strong>.</li>
<li><strong>$51.2B</strong> in data-center revenue, up <strong>66% YoY</strong>.</li>
<li>Guidance for roughly <strong>$65B</strong> next quarter.</li>
<li>Management reiterated AI accelerator demand was <strong>“off the charts.”</strong> (<a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2026" title="NVIDIA Announces Financial Results for Third Quarter ...">NVIDIA Newsroom</a>)</li>
</ul>
<p>Coverage from Reuters, TechCrunch, the Guardian, and others framed it as another emphatic beat that briefly eased AI-bubble jitters and sparked a global relief rally. The earnings release <strong>added</strong> to bullish sentiment; it didn’t cap it. (<a href="https://www.theguardian.com/technology/2025/nov/19/nvidia-earning-report" title="&#x27;We excel at every phase of AI&#x27;: Nvidia CEO quells Wall Street fears of AI bubble amid market selloff">The Guardian</a>)</p>
<hr>
<h2 id="user-content-3-the-macro-backdrop-a-delayed-jobs-report-hits-with-earnings">3. The macro backdrop: a delayed jobs report hits with earnings</h2>
<p>All times below are <strong>US Eastern Time (ET)</strong>.</p>
<h3 id="user-content-thursday-20-november-2025">Thursday, 20 November 2025</h3>
<p>Because a 43-day federal shutdown delayed official statistics, the <strong>September Employment Situation</strong> report landed on <strong>Thursday 20 November at 08:30 a.m. ET</strong>, alongside weekly jobless claims — an unusual overlap with Nvidia’s earnings glow. (<a href="https://www.bls.gov/news.release/empsit.nr0.htm" title="Employment Situation Summary - 2025 M09 Results">Bureau of Labor Statistics</a>)</p>
<p>Headline numbers:</p>
<ul>
<li>Payrolls: <strong>+119k</strong> vs expectations closer to <strong>+50k</strong>.</li>
<li>Unemployment rate: <strong>4.4%</strong>, the highest in several years.</li>
<li>Prior months <strong>revised down</strong>.</li>
</ul>
<h3 id="user-content-friday-21-november-2025">Friday, 21 November 2025</h3>
<ul>
<li><strong>08:30 a.m. ET</strong> - BLS releases <strong>September Real Earnings</strong>.</li>
</ul>
<h3 id="user-content-crucial-timing">Crucial timing</h3>
<ul>
<li><strong>All the macro prints hit at 08:30 a.m. ET.</strong></li>
<li>The <strong>US cash equity session opens at 09:30 a.m. ET</strong>.</li>
<li>On both days, the macro data were public for <strong>an hour</strong> before regular-session trading in NVDA began.</li>
</ul>
<hr>
<h2 id="user-content-4-the-sequence-earnings-pop-jobs-data-then-a-positioning-unwind">4. The sequence: earnings pop, jobs data, then a positioning unwind</h2>
<h3 id="user-content-wednesday-evening-into-thursday-pre-market">Wednesday evening into Thursday pre-market</h3>
<ul>
<li>After hours on <strong>Nov 19</strong>, NVDA jumped about <strong>5%</strong>, adding well over <strong>$200B</strong> in market value within minutes. (<a href="https://www.reuters.com/world/china/global-markets-trading-day-graphic-2025-11-20/" title="Trading Day: Nvidia sparkle vanishes, Wall Street slammed">Reuters</a>)</li>
<li>Between <strong>8–9 a.m. ET</strong> on <strong>Nov 20</strong>, pre-market updates repeatedly showed NVDA <strong>up 5%+</strong>, dragging Nasdaq futures higher (Nasdaq futures <strong>>+1.5%</strong>). (<a href="https://www.reuters.com/world/china/global-markets-trading-day-graphic-2025-11-20/" title="Trading Day: Nvidia sparkle vanishes, Wall Street slammed">Reuters</a>)</li>
<li><strong>08:30 a.m. ET</strong> — Jobs report and claims drop. Markets digest a “no quick rate cuts, but no collapse either” narrative. Nvidia remains elevated into the open; there is <strong>no evidence of a dump before 8:30</strong>. (<a href="https://www.bls.gov/news.release/empsit.nr0.htm" title="Employment Situation Summary - 2025 M09 Results">Bureau of Labor Statistics</a>)</li>
</ul>
<hr>
<h3 id="user-content-thursday-cash-session-20-nov">Thursday cash session (20 Nov)</h3>
<ul>
<li><strong>09:30 a.m. ET</strong> — NVDA opens <strong>$195.95</strong>, ticks <strong>$196.00</strong> near the open. Indices gap up hard: Dow +700, Nasdaq +2.5% within minutes. (<a href="https://www.investopedia.com/dow-jones-today-11202025-11853411" title="Markets News, Nov. 20, 2025: Major Stock Indexes Post ...">Investopedia</a>)</li>
<li><strong>Late morning (≈10:30–10:45 a.m.)</strong> — Nvidia and the indices <strong>roll over</strong>. The S&#x26;P 500 and Nasdaq slide from early gains of ~2% to losses over 1.5%, a rare intraday reversal more typical of stress days.</li>
<li><strong>Low:</strong> NVDA hits <strong>$179.85</strong>, an <strong>~8.2%</strong> drop from the <strong>$196</strong> high.</li>
<li><strong>Close:</strong> <strong>$180.64</strong>, down <strong>~3.1%</strong> on the day and <strong>~7.8%</strong> below the pre-market high. (<a href="https://stockanalysis.com/stocks/nvda/history/" title="NVIDIA (NVDA) Stock Price History 1999-2025">StockAnalysis</a>)</li>
</ul>
<p>Market color during the slide focused on profit-taking in a crowded AI leader, options-driven hedging (dealers short gamma selling into weakness), and a reassessment of how much to pay for a stock that had already doubled on the AI story. There were <strong>no major new headlines</strong> driving the 10:30–4:00 tape.</p>
<hr>
<h3 id="user-content-friday-cash-session-21-nov">Friday cash session (21 Nov)</h3>
<ul>
<li><strong>Open:</strong> <strong>$181.24</strong>, a touch above Thursday’s close.</li>
<li><strong>Range:</strong> <strong>$184.56</strong> high to <strong>$172.93</strong> low.</li>
<li><strong>Close:</strong> <strong>$178.88</strong>, about <strong>-1%</strong> on the day and <strong>~11.8%</strong> below Thursday’s high. (<a href="https://stockanalysis.com/stocks/nvda/history/" title="NVIDIA (NVDA) Stock Price History 1999-2025">StockAnalysis</a>)</li>
</ul>
<p>The broader market stabilized and even finished green, helped late by Fed commentary. Nvidia stayed weak, pointing to stock-specific positioning rather than fresh macro shock.</p>
<hr>
<h2 id="user-content-5-why-people-think-it-started-before-the-jobs-report">5. Why people think “it started before the jobs report”</h2>
<p>Two mundane issues create the illusion of pre-data selling:</p>
<ol>
<li>
<p><strong>Time zones.</strong> Economic calendars quote <strong>08:30 a.m. ET</strong>; many charts default to <strong>UTC</strong>. <strong>08:30 ET = 13:30 UTC.</strong> If you compare a UTC-stamped calendar to an ET-stamped chart, it can look like the stock moved first when you are staring at the same moment in two clocks. (<a href="https://www.bls.gov/news.release/empsit.nr0.htm" title="Employment Situation Summary - 2025 M09 Results">Bureau of Labor Statistics</a>)</p>
</li>
<li>
<p><strong>Session coverage.</strong> Many retail charts show only the <strong>9:30–16:00 ET</strong> cash session. The first reactions to both the earnings release (after hours) and the jobs data (pre-market) happen <strong>before</strong> those charts even draw a candle. By 9:30, the info is already reflected; what you see next is the positioning unwind.</p>
</li>
<li>
<p><strong>Instrument mix.</strong> Futures and FX trade almost 24/5; single-stock pre-market liquidity is thinner and cash opens later. Overlaying futures with cash equities without adjusting time windows can create false leads/lags.</p>
</li>
</ol>
<p>Once you align clocks and sessions, the sequence looks straightforward: <strong>earnings pop → jobs data digested → late-morning unwind in a crowded trade.</strong></p>
<hr>
<h2 id="user-content-6-what-actually-drove-the-reversal">6. What actually drove the reversal</h2>
<p>Beyond the calendar mechanics, the drivers were classic:</p>
<ul>
<li><strong>Crowded positioning.</strong> Nvidia was the flagship AI trade, the world’s most valuable listed company, and a core holding in AI-focused ETFs. Even “great” numbers can trigger “good enough to take profits.”</li>
<li><strong>Options and systematic flows.</strong> Elevated implied volatility around earnings + macro means dealers hedging short gamma can sell stock as it drops, reinforcing the move. Stop-loss orders and CTA/vol-targeting rules can add momentum.</li>
<li><strong>Valuation check.</strong> After doubling (and then some) on the AI story, investors reassessed how much they would pay in a session that started euphoric and flipped to risk-off.</li>
<li><strong>No fresh negative headline.</strong> Media coverage at the time explicitly noted the reversal came with “no major headlines” — it was flow- and positioning-driven. (<a href="https://www.investopedia.com/dow-jones-today-11202025-11853411" title="Markets News, Nov. 20, 2025: Major Stock Indexes Post ...">Investopedia</a>)</li>
</ul>
<hr>
<h2 id="user-content-7-is-there-evidence-of-manipulation">7. Is there evidence of manipulation?</h2>
<p>To move from “ugly tape” to “illegal manipulation,” you’d expect specifics: non-public access to the jobs report before 8:30, documented spoofing/layering patterns, or regulator/exchange notices flagging abnormal NVDA trading. (<a href="https://www.bls.gov/news.release/empsit.nr0.htm" title="Employment Situation Summary - 2025 M09 Results">Bureau of Labor Statistics</a>)</p>
<p>Based on public information, none of those exist for <strong>19–21 November 2025</strong>, despite intense media attention. Major outlets explained the move with standard factors: positioning, valuations, AI-bubble concerns, and options hedging. (<a href="https://www.reuters.com/world/china/global-markets-trading-day-graphic-2025-11-20/" title="Trading Day: Nvidia sparkle vanishes, Wall Street slammed">Reuters</a>)</p>
<p><strong>Conclusion:</strong> The price path is consistent with a crowded trade being unwound into an earnings + macro window, not a smoking gun of collusion.</p>
<hr>
<h2 id="user-content-8-practical-takeaways">8. Practical takeaways</h2>
<ul>
<li><strong>Don’t anchor on the close.</strong> “Nvidia down 3%” misses an <strong>8% intraday swing</strong> and an <strong>~12% peak-to-trough</strong> over two sessions.</li>
<li><strong>Great fundamentals can still be “sell the news.”</strong> When expectations and ownership are stretched, excellent prints can be liquidity events for de-risking.</li>
<li><strong>Align clocks and sessions.</strong> Check time zones and whether charts include pre-/post-market before inferring causality.</li>
<li><strong>Expect big moves around stacked catalysts.</strong> Earnings + delayed macro + crowded positioning + elevated vol is a recipe for two-sided 5–10% swings, even in mega-caps.</li>
<li><strong>Assume flow before conspiracy.</strong> In the absence of regulatory flags, the base case is positioning and hedging mechanics.</li>
</ul>
<p>If you like data-driven breakdowns like this, explore the rest of the <a href="/blog">blog</a>.</p>
<hr>
<p><em>Standard disclaimer: this article is for informational and educational purposes only and is not investment advice or a recommendation to buy or sell any security.</em></p>


