What If Every Law Were Enforced, Every Time, Against Everyone? The Perfect Enforcement Test

"The law is the law." "If you don't want a fine, follow the rules." "If you have nothing to hide, you have nothing to fear." Almost all of us have said something like this, and the idea behind it looks flawless: if a law exists, it should be enforced always, against everyone, in every case.

This text takes that idea seriously and carries it to the end as a thought experiment. Automated surveillance and artificial intelligence are bringing the experiment closer to reality, so it deserves a more rigorous examination than an opinion. An earlier version of this article argued that enforcing everything would bring chaos first and arbitrariness second. Under criticism, that conclusion turned out to be only part of the truth, and a badly framed part. What follows is the reconstruction.

Summary

The thesis. Perfect enforcement is not the law brought to its fullness but a change of regime. The law orders its own application always, yet its penalties, procedures, and mechanisms of correction were calibrated for a world where enforcing it was expensive and scarce. That scarcity did work nobody ever assigned it in writing. Remove it, and the same written law produces a different system.

The result. Raising detection works very well against deliberate conduct, and its defenders are right about that. But it leaves a residual no deterrence can remove, and that residual brings a trilemma: you cannot have every violation carry a consequence, every consequence open to challenge, and finite resources all at once. Give up the first and you get selection; the second, automated injustice; the third, collapse. The main danger is not noisy chaos but silent injustice, plus an archive that gives power over anyone.

The proposal. A rule of conservation of severity: the more complete the detection, the lighter the default consequence; the graver the consequence, the more human the process; and what is detected and not used must be forgotten. A six-criterion test then shows which laws can move into a high-detection regime.

The figures are interactive. Unless a source is stated, they are illustrative simulations: they teach a mechanism, they do not measure the world.

Table of contents

1. Defining the experiment

What "enforce" means

The phrase "enforce every law" seems clear until you try to program it. It blends at least four quantities that can vary independently:

  • Detection: what share of violations becomes known.
  • Sanction: what share of those detected gets a consequence.
  • Rigidity: how much the concrete case is looked at when deciding.
  • Review: how far the result can be challenged.

Most of the confusion in this debate comes from jumping between them without warning. A system can reach every case and apply equity in each, or reach few and be merciless. The usual examples don't measure the same thing either: some prove error, others rigidity, others indeterminacy. And one distinction carries everything else: detecting everything and sanctioning little is a very different world from detecting everything and sanctioning everything.

Figure 1

Enforcement is not one quantity but at least two

Crossing detection and sanction gives four regimes. The thought experiment lives at top right. The most dangerous one, at bottom right, almost never appears in the debate.

Which laws count

"Every law" doesn't mean what it seems to either. Hart separated rules that impose duties from rules that confer powers, such as those governing a contract or a will, which are not "enforced" on anyone by surprise. This thesis is limited to rules of conduct backed by a public sanction.

Within that domain lies a paradox the defender of total enforcement rarely considers. "Every law" includes the laws that limit the enforcer: data protection, the inviolability of the home, the presumption of innocence, statutes of limitations. Applied with equal rigor, they cut detection. The EU's data protection regulation restricts decisions based solely on automated processing (GDPR, art. 22). So literally total enforcement contradicts itself: you cannot apply every rule of conduct and also every rule protecting people from its application.

A position that isn't absurd

Spanish law, the example I know in most detail, doesn't support the comfortable idea that laws are made to be enforced "sometimes." On criminal matters it formally says the opposite. Prosecutors are obliged to bring the criminal actions they consider proper (Spanish Criminal Procedure Act, art. 105), and an authority who intentionally fails to pursue crimes commits one (Spanish Criminal Code, art. 408). The earlier version of this article said laws "were written to be enforced sometimes." As a claim about their text, that was false.

The correct claim is different. The law orders totality, while budgets and the calibration of its sanctions assume the opposite. George Stigler put it in 1970: society must forgo complete enforcement, because enforcement is costly (Stigler, 1970). That gap between the mandate and its calibration is the real subject of this thesis.

