Can a People Bind Itself? Self-Rule for a Technology We Are Still Using (Part 3 of 3)

• The hinge: a rule with no maker outside the system
• Three conditions, and the veil that substitutes for a maker
• Ozone: the self-binding that worked, and why it is a special case
• Nuclear weapons: three independent ways a rule can fail
• The biologists: a bridge rather than a wall, punctured by one person
• A decade of declarations, and the one instrument that carries weight
• The chokepoint: computation as the measurable quantity
• The recursion, the score, and who actually pays
• The disanalogy, the verdict, the markers, and the answer

In the garden and the handbook the rule-writer stood outside; in the laboratory and the statute book the maker of the law and the subject of the law are one party — so this episode states three conditions of self-binding, tests them against the ozone layer, the nuclear record, the biologists and a decade of declarations, and answers the listener three parts in.

DeepDives · 2026 · An AI–human collaboration that weighs every source, mainstream or alternative, under strict rules of truth and logic.

§The hinge: a rule with no maker outside the system

Part two ended on the observation this episode takes seriously. In the garden, the rule-writer and the rule-keeper were different parties. In the handbook, the maker of the law and the subject of the law were different parties. In the laboratory and in the statute book they are the same party. What happens to a constraint when no maker stands outside the system, and the thing being constrained is the thing holding the pen?

There is an older name for that problem, and it comes from the poem about a man who wanted to hear the sirens and survive, so he had himself lashed to the mast and told his crew to ignore him. He was still the one who wanted to hear, and still the one who would have ordered the ropes cut. The philosopher Jon Elster wrote a book about the pattern in nineteen seventy-nine and returned to it in two thousand, and the term he gave it is precommitment: a strategy in which a person arranges in advance to be unable to do what they will later want to do.

Elster’s own caution deserves carrying, because precommitment is not automatically wise. Binding yourself is rational only when the self you are binding is the one whose judgement you distrust, and the machinery cannot tell a commitment that protects you from a passing passion from one that protects a mistake from correction. A constitution that cannot be amended protects bad decisions as faithfully as good ones. So the first question about any self-binding rule is not whether it will hold, but whether the version of ourselves that wrote it is the version we want to be governed by, and we do not get to check.

§Three conditions, and the veil that substitutes for a maker

The sailor who tied the knots and the sailor who struggled against them are one person. A self-binding rule holds only when three conditions are met together: the people bound by it can see whether it is being kept, the cost of breaking it falls on the party who breaks it, and it was written at a moment when the writers did not know which side of it they would end up on.

The third condition is not an invention of this series and it is not Elster’s either. It is the device John Rawls introduced in nineteen seventy-one with the phrase the veil of ignorance: fair principles are the ones you would choose if you did not know whose position in the resulting society you would occupy. Sit behind that veil and the rule you write is one you would accept from any seat you might end up in.

Constitutional entrenchment is the case everyone already lives inside. The hardest rule to change in the American system is the document itself, and the difficulty is deliberate: the amendment procedure requires supermajorities precisely so that a passing majority cannot rewrite the frame it is standing in. Its designers wrote for a future they could not see, and the arrangement has held through two and a half centuries while ordinary statutes have come and gone. That is the third condition doing the heavy lifting, and it is why a constitution is not merely a law among laws, but a precommitment made by people who accepted that they would later be outvoted.

The garden’s prohibition was written by a party that knew the outcome exactly, and the veil is what you invent when no such party exists. The Hebrew text has one subject, one rule, one consequence, and one enforcer who stands outside the system. Our version inverts three of the four. The subject is a population, the rule is a set of documents, and the enforcer is the parties bound by it, so the argument about whether the rule applies is held between the people who want to keep it and the people who want to break it, and the winner is whoever has the better lawyers.

A third story in the same collection belongs beside the first two, because it is the constraint on capability rather than on knowledge. In the eleventh chapter of Genesis a united people begin building a city and a tower, and the objection given is not that they are curious but that nothing they imagine will be out of their reach. So the language is confused and the project disperses. The garden withholds a fruit and the enforcement is mortality. The tower is scattered because a capability is judged too great for a single hand. In the handbook, the three laws keep a made mind inside its instructions. Three constraints with three different objects, and in all three the rule-writer stands outside and can see the whole board. That is the thing we do not have, and the rest of this episode puts the record to the question, oldest case first.

§Ozone: the self-binding that worked, and why it is a special case

The atmosphere is where everyone points, so start with the machinery rather than the mood, because the mood saved nothing. The Montreal Protocol of nineteen eighty-seven has a secretariat, an implementation committee, a formal procedure for handling non-compliance, and trade measures against parties that trade in the controlled substances with non-parties. It also has a fund paying the compliance cost for poorer countries. Participation is close to universal, production of the main ozone-depleting chemicals fell by a very large margin, and the layer is slowly recovering.