Nvidia stock ripped on record Q3 FY26 earnings, then whipsawed nearly 12% from Nov 19-21, 2025. This post lines up the exact trading timeline with the delayed jobs report, futures action, and positioning to show why the move looked ugly but not rigged — plus practical lessons for traders.

Nvidia's November 2025 Flash Crash: Did Wall Street Rig NVDA or Was It Just a Crowded Trade Unwinding?


## TL;DR: The short answer

* **Gmail absolutely scans your emails with AI and machine learning** for spam filtering, security, and smart features like Smart Compose and email categorization. ([Privacy International][1])
* **According to Google's own documentation,** your Gmail/Workspace data is **not used to train the underlying generative AI models** that power Gemini and Search, unless you explicitly share data or give permission. ([Google Workspace][2])
* **Your Gmail content can be used to improve Gmail's own smart features** (spam filters, Smart Compose, etc.) when certain smart features settings are turned on. That is still AI training, just for Gmail's own models. ([Google Help][3])
* **New Gemini features (like Deep Research) can use Gmail, Drive, and Chat as context** to answer questions, but Google says this contextual use is separate from training the base models. ([Fox News][4])
* You can **significantly reduce how much of your email is used to train or improve AI** by turning off:
  * Smart features in Gmail, Chat, and Meet
  * Smart features in Google Workspace and in other Google products
  * Gemini Apps Activity / "Help improve Google AI"
* You **cannot stop Gmail scanning for spam and security** without leaving Gmail entirely.

This article walks through the "why," "how," and exact settings to change.

## 1. Why Gmail uses AI on your emails at all

Gmail has always relied on machine learning and AI-style systems for three main reasons:

1. **Security and spam filtering**
   * Gmail uses AI-enhanced spam and phishing detection to block a huge volume of junk and malicious emails before they reach your inbox. ([Google Workspace][5])
   * This requires scanning **message content, metadata, links, and attachments** to decide what is safe and what is spam.

2. **Convenience and smart features**
   When you enable smart features, Gmail uses your email content to power things like: ([Google Help][3])
   * Smart Compose (inline text suggestions)
   * Smart Reply (suggested one-tap replies)
   * Primary/Social/Promotions auto-categorization
   * Travel and package summary cards above emails
   * Events auto-added to Calendar, tickets into Wallet, etc.
   Google's own help pages say that when these settings are on, **your Workspace content and activity may also be processed to improve these features**. That is a polite way of saying "used to train or refine the models that power them." ([Google Help][3])

3. **New generative AI features (Gemini in Gmail)**
   * If you have Google AI/Gemini features in Workspace or Gmail, Gemini can **summarize threads, draft replies, and find info in your Workspace content**. ([Google Help][6])
   * For paying Workspace AI and Google One AI plans, Gemini models operate inside Workspace with additional privacy guarantees.

So yes, **AI is deeply baked into Gmail**, but different layers of AI use your emails in different ways.

## 2. Does Gmail actually train AI on your emails?

There are three separate questions here.

### 2.1 Training for spam filters and smart features

This is the part almost nobody argues about.

Google's own smart-features help page says that if you turn on any smart-features settings, your "Workspace content and activity" can be processed **to improve those features** and that these learnings may persist even after you turn the settings off. ([Google Help][3])

That is **AI training**, just not the flashy "big Gemini model" kind. It is training and refining:

* Spam and phishing detection
* Email categorization
* Smart Compose / Smart Reply
* Workspace search suggestions, etc.

So: **if smart features are on, Gmail is absolutely feeding your emails into machine-learning systems to improve those features over time.**

### 2.2 Training the big generative models (Gemini, Search, etc.)

This is where the controversy exploded in late 2025.

Headlines and viral posts claimed that "Google is using your Gmail to train Gemini" and that users were auto-opted in. Fact-checks and follow-up reporting found something more nuanced: ([The Verge][7])

* Google and Workspace privacy pages say **Workspace data (including Gmail) is not used to train or improve the underlying generative AI models that power Gemini and Search without permission.** ([Google Workspace][2])
* A specific Gmail+Gemini help page adds that **Gemini in Gmail/Workspace uses your content to answer prompts but does not use that content to train or improve Gemini or other generative AI models**, and does not store prompts or outputs without your permission. ([Google Help][6])
* Multiple outlets (Verge, Android Central, Snopes, etc.) summarize Google's position the same way: Gmail content itself is not fed into Gemini's base model training pipeline, even though ML is used heavily for Gmail's own features. ([The Verge][7])

So: **according to Google's own docs and public statements, your Gmail messages are not used to train the core Gemini models or Search ranking models by default.**

You might or might not fully trust that, but that is the official position.

### 2.3 What about Gemini Deep Research "scanning" Gmail, Drive, and Chat?

In November 2025, Google announced **Gemini Deep Research can use context from Gmail, Drive, and Chat** to answer questions and perform research-style tasks. ([Fox News][4])

Key points:

* Deep Research can **read your emails and files as context at inference time** to generate answers.
* Coverage notes that Google says **using this context does not mean your Gmail content is automatically added to Gemini's training data**, and that core Gmail content still is not used to train Gemini unless explicitly shared with the AI. ([Fox News][4])

However:

* If you use **Gemini Apps** (the general Gemini web/mobile/chat experience) and explicitly give it data by pasting email content, uploading files, or sharing your screen, **that data is treated like any other Gemini prompt.**
* The Gemini Apps Privacy Hub states that Gemini **uses your activity to provide, maintain, and improve its services, including training generative AI models**, unless you change Gemini Apps Activity settings. ([Google Help][8])

So the realistic picture:

* **Gmail always uses ML on your mail for security and smart features.**
* **Generative AI (Gemini) can now read Gmail content as context to help you, but, per Google, that context is not directly thrown into base model training.**
* Any time you **explicitly share Gmail content with Gemini Apps**, it is treated like normal Gemini prompts and may be used for training unless you opt out in Gemini settings.