2. What strict enforcement gets right

Those who defend rigorous enforcement are more right than their critics usually admit.

Equality. A law that falls on the weak and forgives the powerful is an injustice. And leniency is the most opaque form of power: the decision not to act leaves almost no trace and is rarely reviewed. Joseph Goldstein made this point in 1960 about low-visibility police decisions (Goldstein, 1960). An automated log, by contrast, can be audited.

Deterrence. The likelihood of being caught matters far more than the harshness of the punishment (National Institute of Justice). And it works. The Cochrane review of speed cameras found, near the cameras, reductions of 11% to 44% in fatal and serious-injury crashes (Wilson et al., 2010).

The tax evidence. According to the IRS, net misreporting in the United States is 1% on income with withholding and substantial third-party reporting, and 55% on income with little or no third-party reporting (IRS). Being seen changes what you do.

Pressure on bad laws. Ulysses S. Grant said in 1869 that he knew no method to secure the repeal of bad laws so effective as their stringent execution (Grant, 1869). It is this position's best argument: universal enforcement makes visible a cost that is spread silently today, and so forces politics to respond.

Any serious critique has to explain why, given all this, the conclusion isn't "more enforcement, always." The following sections try to do so through eight hypotheses.

3. H1. Non-compliance is massive, but uneven

Hypothesis. With enough exposure and regulatory detail, even someone who almost always complies ends up in breach.

The arithmetic is relentless. If someone performs ten regulated acts a day and complies with 99.9% of them, the probability of a full year without a single violation is 2.6%.

Figure 2 · H1

Complying almost always isn't enough to never breach

Each dot is a person. The probability of a clean record is compliance raised to the number of acts. Illustrative model assuming independent errors.

Traffic has data. Spain's DGT estimates that in free-flowing traffic about 60% of vehicles exceed the limit on conventional roads, and just under 40% on dual carriageways (DGT, 2026). This is no antisocial minority.

Two corrections limit the hypothesis. First, the evidence is good in high-exposure fields such as traffic or tax, and much weaker elsewhere. Claiming "almost everyone" breaches in every field is an extrapolation. Second, and more important, the earlier version assumed each lapse would end in a sanction, which forgets that the law already contains filters. Spanish administrative law sanctions only those liable through intent or fault (Law 40/2015, art. 28). In tax, there is no penalty if due diligence was exercised, for instance through a reasonable interpretation of the rule (General Tax Act, art. 179.2.d).

The thought experiment must therefore choose. Either it applies the letter without those filters, and then it isn't applying "the law," or it keeps them, and then someone must decide case by case whether there was fault. And deciding is precisely what doesn't scale.

4. H2. Enforcing requires deciding

Hypothesis. Part of every rule stays undetermined until someone applies it, and automation doesn't remove that decision. It moves it to whoever writes the code.

Hart called this the open texture of law: words have a clear core and a penumbra (Hart, 1961). Frederick Schauer added that rules are entrenched generalizations, doomed to reach too far or not far enough (Schauer, 1991). No drafting fixes this entirely.

The surprise is that it happens even with the simplest rule. Researchers recorded a car's actual speed over one hour and asked 52 programmers to automate the speed limit, each working alone. With the same data and the same rule, their programs issued anywhere from no tickets to one per reading above the limit (Shay et al., 2016). There was also a telling detail: on hills, the car exceeded the limit with cruise control set exactly at it (Practical Ethics).

Figure 3 · H2

Same trip, same law, a different number of tickets

Simulated trip, not the study's data. The literal setting gives 266 tickets. Counting by episode leaves 21, and a 7 km/h margin with 30 seconds of grace leaves none. The text of the rule doesn't choose between them.

This experiment needs fair treatment. The programmers had no legal training, and a subgroup given a legally reviewed specification deviated less. That's a point for those demanding strict rules: precision helps. What it doesn't do is bring the deviation to zero. The correct conclusion is that a residual indeterminacy remains and someone always resolves it. The question is whether that's a legislator in public or a programmer nobody sees.