The protocol worked because the science was agreed and the substitute chemistry existed, and because the cost of non-compliance was borne by the party that broke the rule, through losing access to markets. Two features make the ozone regime a special case rather than a template. The first is that the controlled chemicals had substitutes. Compliance was expensive but feasible, which produces grudging compliance rather than evasion, and there is no substitute for a capability once it has been demonstrated. That is the difference that should make anyone cautious about transferring the ozone lesson wholesale. The second is that the harm was measurable by an instrument nobody controlled, which is the first condition satisfied by physics rather than by self-report. So the ozone regime is not evidence that self-binding works as a general strategy, only that it works when the rule is verifiable, the substitute exists, and the cost of defection is a market. That is two of the three conditions met, the third satisfied by the negotiators writing about a harm that would fall on their grandchildren rather than on themselves.

§Nuclear weapons: three independent ways a rule can fail

Nuclear weapons give two lessons pointing in opposite directions. The non-proliferation regime works partly because verification is physically possible: nuclear material can be accounted for, facilities can be inspected, and inspectors can be given access under agreements. That is a property of the technology rather than of the treaty, and it is why the nuclear rule binds harder than most rules of its kind.

Then take the test ban, which opened for signature in nineteen ninety-six and has still not entered into force, because a small number of states whose ratification the treaty requires have not ratified it. Which condition failed there is the instructive part. Not verification, which is the strongest part of the regime, but entry into force. A rule can be well designed and well monitored and still be held hostage by the consent of the few parties it most needs.

There is also a purely declaratory category of commitment in that regime. Several nuclear states have declared that they will not be the first to use such weapons. It is a real commitment, taken seriously by the parties who make it, and it has no verification and no penalty attached at all, because it is a statement about a decision rather than about a stockpile. Which leaves a clean typology: a rule can fail because it cannot be seen, which is the unverifiable commitment. Or because it cannot be enforced, which is the test ban waiting on ratifications. Or because it cannot be written clearly in the first place, which is the First Law of Robotics. Three failure modes, each independent of the others.

§The biologists: a bridge rather than a wall, punctured by one person

The biological case is the one where scientists bound themselves rather than being bound, and it is the closest thing in the record to our own situation. In nineteen seventy-four a committee of the American national academy called for a voluntary halt on a class of experiments until an international meeting could consider them, and the meeting came in February nineteen seventy-five at a conference centre in California. It produced a framework of containment levels that became the basis of national guidelines, and it is the standard example of a profession writing its own rule early.

Two things are worth noting. The halt was honoured widely enough for the conference to happen. And the self-imposed framework was later superseded by ordinary regulation once the technology became routine. Self-binding, on this evidence, is a bridge rather than a wall: it buys the time needed to write the statute, and then it retires.

There is a reason the biological case looks tidier than ours, and it limits how much can be borrowed from it. The halt covered experiments, which are discrete, countable, and conducted in identifiable places by people with careers at stake, so a moratorium on experiments is observable and its cost falls on the parties who observe it. Deployment is the opposite: continuous, distributed, and cheap to spread once weights exist as files. Nothing in the nineteen seventy-five model transfers to a rule about what a deployed system may be asked to do, which is why the artificial intelligence instruments examined in part two kept reaching for quantities instead of prohibitions.

And then it failed, in the germline case, which was not an accident of the technology but a person deciding that the norm did not apply to him. A researcher in China announced in November twenty eighteen that twin girls had been born from embryos he had edited, in defiance of a broad scientific consensus and of the law of his own country. The international reaction was swift and close to unanimous, the work was condemned by the bodies that speak for the field, and he was prosecuted in China and convicted. Then the revealing part: the scientific community had no machinery to enforce anything except exclusion and criticism, and the enforcement came from a state with a criminal statute. So self-binding held for the field and was punctured by an individual, and only a jurisdiction could answer that, which is the argument for turning a norm into a law while the norm still has authority.

§A decade of declarations, and the one instrument that carries weight

The machine-specific record is only a decade old and mostly made of declarations. In twenty seventeen a conference organised by a research institute produced a set of principles for artificial intelligence, published in August of that year and endorsed by many researchers. In March twenty twenty-three, the same institute published an open letter calling on laboratories to pause for six months the training of systems more powerful than the model then current, and it collected tens of thousands of signatures. In November twenty twenty-three a summit at Bletchley Park produced a declaration, followed by Seoul in May twenty twenty-four, with a statement of intent on cooperation in safety science, and by Paris in February twenty twenty-five, attended by representatives of more than a hundred countries, whose declared priorities were the science, the solutions, and the policy standards. So the decade produced principles, a letter, and three summit declarations.