## 3. How to stop Gmail and Google AI from using your emails as much as possible

You cannot make Gmail stop scanning for spam and security. But you can:

* Disable or restrict smart features
* Limit how much data can be used to "improve" AI models
* Block Gemini Apps from using your activity for training
* Avoid feeding sensitive content into generative AI at all

### 3.1 Turn off "Smart features in Gmail, Chat, and Meet"

This reduces how much of your email is used to personalize and improve Gmail's AI-powered features.

On desktop Gmail: ([Google Help][3])

1. Open Gmail.
2. Click the gear icon -> "See all settings."
3. In the **General** tab, scroll to **"Smart features in Gmail, Chat and Meet."**
4. Uncheck the box or turn this setting off.
5. Scroll down and click **Save changes**.

On mobile (Android/iOS):

1. Open the Gmail app.
2. Menu -> **Settings** -> choose your account.
3. Under **General**, find **"Smart features."**
4. Uncheck or toggle off.

What you lose:

* Smart Compose and Smart Reply in Gmail
* Automatic Primary/Social/Promotions categorization
* Summary cards like package tracking, travel cards, etc.

What you gain:

* Your Gmail, Chat, and Meet content is no longer used to drive or improve these smart features going forward. Existing learnings may persist, but new data will not be used for improvement once the setting is off. ([Google Help][3])

### 3.2 Turn off "Smart features in Google Workspace" and "Smart features in other Google products"

This is the second layer many people miss. It controls how your Gmail/Workspace data flows into other apps and AI features across Google. ([Google Help][3])

In the Gmail app or on the web:

1. Open Gmail -> Settings.
2. Look for **"Google Workspace smart features"** or a link like **"Manage Workspace smart feature settings."**
3. You should see two options:
   * **Smart features in Google Workspace**
   * **Smart features in other Google products**
4. Turn **both** off.
5. Save if prompted.

What this shuts down:

* Workspace-wide smart features that use Gmail content, such as:
  * Events pulled from Gmail into Calendar
  * Personalized search in Drive based on email/file activity
  * Gemini features inside Workspace that cross-reference Gmail, Drive, Calendar, etc. ([Google Help][3])
* Smart features in other Google products, including:
  * Reservations and tickets pulled into Maps and Wallet
  * Suggestions and answers in the Gemini app based on Workspace content ([Google Help][3])

The same help page makes clear that when any smart-feature setting is on, Workspace content and activity **may be processed to improve those features**. Turning these off stops future use of your mail for that improvement. ([Google Help][3])

### 3.3 Lock down Gemini Apps Activity and training

Even if you lock down Gmail's own settings, Gemini Apps itself can still use what you explicitly give it.

From the **Gemini Apps Privacy Hub**: your prompts, file uploads, and connected-app data can be used to **maintain and improve Gemini, including training its generative models**, and some of that data is reviewed by humans. ([Google Help][8])

To reduce that:

1. Go to **Gemini Apps Activity** (via your Google Account or myactivity.google.com). ([Google Help][8])
2. Turn **Gemini Apps Activity** off, or:
   * Set auto-delete to the shortest interval (for example, 3 months).
   * Explicitly disable the option that lets Google use your activity to improve Google AI, where available.
3. Manually delete past Gemini activity that might contain sensitive email content.

Important nuance: even with this off, Google says some anonymized processing of chats can still be used to improve safety and basic systems, but training of the main models on your identifiable activity is controlled by this setting. ([Google Help][8])

### 3.4 Disable or limit Gemini Deep Research access to Gmail/Drive/Chat

Gemini Deep Research can now pull context from Gmail, Drive, and Chat. Reports note that you can disable this from Google AI/Workspace settings so that Deep Research cannot scan those sources. ([Fox News][4])

Actions to take:

* In **Google Workspace or account-level Gemini settings**, look for controls that enable Gemini access to Gmail, Drive, and Chat and toggle them off where possible.
* Do not enable Labs or experimental Deep Research features that explicitly mention Gmail/Drive context if you want to avoid this entirely.

### 3.5 General Google account privacy tightening

Beyond Gmail-specific toggles, you should:

* Review your **Google Privacy Checkup** and:
  * Turn off or limit **Web & App Activity**.
  * Turn off **Ad personalization**.
  * Consider turning off **Location History**, if not needed. ([Google Transparency][9])

This does not stop Gmail's core spam filtering, but it limits how broadly Google can cross-link your behavior across services.

## 4. Can you completely stop Gmail from scanning your emails?

Short answer: **no, not if you keep using Gmail.**

Even with every smart feature and AI add-on disabled:

* Gmail must still **scan emails for spam, phishing, malware, and abuse**.
* That scanning uses heuristics and machine learning. ([Google Workspace][5])

If you want:

* End-to-end encryption
* Minimal server-side processing
* No mass model-training on content

...you need to consider **a different provider and/or end-to-end encrypted tools**, and possibly host your own mail or use services designed around strong encryption and limited metadata retention.

But for many people, turning off smart features and Gemini training is a pragmatic middle ground: **you keep Gmail's basic security but significantly reduce how much of your email is used to improve AI systems.**

## 5. Practical privacy habits for any email provider

Regardless of how you configure Gmail:

1. **Do not treat email as a vault for secrets.** Avoid storing raw passwords, highly sensitive financial data, or confidential documents directly in your inbox.
2. **Use other channels for truly sensitive content.** For medical, legal, or highly confidential work, consider end-to-end encrypted communication or encrypted file exchange rather than plain email.
3. **Minimize what you feed to generative AI.** Even if policies say data is anonymized, do not paste entire confidential threads into Gemini, ChatGPT, or any other chatbot unless you are comfortable with the associated risks.
4. **Strip metadata before sharing files.** If you send images, remove EXIF and GPS data with the [Media Metadata Scrubber](/tools/media-metadata-scrubber) to avoid leaking location data.
5. **Segment your accounts.** Use one email account for "AI-heavy convenience" (smart features on) and a separate, more locked-down account (smart features off) for critical communications if you must stay within Gmail.

Bottom line: Gmail always runs ML for security and convenience, but you can shut off most training-heavy features, limit Gemini’s use of your data, and keep sensitive content out of any generative AI workflow.

---

* [The Verge](https://www.theverge.com/news/826902/gmail-ai-training-data-opt-out)
* [Android Central](https://www.androidcentral.com/apps-software/why-you-should-care-about-what-companies-are-doing-with-your-clicks)

[1]: https://privacyinternational.org/examples/2170/google-ceases-scanning-email-ad-targeting "Google ceases scanning email for ad targeting"
[2]: https://workspace.google.com/blog/identity-and-security/protecting-your-data-era-generative-ai "Data privacy protections with Duet AI in Google Workspace"
[3]: https://support.google.com/mail/answer/15604322?co=GENIE.Platform%3DAndroid&hl=en "Learn about smart features and controls for Google Workspace and other Google products - Android - Gmail Help"
[4]: https://www.foxnews.com/tech/how-stop-google-ai-from-scanning-your-gmail "Google Gemini now scans Gmail messages - here's how to disable it | Fox News"
[5]: https://workspace.google.com/blog/identity-and-security/an-overview-of-gmails-spam-filters "An overview of Gmail's spam filters"
[6]: https://support.google.com/mail/answer/14615114?hl=en "Learn how Gemini in Gmail, Calendar, Chat, Docs, Drive, Sheets, Slides, Meet and Vids protects your data - Gmail Help"
[7]: https://www.theverge.com/news/826902/gmail-ai-training-data-opt-out "Google denies 'misleading' reports of Gmail using your emails to train AI"
[8]: https://support.google.com/gemini/answer/13594961?hl=en "Gemini Apps Privacy Hub - Gemini Apps Help"
[9]: https://transparency.google/intl/en_us/our-policies/privacy-policy-terms-of-service/ "Privacy Policy, Terms of Service, and AI - Transparency Center"


Does Gmail use your emails to train AI like Google Gemini? This guide explains how Gmail's AI works, what "smart features" really do, and step-by-step how to stop Gmail and Google AI from using your emails wherever possible.