The work-to-rule strike, which the earlier version offered as proof of chaos, shows no such thing. There nobody breaches and here almost everyone does, so they are opposite worlds. What it does show is something else: Spain's Royal Decree-Law 17/1977 lists such strikes among "unlawful or abusive" acts (BOE), and James C. Scott made the same point: any organization depends on practices its regulations don't describe (Scott, 1998). Rules presuppose good faith and interpretation. They don't comply by themselves.

5. H3. Sanctions were calibrated for scarcity

Hypothesis. Raising detection without changing sanctions is a penalty increase nobody voted for.

This is probably the most important idea missing from the earlier version. Law and economics has had it since Gary Becker: what deters is the expected sanction, the probability times the amount (Becker, 1968; Polinsky and Shavell, 2000). If a €200 fine is applied once in a hundred violations, its expected sanction is €2. If it's applied every time, it's €200. The law hasn't changed a comma and the system is a hundred times harsher.

Figure 4 · H3

The same fine, a hundred times harsher

Expected sanction is probability times amount. The 1% starting point is illustrative. The last figure shows the fine that would keep today's severity.

This produces a serious objection to my earlier version, which always located the problem in the rule. An economist would answer that often the rule is fine and what must change is the amount. That's right. Hartzog and coauthors defend something close: many laws were drafted assuming leniency (Hartzog et al., 2015).

Rescaling has two limits. Not every punishment is divisible: license points, criminal records, and dismissal can't be divided by a hundred. And a flat fine weighs very differently depending on income, so at full certainty it becomes regressive. Spanish criminal law already knows the partial solution, the day-fine, with a daily amount tied to economic circumstances (Spanish Criminal Code, art. 50). Total enforcement would require taking it much further.

6. H4. People adapt

Hypothesis. Total enforcement sharply reduces deliberate violations, but leaves a residual no deterrence removes. The whole problem lies in that residual.

This was the earlier version's gravest error. It conceded that certainty deters and then computed fines as if nobody changed their conduct. That's a snapshot of day one. The experiment must look at the equilibrium, and in equilibrium most people would have learned.

Two kinds of violation must be separated. Deterrable ones respond to the probability of being seen: the phone at the wheel, hidden income. Residual ones don't: error, ignorance of a rule nobody can know, instrument imprecision, necessity, mutually incompatible rules. Nobody stops making mistakes because they are watched.

Figure 5 · H4

What remains once deterrence has done all its work

Illustrative model. The size of the residual is unknown and varies by field. It is the empirical variable almost this entire thesis depends on.

Certainty isn't magic either. Hawaii's HOPE program, with swift, certain, mild sanctions, cut arrests from 47% to 21% in its original trial (Hawken and Kleiman, 2009). But the four-site replication, with more than 1,500 people, found no difference in arrests, except for drug arrests at one site (Lattimore et al., 2016). And on domestic violence, the Milwaukee experiment found that arrest had different effects depending on the profile, and could make things worse among the unemployed (Sherman et al., 1992). Sherman himself theorized the mechanism: a sanction perceived as unfair breeds defiance, not obedience (Sherman, 1993).

The conclusion is that defenders of strict enforcement are right about the deterrable part and wrong about the residual. That gives the first criterion of the whole analysis: for a law with a small residual, raising detection is nearly all upside. For one with a large residual, raising it punishes diligent people.

7. H5. The trilemma of total enforcement

Hypothesis. On the residual, three things cannot be kept at once: every violation having a consequence, every consequence being open to challenge, and resources being finite.

This replaces the "first chaos, then arbitrariness" sequence, which had a logical flaw. If you suppose that everything is enforced, you can't conclude that someone chooses, because choosing denies the supposition. There are, instead, three alternative exits.

Figure 6 · H5

Three desirable properties; only two fit

Pick the property you give up. Each sacrifice produces a different regime with real examples.

Giving up limits: collapse

If you want completeness with safeguards, the numbers don't work. Spain had 28,142,469 licensed drivers at the end of 2024 (Vozpópuli, citing DGT data), and the DGT issued 6.1 million reports in 2025 (Xataka). One fine per driver per year would mean 28 million. Spain's administrative courts took in 198,961 cases in 2025 and resolved 212,680 (CGPJ).

