The Perfect Enforcement Test

Almost anyone who has been driving well for twenty years would fail the driving test today. Anyone who has gone back for a second license after years behind the wheel knows this: you take the exam again and the lottery is the same as ever. You make faults. You fail. Not because you've forgotten how to drive, but for a more uncomfortable reason: the test doesn't measure whether you drive well. It measures the distance between your actual driving and an ideal driving that nobody practices. That distance exists in every driver, every day: a turn signal that doesn't last the full three seconds, a following distance that shrinks for an instant, a solid line grazed while merging. Given enough exposure to the road, an infraction is not a possibility. It's a statistical certainty.

Traffic is merely where this is easiest to see, because we all drive in full view of everyone else. But the structure repeats in every corner of personal and professional life. Anyone with a business or moderately complex finances knows that a flawless tax return is improbable: the tax code is so vast and so changeable that even the advisors disagree about what's deductible. Data protection rules require every organization to maintain documented, continuously updated compliance across every process; among the compliance professionals themselves, it's a commonplace that full, sustained conformity is more an aspiration than a state. No employee follows to the letter the internal handbook they signed without reading on day one. Rare is the business — the construction site, the bar, the shop — that simultaneously satisfies every ordinance, license, and technical requirement that applies to it. And nearly all of us have "accepted" thousands of pages of terms of service that it would be astonishing not to be violating in some way. The general rule of regulated life is this: given enough exposure to any regulated domain, and enough detail in the rules, almost everyone ends up in breach.

“Given enough exposure to the road, an infraction is not a possibility. It's a statistical certainty.”

And yet everything works. Why?

Because the rule we actually lived under was never the written rule. The real rule was the text multiplied by two factors that appear in no gazette, no contract, no handbook: the probability of being observed, and the judgment of the human observing you. The tax inspector picked a few returns out of millions. The labor inspector visited one company among thousands. The floor manager saw someone arrive late and decided, based on context and person, whether it mattered. The traffic officer applied experience-trained judgment to a concrete situation right in front of him. You didn't have to comply perfectly; you had to behave well enough not to draw the attention of someone reasonable. The tolerance was written nowhere, but it was part of the design.

Put another way: until now we have been free, in part, because power could not see us. And I say power, not just the state, because the authors of impossible rules are not only legislators. They are also employers, sector regulators, digital platforms. Each of them lived under the same scarcity of eyes.

“Until now we have been free, in part, because power could not see us.”

Artificial intelligence eliminates exactly that scarcity, and it eliminates it everywhere at once. On the road, Spain's DGT says 232 AI-enabled cameras detect offenses such as holding a phone while driving or failing to wear a seatbelt. The images are reviewed automatically and then by specialized staff before any reports are validated. (DGT) In taxation, Spain's new invoicing-system rules will require tamper-resistant records protected through hashing and chaining. VERI*FACTU will send them immediately to the tax authority; the alternative mode will retain them locally with additional safeguards for integrity and traceability. (Spanish Tax Agency) At the same time, the agency analyzes growing volumes of tax and financial information with risk-analysis tools that help officials identify cases for possible review, but do not independently decide whom to audit. (Spanish Tax Agency) At work, monitoring software measures keystrokes, screenshots, response times, and absences from the desk with a thoroughness no supervisor ever had. Online, automated moderation enforces terms of service at planetary scale, closing accounts without a human ever looking at the case. Each of these advances, taken alone, looks like an efficiency gain. Together they are something else: the simultaneous closing, on every front, of the gap between the written rule and the applied rule — which was where, without knowing it, much of our everyday freedom lived.

Here it's worth running a thought experiment, because technology is turning it into a practical question. Imagine that for one month, one hundred percent of all rules in force were enforced. Not a comma needs changing: just enforce. Every driver would receive several fines a day. A large share of tax returns would be sanctionable for some error. Most companies would be fined for some breach of data protection, workplace safety, or sector regulation. Employees would rack up disciplinary infractions under their own internal handbooks. An enormous fraction of internet users would see some account suspended. Almost no one — not the diligent citizen, not the well-run company — would finish the month with a clean record.