Is Gmail Training AI on Your Emails? What's Really Happening (And How to Lock It Down)


Paul Graham's essays look disarmingly simple: plain language, drifting structure, and then a line that suddenly clicks everything into place. Under the surface, though, he is following a very specific recipe: surprising question x exploratory drafting x brutal rewriting, all aimed at pieces that are "true, important, new, useful." Think of this less as a slogan and more as a workflow you can steal. As a writing coach, I'll walk you through that workflow in a way you can actually apply to your own drafts.

For a quick mental picture, imagine his process as three intertwined moves rather than separate stages. Surprising questions give the essay its energy, exploratory drafts generate unexpected ideas on the page, and brutal rewriting refines both the thinking and the sound of the sentences. Put differently: you collect what surprises you, ask one real question, write to discover what you think, then revise hard until your ideas and your prose finally match.

**Step 1 - Collect your surprises and obsessions.** Graham's best essays usually begin not from grand topics like "education" or "technology" but from little feelings that something is off: why school essays feel fake, why some programmers are wildly more productive, why certain kinds of work feels meaningful and others deadening. Your job is to treat those irritations and fascinations as raw material, jotting them into notebooks or files whenever they pop up. The more faithfully you collect these anomalies, the easier it becomes to spot questions that haven't already been answered to death.

**Step 2 - Turn a topic into a question.** Instead of thinking "I should write about school," you ask "Why does so much school writing feel dead?" or "What's actually going on when students fake essays?" In "The Age of the Essay," Graham sums this up by saying that an essay is something you write "to try to figure something out," not to dress up a conclusion you already have. Framing your piece as a question gives it direction without locking you into a pre-decided answer, which in turn makes it easier to stay honest when the draft surprises you.

**Step 3 - Write a fast, bad version one.** At this point you sit down and pour out a rough draft as quickly as you can, with zero expectation that it will be coherent or persuasive. In his short piece "Writing, Briefly," Graham describes spending more time on rewriting than on the initial pass, treating the first draft mainly as a way to drag half-formed thoughts into the open where he can see them. You should expect ill-fitting examples, shaky claims, and detours that go nowhere; the goal of version one is simply to get wet clay onto the table so you have something to shape.

**Step 4 - Let the draft meander by interest.** As you draft, you follow the most interesting thought from one paragraph to the next, even if it does not match the outline you imagined in your head. In his essays on the form, Graham argues that this kind of wandering is a feature of real inquiry: if you truly follow the question, the path will bend. When you hit a dead end, you back up and try a different branch, trusting that you can clean up the path later in revision.

**Step 5 - Extract what's true, important, and new.** Once that messy draft exists, you switch into treasure-hunt mode and start looking for the few parts that are actually worth keeping. Here it helps to keep three questions in front of you: what in here genuinely surprised me, what would a smart reader actually care about, and what could only you have written in this particular way. In "How to Write Usefully," Graham defines useful writing as telling people something "true and important" that they didn't already know, and says the job is to say it as unequivocally as possible. Those criteria are demanding, but they are also a powerful filter for trimming away fluff.

**Step 6 - Rewrite in loops.** This is the brutal part, and it is where most of the magic happens. On each pass you try to read your own prose as a skeptical, slightly impatient stranger, watching for two kinds of trouble: boredom, where your attention drifts, and friction, where your claims do not really convince you. Drawing on his essay "Putting Ideas into Words," Graham's view is that if a sentence sounds wrong in your ear, it usually means the thought underneath is still half-baked, so you fix the idea rather than just adding adjectives. A useful quick pass is: cut anything that bores you, rethink anything that feels fuzzy, and simplify any sentence you would be embarrassed to say out loud.

Part of this looping rewrite is making the language sound like an actual person. In "Write Like You Talk," Graham urges writers to "write in the same language you'd use if you were talking to a friend," on the grounds that this is both clearer and harder to fake. As a quick check, read a paragraph aloud and ask, "Would I really say this to someone I respect?" If not, you keep tightening until the sentences sound like a calm, thoughtful version of your real voice rather than a committee memo.

**Step 7 - Use external readers as a reality check.** At some point you have taken the draft as far as you can alone, and your private sense of what works stops being trustworthy. This is when you hand it to one or two smart friends and ask very specific questions: where did you get bored, where did you get lost, and which one sentence do you think I'll most regret later. Graham thanks readers on many of his essays and has described using them exactly this way, not to rubber-stamp his ideas but to locate blind spots he can no longer see. When two or more readers highlight the same weak spot, that almost always marks a real problem in either the idea or the explanation.

**Step 8 - Do a final shaping pass.** By now you know what the essay is really about, so you can prune the wandering path into something more direct while keeping the feeling of live thinking. Graham often keeps his digressions in the piece but moves them into footnotes, letting the main line of argument run cleanly from question to insight. On this pass you also watch for repetition, overly sharp lines you might regret, and side-ideas that deserve their own essays rather than a rushed paragraph here.

**Step 9 - Publish and feed the loop.** When he posts an essay on his site, Graham is not treating publication as the finish line; he is throwing the idea into the world to see what questions, objections, and misunderstandings come back. Over the years, those reactions have helped shape later essays and even startup-investing ideas, turning the whole archive into a record of thinking in public. For you, the venue might be a blog, newsletter, or shared document, but the principle is the same: once real readers engage, they generate the next round of surprises to toss back into your notebook.

Stepping back, most parts of this loop are strongly supported by what writing teachers and cognitive scientists know about how people think. Using writing to clarify thought, drafting quickly before your internal critic shuts you down, revising multiple times, and seeking feedback from real readers are all standard best practices, not just one hacker-investor's quirks. Treating essays as ways to figure things out rather than as performances to prove what you already know is also in line with research on learning by explanation: trying to teach someone else forces you to sharpen your own understanding.

The catch is that some of Graham's specific habits are tools to try, not commandments. Letting a draft meander freely works well if you are willing to cut ruthlessly later, but for newer writers it can turn into pages of unfocused wandering. Skipping any outline at all may suit him, yet many people benefit from a loose scaffold - a few anchor points or section headings - that they later feel free to rearrange. The right way to use this process is as a laboratory: adopt the moves that make your thinking sharper and discard the ones that consistently leave you stuck.

Here's what to avoid if you want to write in this spirit. Avoid starting from a thesis you feel you must defend at all costs; that makes you a lawyer, not an investigator, and it tempts you to ignore inconvenient evidence. Avoid writing to sound impressive rather than clear, a trap Graham warns about when he urges people to write as they talk. And avoid treating rewriting as cosmetic, tweaking only words and punctuation while leaving shaky ideas untouched, because in his system the whole point of revision is to rethink what you are saying, not just how you say it.

Knowing when to ask for help is part of the craft. If you have been through several rewrite loops and still cannot say what your essay is really about in one clear sentence, that is a sign you need a reader or a coach to ask blunt questions. If you keep avoiding your own draft because it feels boring or muddled, go back to your list of surprises and obsessions and see whether you chose a question that genuinely interests you. Outside eyes can help, but the process still starts with your own curiosity.

Bottom line, Paul Graham's method is a disciplined way to turn curiosity into useful essays: collect the anomalies that bother you, turn them into real questions, draft quickly to see what you actually think, revise in loops until ideas and sentences click, then publish and let readers feed you the next round of puzzles. Do that enough times and your essays stop being school-style performances and become tools for learning in public. In the process, you not only write better; you train yourself to notice more, think more clearly, and share what you find in a way that genuinely helps other people.

If you want the deeper theory behind why writing clarifies thinking, read [Why Good Thinking Sounds Like Good Conversation](/blog/the-symbiosis-of-thinking-and-talking).