Figure 7 · H5

How long it takes the appeal route to break

Total and resolved cases from Spain's CGPJ for 2025. The appeal rate is an assumption: I know of no reliable public figure, and early-payment discounts cut it sharply. Move the second control to see how much depends on it.

At a 1% appeal rate there would be about 281,000 new lawsuits a year, and in five years the queue would pass a million and a half cases. At 0.1% there would be 28,000, around 14% of current intake: a notable load, not a collapse. The earlier version showed only the first number. The honest conclusion is conditional: collapse comes only if challenging is easy. And there lies the trap, because today it usually isn't.

Giving up challenge: automated injustice

This is the likeliest exit and the one the earlier version described worst. There, MiDAS and the Dutch case appeared as examples of chaos, when those systems ran with terrible smoothness. The problem wasn't a jam. It was that there was no jam.

In Michigan, a review found that of 40,195 unemployment-fraud cases decided by the algorithm alone, 85% were incorrect determinations (IEEE Spectrum). In Australia, the Robodebt scheme raised at least A$1.763 billion in debts against about 433,000 people (The Mandarin) and ended in a royal commission. In the Netherlands, the administrative division of the Council of State upheld an "all-or-nothing" line until October 2019: parents who could not account for a small part of their advance had to repay all of it (Advocatie). Mark Rutte's government resigned over the scandal in January 2021 (IamExpat).

That last case is above all an example of rigidity, not coverage. And it has an important nuance: the court changed its line without the law changing (Dutch Council of State). The rule wasn't compelling the result; a strict interpretation was. Whoever demands "the law, period" is also choosing an interpretation.

Why is there no jam? Because almost nobody fights. Pamela Herd and Donald Moynihan have documented that administrative burdens fall hardest on those with the fewest resources to bear them (Herd and Moynihan, 2018). Paying usually costs less than appealing. The system reads silence as accuracy.

Scale adds a mathematical problem, and here H4 turns against itself. Suppose a system that catches 95% of violators and wrongly flags one innocent act in a thousand. If 1% of acts are violations, about 9 of every 100 accused are innocent. But if deterrence works and only one act in a thousand is a violation, more than half of the accused are innocent.

Figure 8 · H5

The better the system deters, the worse it accuses

Bayes' rule with 95% sensitivity. With high compliance, most alarms can be false even if the detector is excellent. Illustrative parameters.

I had not seen this consequence stated in any of the sources, and it strikes me as central: the success of deterrence degrades the reliability of accusation. A high-detection system needs more review as it works better, not less.

I shouldn't overstate this. The honest comparison is with human error, which also exists and isn't always measured. And false negatives do harm too: an unprotected victim is a real cost. The objection isn't to automating but to automating without review.

Giving up completeness: selection

The third exit is the world we already live in. With safeguards and finite resources, only part gets pursued. Michael Lipsky showed that front-line officials make de facto policy through that choice (Lipsky, 1980).

The earlier version made another error here: it treated selection and arbitrariness as synonyms. They aren't. Scarcity entails selection, but a public lottery or a published ranking of severity is selection without arbitrariness. Arbitrariness needs three more ingredients: a hidden criterion, an unreviewable decision, and room for improper motive. Kenneth Culp Davis proposed over half a century ago what to do with discretion: confine it, structure it, and check it (Davis, 1969).

Figure 9 · H5

Everyone breached. Who gets sanctioned?

A hundred people with the same violation and capacity to process only a few. Non-compliance doesn't change between rules; only who pays does. Illustrative simulation.

8. H6. The archive

Hypothesis. The most dangerous regime isn't sanctioning everything but seeing everything and sanctioning little. That alone creates power over anyone.

Many of the earlier version's predictions actually concerned this regime, not universal enforcement. When almost everyone has a recorded violation, the question stops being who breached. It becomes whom it is convenient to pursue.

Philip Pettit explains why this matters even if nobody is sanctioned: you can lack freedom without suffering any interference, if someone else holds the capacity to interfere at will. The kindly master is still a master (Pettit, 1997). That's why the earlier version's central sentence needs clarifying.