That experiment reveals something we prefer not to look at: our rules are written with the structure of the rules of an authoritarian state. I'm not saying we live in one. I'm saying the form is the same: a standard almost nobody can meet however hard they try, combined with discretionary enforcement. That combination turns practically the entire population into latent offenders and leaves it to power to decide whom to punish. In authoritarian regimes this is a deliberate tool. A Soviet record from June 1940 documents the logic: when Lithuania's foreign minister objected that there was no legal basis to prosecute two officials, Vyacheslav Molotov replied that they should be arrested first and that “the articles will be found.” (Russian Academy of Sciences) In democracies it was a harmless inertia, because total enforcement was physically impossible. Impossibility was the insurance. Technology is canceling the policy.

“Impossibility was the insurance. Technology is canceling the policy.”

It's worth saying, too, why rules get written this way, because it isn't just carelessness. Many rules are not drafted to be followed: they are drafted to transfer liability. The eighty-page employee handbook exists so the organization can point to it when something goes wrong. The exhaustive compliance policy exists to protect its author. The unreadable terms of service exist to cover the platform in every scenario. These are rules that presuppose their own violation: their function is not to order conduct but to allocate blame when conduct, inevitably, strays. That trick worked as long as no one could verify compliance. Perfect enforcement blows up this fiction too: once everything is recorded, the liability shield becomes an arsenal of charges against everyone underneath it.

What happens when total enforcement becomes cheap? Three broad paths open up.

The first is to enforce everything: every infraction, always, against everyone, in every domain. It is unworkable. An entire society sanctioned daily for behaving the way everyone has always behaved would generate a backlash no democratic system could absorb. The rules would lose the only thing that makes them work without violence: the perception that they are reasonable. A power that punishes normal behavior is not perceived as justice but as oppression, and perceived oppression at the scale of a whole society stabilizes nothing.

The second path is the one being taken by default, precisely because it requires deciding nothing: keep the capacity to sanction everything, but choose when to use it. It looks like continuity — after all, there was always discretion — but it is a profound mutation. The discretion of the officer, the inspector, or the supervisor is imperfect and it too discriminates, but it is local, contextual, and contestable: a specific human saw a specific situation and decided in the moment, and you can argue with that decision to their face. Centralized discretion is something else. The decision of whom to punish is made far from the facts, on criteria no one can see, over a total archive of infractions that includes nearly everyone. Think about what that means in each domain. A government that has a recorded infraction on every citizen and every company can turn the inspection into a weapon: against the critic, against a friend's competitor, against the inconvenient industry. An employer with a recorded infraction on every employee holds, in practice, a permanently preloaded for-cause dismissal: it doesn't need to use it; it's enough that everyone knows it exists to extract an obedience that goes far beyond the rules. And unlike the bias of an officer on the road, this bias is systemic, opaque, and scalable. The universal latent file is not a regulator: it is an instrument of discrimination and opportunism waiting for its first unscrupulous user.

The third path is the only one compatible with a free society: if technology is going to enforce the rules one hundred percent of the time, the rules have to be written to withstand it. Which means: the normal, reasonable conduct of a diligent person or organization cannot constitute an infraction. Rules should describe the boundary of the tolerable, not the ideal of the perfect.

“Rules should describe the boundary of the tolerable, not the ideal of the perfect.”

Someone will object that intermediate paths exist: tolerance margins programmed into speed cameras, warnings before sanctions, thresholds below which no action is taken. They exist, and they are the best news in the whole argument. A written, public tolerance margin is not a fourth path: it is the third path in miniature. It is the tolerance that used to live, invisibly, in the officer's judgment, moved into the text of the rule, where everyone can see it, debate it, and demand it. Every time an automated system incorporates an explicit margin, it is conceding the entire thesis in miniature: that the rule as written was never meant to be applied literally, and that the honest fix is to rewrite it — not to administer its violation in the shadows.

From this comes a principle that seems to me the most useful conclusion of the whole line of reasoning. Call it the perfect enforcement test: a rule is legitimate only if we would accept it being enforced every single time it is broken. If perfectly enforcing a rule would produce an intolerable result, the problem is not the enforcement. It is the rule. And the test applies to every rule-making power, not just parliaments. It applies to the regulator drafting a technical standard, to the company drafting its internal handbook, to the platform drafting its terms of use. Whoever is unwilling to see their rule enforced one hundred percent of the time is confessing they wrote it for something else.