<p>Paul Graham's essays look disarmingly simple: plain language, drifting structure, and then a line that suddenly clicks everything into place. Under the surface, though, he is following a very specific recipe: surprising question x exploratory drafting x brutal rewriting, all aimed at pieces that are "true, important, new, useful." Think of this less as a slogan and more as a workflow you can steal. As a writing coach, I'll walk you through that workflow in a way you can actually apply to your own drafts.</p>
<p>For a quick mental picture, imagine his process as three intertwined moves rather than separate stages. Surprising questions give the essay its energy, exploratory drafts generate unexpected ideas on the page, and brutal rewriting refines both the thinking and the sound of the sentences. Put differently: you collect what surprises you, ask one real question, write to discover what you think, then revise hard until your ideas and your prose finally match.</p>
<p><strong>Step 1 - Collect your surprises and obsessions.</strong> Graham's best essays usually begin not from grand topics like "education" or "technology" but from little feelings that something is off: why school essays feel fake, why some programmers are wildly more productive, why certain kinds of work feels meaningful and others deadening. Your job is to treat those irritations and fascinations as raw material, jotting them into notebooks or files whenever they pop up. The more faithfully you collect these anomalies, the easier it becomes to spot questions that haven't already been answered to death.</p>
<p><strong>Step 2 - Turn a topic into a question.</strong> Instead of thinking "I should write about school," you ask "Why does so much school writing feel dead?" or "What's actually going on when students fake essays?" In "The Age of the Essay," Graham sums this up by saying that an essay is something you write "to try to figure something out," not to dress up a conclusion you already have. Framing your piece as a question gives it direction without locking you into a pre-decided answer, which in turn makes it easier to stay honest when the draft surprises you.</p>
<p><strong>Step 3 - Write a fast, bad version one.</strong> At this point you sit down and pour out a rough draft as quickly as you can, with zero expectation that it will be coherent or persuasive. In his short piece "Writing, Briefly," Graham describes spending more time on rewriting than on the initial pass, treating the first draft mainly as a way to drag half-formed thoughts into the open where he can see them. You should expect ill-fitting examples, shaky claims, and detours that go nowhere; the goal of version one is simply to get wet clay onto the table so you have something to shape.</p>
<p><strong>Step 4 - Let the draft meander by interest.</strong> As you draft, you follow the most interesting thought from one paragraph to the next, even if it does not match the outline you imagined in your head. In his essays on the form, Graham argues that this kind of wandering is a feature of real inquiry: if you truly follow the question, the path will bend. When you hit a dead end, you back up and try a different branch, trusting that you can clean up the path later in revision.</p>
<p><strong>Step 5 - Extract what's true, important, and new.</strong> Once that messy draft exists, you switch into treasure-hunt mode and start looking for the few parts that are actually worth keeping. Here it helps to keep three questions in front of you: what in here genuinely surprised me, what would a smart reader actually care about, and what could only you have written in this particular way. In "How to Write Usefully," Graham defines useful writing as telling people something "true and important" that they didn't already know, and says the job is to say it as unequivocally as possible. Those criteria are demanding, but they are also a powerful filter for trimming away fluff.</p>
<p><strong>Step 6 - Rewrite in loops.</strong> This is the brutal part, and it is where most of the magic happens. On each pass you try to read your own prose as a skeptical, slightly impatient stranger, watching for two kinds of trouble: boredom, where your attention drifts, and friction, where your claims do not really convince you. Drawing on his essay "Putting Ideas into Words," Graham's view is that if a sentence sounds wrong in your ear, it usually means the thought underneath is still half-baked, so you fix the idea rather than just adding adjectives. A useful quick pass is: cut anything that bores you, rethink anything that feels fuzzy, and simplify any sentence you would be embarrassed to say out loud.</p>
<p>Part of this looping rewrite is making the language sound like an actual person. In "Write Like You Talk," Graham urges writers to "write in the same language you'd use if you were talking to a friend," on the grounds that this is both clearer and harder to fake. As a quick check, read a paragraph aloud and ask, "Would I really say this to someone I respect?" If not, you keep tightening until the sentences sound like a calm, thoughtful version of your real voice rather than a committee memo.</p>
<p><strong>Step 7 - Use external readers as a reality check.</strong> At some point you have taken the draft as far as you can alone, and your private sense of what works stops being trustworthy. This is when you hand it to one or two smart friends and ask very specific questions: where did you get bored, where did you get lost, and which one sentence do you think I'll most regret later. Graham thanks readers on many of his essays and has described using them exactly this way, not to rubber-stamp his ideas but to locate blind spots he can no longer see. When two or more readers highlight the same weak spot, that almost always marks a real problem in either the idea or the explanation.</p>
<p><strong>Step 8 - Do a final shaping pass.</strong> By now you know what the essay is really about, so you can prune the wandering path into something more direct while keeping the feeling of live thinking. Graham often keeps his digressions in the piece but moves them into footnotes, letting the main line of argument run cleanly from question to insight. On this pass you also watch for repetition, overly sharp lines you might regret, and side-ideas that deserve their own essays rather than a rushed paragraph here.</p>
<p><strong>Step 9 - Publish and feed the loop.</strong> When he posts an essay on his site, Graham is not treating publication as the finish line; he is throwing the idea into the world to see what questions, objections, and misunderstandings come back. Over the years, those reactions have helped shape later essays and even startup-investing ideas, turning the whole archive into a record of thinking in public. For you, the venue might be a blog, newsletter, or shared document, but the principle is the same: once real readers engage, they generate the next round of surprises to toss back into your notebook.</p>
<p>Stepping back, most parts of this loop are strongly supported by what writing teachers and cognitive scientists know about how people think. Using writing to clarify thought, drafting quickly before your internal critic shuts you down, revising multiple times, and seeking feedback from real readers are all standard best practices, not just one hacker-investor's quirks. Treating essays as ways to figure things out rather than as performances to prove what you already know is also in line with research on learning by explanation: trying to teach someone else forces you to sharpen your own understanding.</p>
<p>The catch is that some of Graham's specific habits are tools to try, not commandments. Letting a draft meander freely works well if you are willing to cut ruthlessly later, but for newer writers it can turn into pages of unfocused wandering. Skipping any outline at all may suit him, yet many people benefit from a loose scaffold - a few anchor points or section headings - that they later feel free to rearrange. The right way to use this process is as a laboratory: adopt the moves that make your thinking sharper and discard the ones that consistently leave you stuck.</p>
<p>Here's what to avoid if you want to write in this spirit. Avoid starting from a thesis you feel you must defend at all costs; that makes you a lawyer, not an investigator, and it tempts you to ignore inconvenient evidence. Avoid writing to sound impressive rather than clear, a trap Graham warns about when he urges people to write as they talk. And avoid treating rewriting as cosmetic, tweaking only words and punctuation while leaving shaky ideas untouched, because in his system the whole point of revision is to rethink what you are saying, not just how you say it.</p>
<p>Knowing when to ask for help is part of the craft. If you have been through several rewrite loops and still cannot say what your essay is really about in one clear sentence, that is a sign you need a reader or a coach to ask blunt questions. If you keep avoiding your own draft because it feels boring or muddled, go back to your list of surprises and obsessions and see whether you chose a question that genuinely interests you. Outside eyes can help, but the process still starts with your own curiosity.</p>
<p>Bottom line, Paul Graham's method is a disciplined way to turn curiosity into useful essays: collect the anomalies that bother you, turn them into real questions, draft quickly to see what you actually think, revise in loops until ideas and sentences click, then publish and let readers feed you the next round of puzzles. Do that enough times and your essays stop being school-style performances and become tools for learning in public. In the process, you not only write better; you train yourself to notice more, think more clearly, and share what you find in a way that genuinely helps other people.</p>
<p>If you want the deeper theory behind why writing clarifies thinking, read <a href="/blog/the-symbiosis-of-thinking-and-talking">Why Good Thinking Sounds Like Good Conversation</a>.</p>


A practical breakdown of Paul Graham's writing process: surprise-driven questions, exploratory drafts, and hard rewrites that make essays useful.

How Paul Graham Actually Writes Essays

A practical overview of Google DeepMind's SynthID watermarking: which models use it, which do not, and how detection limits affect provenance workflows.

SynthID in 2025: Where Google’s Invisible Watermark Shows Up (and Where It Doesn’t)


What's the relationship between thinking and talking?

The standard picture goes like this: first you think, then you speak. You generate ideas internally, and then you encode them into language for other people. Clean, linear, and wrong.

Watch your own mind closely when you're struggling with something hard: a tough decision, a knotty problem at work, an abstract concept you're trying to grasp. You don't sit there with finished thoughts waiting patiently in a queue to be translated into words. You stumble. You half-form sentences. You argue with yourself. You discover, while trying to say something, that you don't quite know what you mean yet.

That's the interesting part: very often, you don't think first and talk later. You figure out what you think by trying to say it.

Thinking as internalized conversation

When you are really thinking, what actually happens?

For a lot of people, the honest answer is: a running dialogue. You pose questions to yourself. You test explanations. You rephrase. You anticipate objections. The activity feels less like decoding fully formed ideas into words and more like wrestling with messy half-formed sentences until they click.

This isn't just a poetic analogy. A huge amount of what we call "thinking" uses the same machinery as conversation. The difference is mostly where the words go: out into the air, or back into your own head.