Until now we have been free, in part, because power could not see us. Better laws don't fix that. A perfect law under total surveillance still leaves the archive intact, available to whoever comes next.

What technology changes isn't the existence of this abuse. A Soviet record from June 1940 documents it with no machine at all: 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). What changes is the marginal cost. People once had to be assigned to investigate someone until they found something. With a total archive, a query is enough.

The defense is weak for two reasons. The first is legal. Spain's Constitution prohibits arbitrariness by public authorities (Constitution, art. 9.3), yet its Constitutional Court has repeatedly held that there is no right to equality in illegality (Constitutional Court). Proving that others aren't sanctioned generally won't save you, and proving misuse of power requires knowing a criterion you won't see.

The second reason is that the archive isn't neutral: it records best what is cheapest to see. James C. Scott called legibility what a state needs in order to act on a population (Scott, 1998), and Virginia Eubanks documented how automated systems profile and punish the poor above all (Eubanks, 2018). The IRS figure can be read this way too: a salaried worker with withholding can hardly breach; someone with income no third party reports can. Perfect enforcement doesn't reach everyone at once. It reaches first those who were already visible. So the equality argument flips: raising detection without symmetry increases inequality before the law.

The law does have an old antidote: the limitation period. Spain's Workers' Statute bars very serious employee misconduct sixty days after the company learns of it, and in any case after six months (Workers' Statute, art. 60). The earlier version spoke of a "permanently preloaded" dismissal, and that rule largely refutes it. But it suggests the general solution: forgetting is a safeguard. Limitation periods were how the law prevented the archive when archives were made of paper.

9. H7. Excess caution and the dilemma of the margin

Hypothesis. Under high detection, people don't only stop breaching: they also abandon lawful conduct near the line. And publishing a margin doesn't remove the problem, it moves it.

Craswell and Calfee showed that uncertainty about the legal standard can produce over-compliance, because extra care lowers the chance of being held liable (Calfee and Craswell, 1985). If the rule is vague and detection total, the safe place is far from the line. Jonathon Penney observed something similar elsewhere: visits to Wikipedia articles on terrorism fell by 20% to 30%, depending on specification, after the Snowden revelations (Penney, 2016). Nobody had prohibited reading them.

The earlier version proposed publishing margins as a clean solution. It isn't one. A public margin becomes the new limit. On German motorways, Traxler and coauthors found drivers bunching just below the thresholds where the penalty jumps (Traxler et al., 2018). And Meir Dan-Cohen raised the subtlest objection: law may work precisely because the rule citizens hear and the rule judges apply differ. He called this acoustic separation (Dan-Cohen, 1984). Publishing tolerance destroys that separation.

Figure 10 · H7

Where people place themselves against the line

The gray curve is the starting point. With a clear rule people hug the line; with a vague rule they flee it; with a public margin they hug the margin. Illustrative curves.

My answer is that acoustic separation depends on scarcity. It works while decisions are few and barely visible. With high detection and millions of rulings, the real rule can be inferred from experience, and those with advisors will infer it first. The choice is no longer between a hidden margin and a public one, but between a margin everybody knows and one only the sophisticated know. The cost of bunching at the threshold is real, but the threshold can be set accordingly and the sanction graduated. Louis Kaplow offers the general criterion: rules cost more to create and standards cost more to apply (Kaplow, 1992). High detection multiplies applications, so it pushes toward precise rules.

10. H8. The law needs room to be wrong

Hypothesis. Perfect enforcement freezes the law, because part of its correction depends on some people being able to breach it.

This is the most forgotten consequence. Many reforms aren't born in parliament but from a non-compliance that becomes impossible to ignore. In Lawrence v. Texas, the US Supreme Court took into account that sodomy laws were in retreat and showed a pattern of nonenforcement (Lawrence v. Texas, 2003). In Spain, refusal of military service grew from a few hundred young men in the late 1980s to several thousand in the 1990s, by its organizers' own count, before service was suspended effective December 31, 2001 (Royal Decree 247/2001). It wasn't the only cause, but it was part of the process.