“If perfectly enforcing a rule would produce an intolerable result, the problem is not the enforcement. It is the rule.”

Note that this test is not an argument against surveillance or against hard rules. Some rules pass it easily. Nobody objects to detecting one hundred percent of drunk drivers, of workplace harassers, of deliberate fraudsters: those rules describe conduct we want to eradicate entirely, and perfect enforcement strengthens them. The test doesn't weaken good rules; it distinguishes them from the ones that survived only because they weren't enforced. A body of rules that passes the test is harder and more credible at the same time, because every rule in it means what it says.

The problem is that this third path will not happen on its own, because the asymmetry of incentives pushes the other way. A legislator can be blamed for the tragedy his law failed to prevent; no one will ever blame him for the aggregate cost of the laws he passed. Passing a rule is a visible act with concrete beneficiaries who are grateful for it — the protected group, the agency that will administer it, the headline; repealing one is a nearly invisible act whose benefits scatter across millions who will never know, and whose risks concentrate on whoever signs. Inside organizations the same asymmetry operates in miniature: the page added to the handbook protects its author if something goes wrong; the page removed never does. And it's not that deregulatory politics doesn't exist — it periodically wins elections under that very banner — but here is the telling fact: even under governments that promise to prune, the total stock of rules barely recedes, and by the next term it is growing again. A ratchet that survives its declared enemies is not an ideology. It is a structure.

The result is measurable. In Spain — to take the country I know best, though the pattern is Western-wide — the number of rules published each year has roughly quadrupled since the return of democracy; the country's public administrations approved 411,804 rules between 1979 and 2021. (Bank of Spain) In 2024, Spain's national and regional official gazettes published 1,298,086 pages. Keeping up with all of them would require reading about 3,557 pages a day; even limiting the task to the national gazette would mean about 720. (CEOE) Against the full body of rules in force, any serious estimate of reading it all is measured in human lifetimes, not years. The principle that ignorance of the law is no excuse presupposes that the law can be known. We have manufactured a normative universe no human being can know, and we are building machines increasingly capable of watching all of it. It is hard to imagine a more dangerous asymmetry.

That is why the solution is not a technical tweak but a change of incentives. Sunset clauses that force every rule to rejustify its existence or die. Regulatory budgets that forbid adding without removing. Independent evaluation of every rule's effectiveness, with consequences for whoever regulates badly — inside companies too, where an uncompliable handbook should be treated as a liability, not a shield. And, as both an entry filter and a review standard for the existing stock, the perfect enforcement test: no rule should be passed or kept if its complete enforcement by technology would produce a result we are not prepared to accept. There is even an irony available here: the same artificial intelligence that closes the enforcement gap is the first tool in history capable of reading the million pages, detecting the redundancies and contradictions, and auditing the entire body of rules. The technology that creates the problem is also what makes the solution materially possible. What's missing is not capacity. It's will — and will follows incentives.

“The technology that creates the problem is also what makes the solution materially possible.”

That leaves the question of urgency, and here there is no point sugarcoating. This is arriving when trust in national governments remains fragile across many advanced democracies: in the OECD survey conducted in 2025, 40% reported high or moderately high trust, while 43% reported low or no trust. (OECD) Governments and organizations were already using technology to watch, pursue, and sanction with growing efficiency, and the diffuse fear — of the next fine, the next inspection, the next disciplinary file, the next ban, of “what will they come up with now” — is already a daily worry for many people. Artificial intelligence does not create that fear: it industrializes it, and it industrializes it in every domain of life at the same time. A citizenry that knows itself to be permanently in breach — before its government, before its employer, before the platforms it depends on — is not a stable citizenry. And history offers little comfort about where the sustained perception of arbitrariness tends to lead: to delegitimization, from delegitimization to destabilization, and from destabilized democracies, too many times, to conflict.

None of that is inevitable. But avoiding it requires naming the fiction before technology blows it apart: our rules — all of them, the ones in the official gazette and the ones in the company handbook — were written to be enforced sometimes, and we are building machines that enforce them always. Either we rewrite the rules to the standard of perfect enforcement, or perfect enforcement will reveal — in the worst possible way — what kind of rules we had written. The time to choose is now, while the fiction still holds.