We didn't evolve a separate "high-level reasoning module" that just happens to interface with language. Language itself became our main workspace for complex, abstract thought. Instead of building a new system from scratch, evolution leaned on an existing one and pushed it inward. We learned not only to talk to one another, but to talk to ourselves.

How children learn to think out loud

You can see this process in miniature in young children.

Watch a five-year-old doing a puzzle or playing an imaginative game. They often narrate what they're doing: "Okay, this piece goes here... no, that's wrong... maybe the blue one." Psychologists call this "private speech." It looks like chatter, but it plays a real functional role. They are using language to guide their attention, plan actions, and keep track of rules.

Over time, that external monologue goes underground. The child doesn't stop using language to think; they just stop saying as much of it aloud. The voice moves inside. The end result is the inner dialogue many adults experience when they are concentrating.

In other words, for a lot of us, "thinking" is what happens when private speech becomes fast and quiet.

But what about things you can't easily say?

If thinking leans so heavily on words, how do we deal with things that aren't naturally verbal, like rotating a 3D object in your mind, improvising music, or feeling grief?

This is where the first simple story (thinking equals talking) breaks. Not all thought is verbal. We clearly have visual, spatial, auditory, and emotional forms of cognition that don't require full sentences.

However, there's an important pattern:

For immediate perception and skilled action (catching a ball, reading a face, playing a piece you've practiced a thousand times), you mostly don't need language.

For deliberate reasoning about complex or abstract issues (justice, strategy, systems, tradeoffs), you almost always drag in words.

You can feel confused without words. But when you want to debug the confusion, you start to narrate: "What exactly is bothering me here?" As soon as you seriously try to justify, compare, or generalize, you end up talking to yourself if not to others.

Language is not the only form of thought, but it is our default medium for explicit, structured reasoning.

Why explaining things changes your mind

This is why explaining something can suddenly make it clearer, even if you thought you already understood it.

When you explain an idea to someone else, you can't rely on your private shortcuts. Your internal dialogue is allowed to be fragmentary because you already know the missing pieces. The listener doesn't. To make sense to them, you have to:

Fill in the gaps.

Make assumptions explicit.

Order the steps so they actually follow.

Replace fuzzy placeholders ("you know, that thing") with real content.

That process often reveals that your mental model was half-baked. You thought you "knew" it, but it only worked inside the forgiving environment of your own head. By forcing yourself to articulate the idea in a full, public sentence, you either solidify it or expose where it breaks.

Teaching, in that sense, doesn't just communicate understanding; it manufactures it.

Why writing is a thinking tool, not just an output

Writing goes even further.

Speech can be sloppy and ephemeral. You say something, correct yourself mid-sentence, and the previous version disappears. Writing leaves a record. Once a sentence is on the page, you can stare at it and ask:

Is this actually what I mean?

Is there a hidden assumption here?

Does this follow from the previous paragraph?

Would someone else interpret this the way I intend?

When you fail to write something clearly, the usual reason is not lack of literary talent. It's that you haven't finished thinking it through. The problem isn't in the keyboard; it's in the model.

Rewriting isn't "polish" in the cosmetic sense. It is iterative thinking in public. Each draft is a pass at making your internal dialogue more explicit, more precise, and less contradictory.

Training your inner voice

If a lot of your reasoning runs on internal dialogue, then the quality of that dialogue matters.

Two people can have the same raw cognitive horsepower and end up with very different levels of insight simply because one of them has trained a better inner conversational partner. Consider the difference between:

An internal voice that says, "This is probably fine, don't overthink it," versus

One that says, "Spell out what you mean. What's the evidence? What would convince you you're wrong?"

The second person isn't smarter in the IQ-test sense; they just have a more demanding, better-practiced internal critic. They've spent years explaining their own thoughts to themselves, on paper, to other people, and that training shows up as clearer thinking.

The practical implication is simple: you can improve your thinking by deliberately improving how you "talk" about problems.

Concretely:

- When you're stuck, write down what you think is true, then write the objections.
- When something feels off, try explaining the situation to an imaginary colleague who knows nothing about it.
- When you have an intuition, force yourself to state it as a claim and list what would falsify it.

You are upgrading the quality of your inner conversation.

Fragmented communication, fragmented thought

Modern communication tools don't change any of this, but they push on it.

Text messages, group chats, and social feeds reward speed and brevity. They optimize for "reply now" and "fit it in a tiny box," not for building a long chain of reasoning. If most of your expressive life happens in that environment, you may find your thoughts starting to take on the same shape: reactive, clipped, half-finished.

That doesn't mean short form is bad by definition. A tight, 200-character statement can be a model of clarity. The danger is not length; it's habit. If you never practice sustained, careful articulation, you shouldn't expect to have sustained, careful thoughts.

You think the way you train yourself to talk.

The real punchline

The interesting claim is not that thinking and talking are literally the same thing. They aren't. You can have rich, meaningful, nonverbal experience. You can play chess largely in patterns, not sentences.

The real claim is narrower and more useful:

For the kind of complex, explicit reasoning humans care about (plans, arguments, explanations, theories), language is the main workspace.

Your internal dialogue and your external expression are two sides of the same system.

Improving one almost inevitably improves the other.

So better thinking is not a mystical gift reserved for people with exotic brains. It is, to a large extent, the result of a skill you can practice: learning to have better conversations with others, and with yourself.

Talk better to think better. And when in doubt, write.

For a practical writing workflow built on these ideas, see [How Paul Graham Actually Writes Essays](/blog/how-paul-graham-writes-essays).