Mark Granovetter explained the general mechanism: many collective behaviors depend on thresholds, with each person acting only once others have (Granovetter, 1978). If the first movers are removed one by one, the next ones never act.

Figure 11 · H8

A cascade needs its first movers to survive

Illustrative threshold model. With low detection, disobedience spreads and forces reform. Above a certain point it never reaches the threshold. The model doesn't distinguish just causes from unjust ones.

This hypothesis has an extreme form. Michael Rich coined "impossibility structures" for designs that don't punish a violation but prevent it (Rich, 2013), and Jonathan Zittrain warned what is lost with perfect enforcement in digital environments (Zittrain, 2008). A speed limiter or software that refuses an incorrect invoice sanctions nobody, which dissolves the trilemma, but it also removes the option of breaching for a good reason.

And Grant, who held that enforcing a bad law is the best way to repeal it? He is right in one case. Enforcement forces reform when the cost falls on a majority or on those with power. If it falls on a minority, the majority feels nothing and the law survives. The two intuitions reconcile on one fact: who bears the load matters. That's why symmetry isn't a detail but the condition for strict enforcement to correct itself.

11. Objections

"If you don't want discretion, defend rules." Agreed in part. This thesis prefers precise rules as detection rises. But Schauer shows that over- and under-inclusion don't vanish, so a final decision will always remain.

"AI will improve and error will vanish." Measurement error may fall. H2's indeterminacy and H4's residual aren't measurement errors. And Figure 8 shows reliability also depends on the true violation rate, which falls when the system works.

"A human in the loop solves it." Only if review is real. Danielle Citron warned almost twenty years ago about the tendency to trust what the machine says (Citron, 2008). A human signing thousands of decisions a day is an alibi, not a safeguard.

"An impartial machine would be more legitimate." Perhaps. Tyler showed that legitimacy influences compliance independently of fear (Tyler, 1990), and Paternoster and coauthors that those who feel fairly treated reoffend less (Paternoster et al., 1997). But those works concern fair procedures, not good outcomes. We don't know whether people will perceive an automated decision as fair. The OECD survey points to a reservation: fewer people are confident that governments will ensure adequate safeguards when using AI (OECD). The question remains open.

"Fewer laws means less protection." True, and it corrects another error in my earlier version, which asked for "fewer laws." Thresholds, exceptions, and graduated sanctions add text. Many rules protect those without power. The goal is to cut burden and residual, not the number of rules.

"Your remedies worsen your problem." This is the most serious internal objection. Demanding real review and a simple appeal raises per-case cost and can overload the system. So the proposal can't be more safeguards for everything. It has to be that safeguards depend on what's at stake, and that most cases never become a sanction.

12. Synthesis: conserving severity

The eight hypotheses point to one rule with three parts:

  1. The default consequence must fall as detection rises. If detection multiplies by a hundred, expected severity shouldn't do so without an explicit vote.
  2. Process must grow with consequence. In Mathews v. Eldridge, the US Supreme Court held that safeguards depend on the interest affected, the risk of error, and the cost (Mathews v. Eldridge, 1976). An automated notice doesn't need a trial. A conviction does.
  3. What is seen and not used must be forgotten. Limitation periods should cover data too: what is detected without action within a short period gets deleted.

This also breaks the trilemma, by changing one word. A detected violation doesn't need a sanction: a response is enough. Ian Ayres and John Braithwaite proposed an enforcement pyramid: start with soft measures and escalate only when necessary, with a credible threat at the top (Ayres and Braithwaite, 1992). Bardach and Kagan reached a kindred conclusion from studying rigid regulation (Bardach and Kagan, 1982).

Figure 12 · Synthesis

Total response, exceptional sanction

Every detected violation gets a response, and only a fraction reaches a sanction. Press a level to see its safeguards. Illustrative escalation rates.

The pyramid has a cost too: someone decides who escalates, so discretion returns. That's unavoidable, and this thesis accepts it. No rule eliminates judgment. What can be done is to narrow it, log it, and put it under control, as Davis asked.