<p>What's the relationship between thinking and talking?</p>
<p>The standard picture goes like this: first you think, then you speak. You generate ideas internally, and then you encode them into language for other people. Clean, linear, and wrong.</p>
<p>Watch your own mind closely when you're struggling with something hard: a tough decision, a knotty problem at work, an abstract concept you're trying to grasp. You don't sit there with finished thoughts waiting patiently in a queue to be translated into words. You stumble. You half-form sentences. You argue with yourself. You discover, while trying to say something, that you don't quite know what you mean yet.</p>
<p>That's the interesting part: very often, you don't think first and talk later. You figure out what you think by trying to say it.</p>
<p>Thinking as internalized conversation</p>
<p>When you are really thinking, what actually happens?</p>
<p>For a lot of people, the honest answer is: a running dialogue. You pose questions to yourself. You test explanations. You rephrase. You anticipate objections. The activity feels less like decoding fully formed ideas into words and more like wrestling with messy half-formed sentences until they click.</p>
<p>This isn't just a poetic analogy. A huge amount of what we call "thinking" uses the same machinery as conversation. The difference is mostly where the words go: out into the air, or back into your own head.</p>
<p>We didn't evolve a separate "high-level reasoning module" that just happens to interface with language. Language itself became our main workspace for complex, abstract thought. Instead of building a new system from scratch, evolution leaned on an existing one and pushed it inward. We learned not only to talk to one another, but to talk to ourselves.</p>
<p>How children learn to think out loud</p>
<p>You can see this process in miniature in young children.</p>
<p>Watch a five-year-old doing a puzzle or playing an imaginative game. They often narrate what they're doing: "Okay, this piece goes here... no, that's wrong... maybe the blue one." Psychologists call this "private speech." It looks like chatter, but it plays a real functional role. They are using language to guide their attention, plan actions, and keep track of rules.</p>
<p>Over time, that external monologue goes underground. The child doesn't stop using language to think; they just stop saying as much of it aloud. The voice moves inside. The end result is the inner dialogue many adults experience when they are concentrating.</p>
<p>In other words, for a lot of us, "thinking" is what happens when private speech becomes fast and quiet.</p>
<p>But what about things you can't easily say?</p>
<p>If thinking leans so heavily on words, how do we deal with things that aren't naturally verbal, like rotating a 3D object in your mind, improvising music, or feeling grief?</p>
<p>This is where the first simple story (thinking equals talking) breaks. Not all thought is verbal. We clearly have visual, spatial, auditory, and emotional forms of cognition that don't require full sentences.</p>
<p>However, there's an important pattern:</p>
<p>For immediate perception and skilled action (catching a ball, reading a face, playing a piece you've practiced a thousand times), you mostly don't need language.</p>
<p>For deliberate reasoning about complex or abstract issues (justice, strategy, systems, tradeoffs), you almost always drag in words.</p>
<p>You can feel confused without words. But when you want to debug the confusion, you start to narrate: "What exactly is bothering me here?" As soon as you seriously try to justify, compare, or generalize, you end up talking to yourself if not to others.</p>
<p>Language is not the only form of thought, but it is our default medium for explicit, structured reasoning.</p>
<p>Why explaining things changes your mind</p>
<p>This is why explaining something can suddenly make it clearer, even if you thought you already understood it.</p>
<p>When you explain an idea to someone else, you can't rely on your private shortcuts. Your internal dialogue is allowed to be fragmentary because you already know the missing pieces. The listener doesn't. To make sense to them, you have to:</p>
<p>Fill in the gaps.</p>
<p>Make assumptions explicit.</p>
<p>Order the steps so they actually follow.</p>
<p>Replace fuzzy placeholders ("you know, that thing") with real content.</p>
<p>That process often reveals that your mental model was half-baked. You thought you "knew" it, but it only worked inside the forgiving environment of your own head. By forcing yourself to articulate the idea in a full, public sentence, you either solidify it or expose where it breaks.</p>
<p>Teaching, in that sense, doesn't just communicate understanding; it manufactures it.</p>
<p>Why writing is a thinking tool, not just an output</p>
<p>Writing goes even further.</p>
<p>Speech can be sloppy and ephemeral. You say something, correct yourself mid-sentence, and the previous version disappears. Writing leaves a record. Once a sentence is on the page, you can stare at it and ask:</p>
<p>Is this actually what I mean?</p>
<p>Is there a hidden assumption here?</p>
<p>Does this follow from the previous paragraph?</p>
<p>Would someone else interpret this the way I intend?</p>
<p>When you fail to write something clearly, the usual reason is not lack of literary talent. It's that you haven't finished thinking it through. The problem isn't in the keyboard; it's in the model.</p>
<p>Rewriting isn't "polish" in the cosmetic sense. It is iterative thinking in public. Each draft is a pass at making your internal dialogue more explicit, more precise, and less contradictory.</p>
<p>Training your inner voice</p>
<p>If a lot of your reasoning runs on internal dialogue, then the quality of that dialogue matters.</p>
<p>Two people can have the same raw cognitive horsepower and end up with very different levels of insight simply because one of them has trained a better inner conversational partner. Consider the difference between:</p>
<p>An internal voice that says, "This is probably fine, don't overthink it," versus</p>
<p>One that says, "Spell out what you mean. What's the evidence? What would convince you you're wrong?"</p>
<p>The second person isn't smarter in the IQ-test sense; they just have a more demanding, better-practiced internal critic. They've spent years explaining their own thoughts to themselves, on paper, to other people, and that training shows up as clearer thinking.</p>
<p>The practical implication is simple: you can improve your thinking by deliberately improving how you "talk" about problems.</p>
<p>Concretely:</p>
<ul>
<li>When you're stuck, write down what you think is true, then write the objections.</li>
<li>When something feels off, try explaining the situation to an imaginary colleague who knows nothing about it.</li>
<li>When you have an intuition, force yourself to state it as a claim and list what would falsify it.</li>
</ul>
<p>You are upgrading the quality of your inner conversation.</p>
<p>Fragmented communication, fragmented thought</p>
<p>Modern communication tools don't change any of this, but they push on it.</p>
<p>Text messages, group chats, and social feeds reward speed and brevity. They optimize for "reply now" and "fit it in a tiny box," not for building a long chain of reasoning. If most of your expressive life happens in that environment, you may find your thoughts starting to take on the same shape: reactive, clipped, half-finished.</p>
<p>That doesn't mean short form is bad by definition. A tight, 200-character statement can be a model of clarity. The danger is not length; it's habit. If you never practice sustained, careful articulation, you shouldn't expect to have sustained, careful thoughts.</p>
<p>You think the way you train yourself to talk.</p>
<p>The real punchline</p>
<p>The interesting claim is not that thinking and talking are literally the same thing. They aren't. You can have rich, meaningful, nonverbal experience. You can play chess largely in patterns, not sentences.</p>
<p>The real claim is narrower and more useful:</p>
<p>For the kind of complex, explicit reasoning humans care about (plans, arguments, explanations, theories), language is the main workspace.</p>
<p>Your internal dialogue and your external expression are two sides of the same system.</p>
<p>Improving one almost inevitably improves the other.</p>
<p>So better thinking is not a mystical gift reserved for people with exotic brains. It is, to a large extent, the result of a skill you can practice: learning to have better conversations with others, and with yourself.</p>
<p>Talk better to think better. And when in doubt, write.</p>
<p>For a practical writing workflow built on these ideas, see <a href="/blog/how-paul-graham-writes-essays">How Paul Graham Actually Writes Essays</a>.</p>


Why language and thought are tightly linked, and how better internal dialogue leads to clearer reasoning and better writing.

Why Good Thinking Sounds Like Good Conversation


What does it mean to be human?

Here is the blunt version: you are a fragile animal that knows it is going to die, and you are trying to make that fact bearable.

Everything else is details.

We build religions, write laws, fall in love, write code, start wars, write poems, scroll endlessly on phones, and stare at ceilings at 3 a.m. All of that is one species wrestling with the same basic problem: we are conscious, temporary, and we care.

* * *

## A body that knows it will end

To be human is first to be a body.

You bleed, you ache, you age. You get hungry and tired and sick. Your heart will stop. That is not poetry; it is a medical certainty.

But unlike other animals, you do not just _experience_ this; you can _anticipate_ it. You can imagine your own death in vivid detail. You can picture the funeral, the empty chair, the messages people will send when you are gone.

This is both a curse and a superpower. Without the awareness of death, there would be no urgency, no bucket lists, no sense that time matters. With it, there is anxiety, denial, and the constant temptation to distract ourselves from the obvious.

Being human means carrying that awareness and still doing the dishes.

* * *

## A brain that runs simulations

Humans do not just react; we rehearse.

Your brain runs mental simulations all day: what could go wrong, what might happen, what they might say, how you will respond. You replay past conversations and rewrite them; you rehearse future ones that may never happen.

This ability to imagine, to tell stories, is a defining feature of being human. We build entire worlds that exist only in language:

- nations
- companies
- religions
- markets
- money

None of these are physical objects. They exist because enough people agree to treat them as real.

To be human is to live inside stories, not just inside skin.

* * *

## We become ourselves through others

You do not invent yourself alone.

From birth, you are shaped by other people: parents, siblings, friends, teachers, enemies, strangers online. You learn what is "normal" from the tribe you happen to be born into: what you can say, who you can love, what you should want.

Being human means:

- craving approval and fearing rejection
- constantly reading faces for signals of belonging or threat
- feeling joy or shame because of what other people think, or what you imagine they might think

We like to talk about "being true to yourself," but the self is a social construction. You are, in large part, the reflection of the relationships you have had and the culture you swim in.

To be human is to be entangled.

* * *

## The moral animal

We are not just clever apes; we are judging apes.

Humans obsess over "should." We argue not only about what _is_ happening, but about what _ought_ to happen:

- Is this fair?
- Is this right?
- Who deserves what?

We will sacrifice money, comfort, even our lives for principles, but we will also twist principles to justify what we wanted to do anyway. We are capable of astonishing generosity and breathtaking cruelty, often in the name of the same moral ideal.

To be human is to live within a permanent tension:

- between empathy and indifference
- between justice and self-interest
- between who we think we are and what we actually do

Our moral sense is not clean; it is complicated, biased, and sometimes hypocritical. Yet without it, we would not recognize ourselves.

* * *

## Creativity, and the mess that comes with it

Humans make things that do not exist yet.

We paint, compose, design, engineer, theorize. We build tools that make new tools possible. Fire, wheels, vaccines, nuclear weapons, smartphones, machine learning systems: each step extends our power and our reach.

But every tool is double-edged.

The same ingenuity that writes a symphony builds a concentration camp. The same algorithms that connect families spread disinformation and addiction. The same curiosity that unlocks clean energy can also refine more efficient ways to destroy.

To be human is to create beyond our current level of wisdom, and then scramble to catch up.

* * *

## The search for meaning

Underneath the noise of daily life, there is a quieter question: "What is the point of all this?"

People answer it in many ways:

- religion and faith
- art and beauty
- service and community
- success and status
- family and love
- exploration and discovery

Some lean on ancient traditions; others try to build their own from scratch. Many just try not to think about it too much.

There is no single correct answer. The uncomfortable truth is that meaning is not handed to us; we project it. We decide what matters, and then we live as if that decision is objectively true.

To be human is to participate in this act of meaning-making, consciously or not.

* * *

## Being human is a verb

So what does it mean to be human?