13. The perfect enforcement test

The earlier version said a law is defensible only if we would accept the result of enforcing it always. The formula had three defects: it didn't say who accepts or by what standard, a bad law against a minority could pass it, and good laws failed it through a single tragic case. I restate it like this:

A legal system is ready for a high-detection regime only if, measured at its equilibrium after conduct adapts, it meets six criteria.

Three clarifications. The test evaluates the system, meaning the rule with its sanction, procedure, and capacity, not the rule alone. It's measured at equilibrium and in the mildest case, not the worst or the average. And a public body must apply it and give reasons, not each person's intuition.

Figure 13 · The test

Six questions before saying "just enforce it"

Pick a case and change the answers. The starting assessments are mine and debatable: the test forces you to reason them, it doesn't settle them.

The criteria answer to each hypothesis. Avoidability (H1 and H4): the residual must be small for a diligent person. Determinacy (H2): two honest enforcers must agree. Proportion (H3): the sanction must be acceptable at real detection and in the mildest case. Contestability (H5): it must be challengeable at a cost fitting the stakes. Incidence (H6): it must reach the visible and the hidden alike. Reversibility (H8): the chance to correct the law must survive.

What this test doesn't do must be said.

It doesn't measure justice. It's a necessary condition for raising detection, not proof that a law is good. A well-drafted unjust law could pass several criteria. Reversibility and incidence make that harder without excluding it. Lon Fuller separated law's internal morality from its content (Fuller, 1964), and this test belongs to the former: among his requirements were that rules be practicable and that official action be congruent with the declared rule.

Failing it doesn't delegitimize the law. Domestic violence or juvenile theft may fail because an automated response would harm the very people the law seeks to protect. The verdict isn't that those laws are surplus but that they must stay in the hands of people with judgment.

It doesn't eliminate discretion. It only narrows it.

Easy cases aren't so easy either. I once wrote that drunk driving clears these questions "easily." That's almost true, but someone driving the morning after without knowing they're still over the limit tests avoidability. The correct formula is that it passes with a far better grade than almost any other rule, not that it's beyond discussion.

14. What would have to change

Why this doesn't happen on its own is well known. William Stuntz described in criminal law a logic that pushes toward piling up rules (Stuntz, 2001), and the asymmetry is general: passing a rule is visible and repealing one nearly invisible. In Spain, the number of rules published each year has quadrupled since the return of democracy, and administrations approved 411,804 between 1979 and 2021 (Bank of Spain). In 2024 official gazettes published 1,298,086 pages (CEOE). Not all of those are duties of conduct, since they include notices and appointments, and part of the figure reflects a decentralized state. Yet the Civil Code still says that ignorance of the laws does not excuse compliance (Civil Code, art. 6.1), a rule that presupposes laws can be known.

The tools half exist. Spain's Law 39/2015 asks administrations to review their rules periodically (Law 39/2015, art. 130), although Constitutional Court ruling 55/2018 limited its reach (BOE). Law 40/2015 requires naming who answers for an automated system (Law 40/2015, art. 41). The EU AI Act's high-risk obligations, human oversight included, were postponed to December 2027 (European Commission). What's missing is connecting them in one rule:

  • Detection impact statement. Before multiplying detection of a violation, the regulator publishes its estimated residual, expected sanction before and after, and review capacity.
  • Automatic recalibration. If detection rises, the default sanction falls or becomes a notice, absent an express vote.
  • Income-proportional sanctions in any high-detection system.
  • Mandatory pyramid. A corrective first response for violations with no intent or serious risk.
  • Public, auditable selection criteria, with random sampling to measure the residual and legibility bias.
  • Limitation periods for data. What is detected and not acted on gets deleted within a short period.
  • Symmetry. Detection doesn't rise for one group without a plan for the less visible.
  • Periodic review and independent evaluation, including handbooks and terms of use where they act as private law.

The same artificial intelligence that brings total enforcement closer can read millions of pages, flag contradictions, and help estimate which rules have the largest residual, always with lawyers verifying the result.

15. How this thesis could be refuted

A thesis must be able to be wrong. This one would be if the following were observed:

  1. Negligible residual. If trials show that almost every violation vanishes under high detection in complex rules, H4 loses force and the strict-enforcement advocate wins.
  2. No rise in felt severity. If raising detection without touching sanctions doesn't raise the real burden, especially at low incomes, H3 is irrelevant.
  3. Stable precision. If the share of false accusations doesn't rise as violations fall, because systems adapt to the base rate, the argument of Figure 8 weakens.
  4. No legibility bias. If audits with random sampling find no overrepresentation of the most visible, H6 weakens.
  5. Reform without disobedience. If in very high-detection settings unjust laws get corrected just as fast, H8 is false.

These predictions are testable in specific fields. The key variable, the size of the residual by field, can be estimated through random sampling, and it's surprising that almost nobody measures it.

16. Conclusion

If you believe the law should always be enforced, this thesis doesn't ask you to stop. It asks you to accept what that implies.

You are right that impunity corrodes, that certainty deters, and that many laws should be enforced more, above all against those nobody sees today. But the law you have in mind is not just its text. It includes fault, equity, proportion, appeal, and limitation periods. Remove all that and you aren't enforcing the law better: you are replacing it with a harsher one nobody voted for.

Perfect enforcement won't arrive all at once either, or for everyone at the same time. It will come field by field and fall first on the most visible. The risk isn't a day of chaos but growing used to a smooth system that is nearly always right, errs in silence, and keeps a file on every person.

Scarcity protected us without anyone having decided it. Now it has to be decided. If you truly want every law enforced every time, you have to demand that every law can withstand it.

References

Economics of enforcement. Becker (1968), Crime and Punishment: An Economic Approach. Stigler (1970), The Optimum Enforcement of Laws. Polinsky and Shavell (2000), The Economic Theory of Public Enforcement of Law. Calfee and Craswell (1985), Incentives to Comply with Uncertain Legal Standards. Kaplow (1992), Rules Versus Standards. Traxler, Westermaier, and Wohlschlegel (2018), Bunching on the Autobahn?.

Legal and political theory. Hart (1961; 1958 article), The Concept of Law, discussed by Schauer. Fuller (1964), The Morality of Law. Schauer (1991), Playing by the Rules. Dan-Cohen (1984), Decision Rules and Conduct Rules. Davis (1969), Discretionary Justice. Goldstein (1960), Police Discretion Not to Invoke the Criminal Process. Lipsky (1980), Street-Level Bureaucracy. Stuntz (2001), The Pathological Politics of Criminal Law. Pettit (1997), Republicanism. Scott (1998), Seeing Like a State. Granovetter (1978), Threshold Models of Collective Behavior.

Automation. Shay, Hartzog, Nelson, and Conti (2016), Do Robots Dream of Electric Laws?. Hartzog, Conti, Nelson, and Shay (2015), Inefficiently Automated Law Enforcement. Citron (2008), Technological Due Process. Rich (2013), Should We Make Crime Impossible?. Zittrain (2008), Perfect Enforcement on Tomorrow's Internet. Eubanks (2018), Automating Inequality. Herd and Moynihan (2018), Administrative Burden.

Behavior and evidence. Wilson et al. (2010), Speed cameras. National Institute of Justice, Five Things About Deterrence. Hawken and Kleiman (2009), HOPE. Lattimore et al. (2016), HOPE Demonstration Field Experiment. Sherman (1993), Defiance, Deterrence, and Irrelevance. Sherman et al. (1992), Milwaukee. Tyler (1990), Why People Obey the Law. Paternoster et al. (1997), Do Fair Procedures Matter?. Penney (2016), Chilling Effects. Ayres and Braithwaite (1992), Responsive Regulation. Bardach and Kagan (1982), Going by the Book. IRS, Tax Gap Projections.

00
Previous

Previous article

Best WordPress hosting with New Relic: 2026 comparison

Next article

Recent Country Border Changes Since 2000: Official and De Facto

Next

Explore the tools or browse interactive maps for more experiments.

Back to Blog Posts