It is not just belonging to the species Homo sapiens or having a highly developed brain. Those are technicalities. The deeper reality is:

To be human is to be:

- a mortal body that knows it is mortal,
- a mind that tells stories,
- a social being who becomes itself through others,
- a moral agent struggling with its own contradictions,
- a creator of tools and meanings that can outgrow our control.

And crucially, it is not a static state. Being human is something you do every day:

- in the way you treat the people who cannot give you anything
- in what you pay attention to when nobody is watching
- in the stories you choose to believe and pass on
- in how you respond to your own fear, confusion, and desire

We are not just the species that asks, "What does it mean to be human?" We are the species that answers that question with our behavior, collectively, in real time.

The definition is not finished. It is being revised every time we choose to act with courage instead of cowardice, curiosity instead of indifference, honesty instead of comfort.

That, more than any cosmic metaphor, is what it means to be human.

If you want a companion essay on how language shapes thought, read [Why Good Thinking Sounds Like Good Conversation](/blog/the-symbiosis-of-thinking-and-talking).


<p>What does it mean to be human?</p>
<p>Here is the blunt version: you are a fragile animal that knows it is going to die, and you are trying to make that fact bearable.</p>
<p>Everything else is details.</p>
<p>We build religions, write laws, fall in love, write code, start wars, write poems, scroll endlessly on phones, and stare at ceilings at 3 a.m. All of that is one species wrestling with the same basic problem: we are conscious, temporary, and we care.</p>
<hr>
<h2 id="user-content-a-body-that-knows-it-will-end">A body that knows it will end</h2>
<p>To be human is first to be a body.</p>
<p>You bleed, you ache, you age. You get hungry and tired and sick. Your heart will stop. That is not poetry; it is a medical certainty.</p>
<p>But unlike other animals, you do not just <em>experience</em> this; you can <em>anticipate</em> it. You can imagine your own death in vivid detail. You can picture the funeral, the empty chair, the messages people will send when you are gone.</p>
<p>This is both a curse and a superpower. Without the awareness of death, there would be no urgency, no bucket lists, no sense that time matters. With it, there is anxiety, denial, and the constant temptation to distract ourselves from the obvious.</p>
<p>Being human means carrying that awareness and still doing the dishes.</p>
<hr>
<h2 id="user-content-a-brain-that-runs-simulations">A brain that runs simulations</h2>
<p>Humans do not just react; we rehearse.</p>
<p>Your brain runs mental simulations all day: what could go wrong, what might happen, what they might say, how you will respond. You replay past conversations and rewrite them; you rehearse future ones that may never happen.</p>
<p>This ability to imagine, to tell stories, is a defining feature of being human. We build entire worlds that exist only in language:</p>
<ul>
<li>nations</li>
<li>companies</li>
<li>religions</li>
<li>markets</li>
<li>money</li>
</ul>
<p>None of these are physical objects. They exist because enough people agree to treat them as real.</p>
<p>To be human is to live inside stories, not just inside skin.</p>
<hr>
<h2 id="user-content-we-become-ourselves-through-others">We become ourselves through others</h2>
<p>You do not invent yourself alone.</p>
<p>From birth, you are shaped by other people: parents, siblings, friends, teachers, enemies, strangers online. You learn what is "normal" from the tribe you happen to be born into: what you can say, who you can love, what you should want.</p>
<p>Being human means:</p>
<ul>
<li>craving approval and fearing rejection</li>
<li>constantly reading faces for signals of belonging or threat</li>
<li>feeling joy or shame because of what other people think, or what you imagine they might think</li>
</ul>
<p>We like to talk about "being true to yourself," but the self is a social construction. You are, in large part, the reflection of the relationships you have had and the culture you swim in.</p>
<p>To be human is to be entangled.</p>
<hr>
<h2 id="user-content-the-moral-animal">The moral animal</h2>
<p>We are not just clever apes; we are judging apes.</p>
<p>Humans obsess over "should." We argue not only about what <em>is</em> happening, but about what <em>ought</em> to happen:</p>
<ul>
<li>Is this fair?</li>
<li>Is this right?</li>
<li>Who deserves what?</li>
</ul>
<p>We will sacrifice money, comfort, even our lives for principles, but we will also twist principles to justify what we wanted to do anyway. We are capable of astonishing generosity and breathtaking cruelty, often in the name of the same moral ideal.</p>
<p>To be human is to live within a permanent tension:</p>
<ul>
<li>between empathy and indifference</li>
<li>between justice and self-interest</li>
<li>between who we think we are and what we actually do</li>
</ul>
<p>Our moral sense is not clean; it is complicated, biased, and sometimes hypocritical. Yet without it, we would not recognize ourselves.</p>
<hr>
<h2 id="user-content-creativity-and-the-mess-that-comes-with-it">Creativity, and the mess that comes with it</h2>
<p>Humans make things that do not exist yet.</p>
<p>We paint, compose, design, engineer, theorize. We build tools that make new tools possible. Fire, wheels, vaccines, nuclear weapons, smartphones, machine learning systems: each step extends our power and our reach.</p>
<p>But every tool is double-edged.</p>
<p>The same ingenuity that writes a symphony builds a concentration camp. The same algorithms that connect families spread disinformation and addiction. The same curiosity that unlocks clean energy can also refine more efficient ways to destroy.</p>
<p>To be human is to create beyond our current level of wisdom, and then scramble to catch up.</p>
<hr>
<h2 id="user-content-the-search-for-meaning">The search for meaning</h2>
<p>Underneath the noise of daily life, there is a quieter question: "What is the point of all this?"</p>
<p>People answer it in many ways:</p>
<ul>
<li>religion and faith</li>
<li>art and beauty</li>
<li>service and community</li>
<li>success and status</li>
<li>family and love</li>
<li>exploration and discovery</li>
</ul>
<p>Some lean on ancient traditions; others try to build their own from scratch. Many just try not to think about it too much.</p>
<p>There is no single correct answer. The uncomfortable truth is that meaning is not handed to us; we project it. We decide what matters, and then we live as if that decision is objectively true.</p>
<p>To be human is to participate in this act of meaning-making, consciously or not.</p>
<hr>
<h2 id="user-content-being-human-is-a-verb">Being human is a verb</h2>
<p>So what does it mean to be human?</p>
<p>It is not just belonging to the species Homo sapiens or having a highly developed brain. Those are technicalities. The deeper reality is:</p>
<p>To be human is to be:</p>
<ul>
<li>a mortal body that knows it is mortal,</li>
<li>a mind that tells stories,</li>
<li>a social being who becomes itself through others,</li>
<li>a moral agent struggling with its own contradictions,</li>
<li>a creator of tools and meanings that can outgrow our control.</li>
</ul>
<p>And crucially, it is not a static state. Being human is something you do every day:</p>
<ul>
<li>in the way you treat the people who cannot give you anything</li>
<li>in what you pay attention to when nobody is watching</li>
<li>in the stories you choose to believe and pass on</li>
<li>in how you respond to your own fear, confusion, and desire</li>
</ul>
<p>We are not just the species that asks, "What does it mean to be human?" We are the species that answers that question with our behavior, collectively, in real time.</p>
<p>The definition is not finished. It is being revised every time we choose to act with courage instead of cowardice, curiosity instead of indifference, honesty instead of comfort.</p>
<p>That, more than any cosmic metaphor, is what it means to be human.</p>
<p>If you want a companion essay on how language shapes thought, read <a href="/blog/the-symbiosis-of-thinking-and-talking">Why Good Thinking Sounds Like Good Conversation</a>.</p>


A concise reflection on what makes us human: mortality, narrative thinking, social identity, moral tension, and meaning-making.

Browser-based tools, interactive maps, and product experiments built with a craftsman's bias for clarity.

Signal board

Tools

Dogs Explorer

Subvenciones ONGs

Maps

Things you can use today

Tools

Dogs Explorer

Subvenciones ONGs

Maps

Audio Converter

Video Converter

Image Converter

Blog

A shelf of independent bets

Sondeas

WhatYouEat

PluginsDB

Subscription Land

Build logs and ideas

AI Slop Is Eating the Internet: The 2026 Guide to Spotting It, Avoiding It, and Not Publishing It

OpenClaw Explained: The Complete Story of the Viral AI Agent That Broke the Internet

X Deleted Grok's Image Tab - The Search Hack That Still Shows All Grok Images

Direct Provider Pricing for Image + Video Generation APIs (USD)

Latest updates

Audio Converter

Video Converter

Image Converter

Dogs Breed Explorer

Guía Subvenciones Digitales

Professional Rates Calculator

WordPress Plugins