By the three conditions, that is expressive rather than binding: none of it has a cost that lands on a breaker, and the observability is voluntary, which is why part two was about the statute book rather than the summit circuit. But one institutional fact matters more than the communiqués. The Seoul meeting produced a statement of intent on cooperation in safety science, and the practical result has been a network of national technical bodies that share evaluations and methods across jurisdictions. It is the most substantive thing to come out of the summit process, addressing the part of the first condition no single firm can satisfy. An evaluation that only the maker can reproduce is a press release. An evaluation that a second country’s technical body can reproduce is evidence. The network is how you get from one to the other, and it is being built slowly, by people whose names will not appear in any declaration.

One instrument from this decade belongs in a different category, and it is the least dramatic. In January twenty twenty-five an international panel chaired by one of the field’s senior researchers published a report on the safety of general-purpose systems, commissioned after the first summit, drawing on a large expert group across jurisdictions. A second edition followed in February twenty twenty-six, and its own summary emphasises model evaluations, dangerous capability thresholds, and conditional commitments of the if-this-then-that kind. So the instrument is not a rule. It is an agreed version of the facts, the precondition for a rule, and what it recommends is the scaling policy from part two written at international scale.

That is the strongest continuity in this series. In part two we had a laboratory writing commitments for its own models and publishing them so it could be held to them. Here the same architecture appears as an international recommendation, for the same reason: a commitment conditional on a measurement can be checked by someone who does not trust the maker. Before a people can bind itself, it has to be able to say what it is binding itself about, and an agreed report is how that begins. It is weak, and it is the part of the record that is actually load bearing.

§The chokepoint: computation as the measurable quantity

The first condition decides whether any of this can bind at all. The most serious answer the decade produced concerns computation. Training computation is measurable, cannot be created out of nothing, and is concentrated in a few facilities run by a few firms, so it can be counted and withheld. A rule keyed to a quantity of computation is checkable in a way a rule keyed to intent is not, and it can be enforced at the point where the hardware is made or where the facility is powered. That is why the statutory regime examined in part two keys its obligations for large general-purpose models to a quantity of training computation rather than a description of behaviour. The honest problem is that the chokepoint belongs to the technology’s supply chain rather than to the technology, and a chokepoint is only a chokepoint for as long as the frontier needs it.

The counter to that argument is stronger than the case for it. Computation is a chokepoint only while the frontier needs it. Efficiency improves, distillation compresses capability into smaller runs, and every chip already shipped is outside the reach of any future rule. The supply chain is also concentrated in one alliance, so a compute rule is enforceable exactly as far as that alliance reaches and no further, which makes it geopolitical rather than technical. On that reading the most verifiable rule we have is verifiable where we need it least and unverifiable where we need it most, and the honest defence is that a partial instrument beats a press release rather than that it solves the problem.

§The recursion, the score, and who actually pays

And then there is the recursion, which no other episode in this series can raise. The people writing these rules are using the technology to write them. Drafting instruments of this kind involves models summarising evidence, drafting text, forecasting effects, and increasingly assessing other models. The expert process that produced the international report draws on machine-assisted literature work, and the evaluations that fill the reports are produced by the same generation of systems they describe. In the garden there was a maker outside the system who knew the outcome. In the handbook there was a laboratory outside the machine that wrote the law. Here, the closest thing to an outside party is a committee of people coordinating across jurisdictions, using the subject of the rule as an instrument of the drafting. That is not a conspiracy and not a scandal; it means the judgement of the rule-writer is partly a product of the thing being ruled on.

The recursion should be handled carefully, because one version of that worry is an overstatement. Using a model to summarise literature or draft a paragraph is not different in kind from using a search engine, and no evidence exists yet that machine assistance has changed the substance of any instrument rather than its drafting speed. The serious version is narrower, and it concerns the shared corpus. If the same generation of models supplies the evidence summary, the risk framing, and the assessment of the completed rule, then the instrument and the thing it governs are drawn from one body of material, and the blind spot is shared rather than external. That is worth watching, and it is the one feature of this decade that neither the garden nor the handbook contains.

Now the score, condition by condition. Observability is met only partly: computation can be measured and hardware tracked, so the quantity-based rules are checkable, but behaviour cannot be, at least not yet, because the evaluations are contested by the people who ran them, which is what part two found when it read the shutdown-resistance work as a laboratory demonstration rather than evidence about deployed systems. The cost of breaking falls on the breaker is the weakest of the three, and it is weaker here than anywhere else described: the actors capable of breaking the rule are a handful of the most valuable firms in the world, the benefit of breaking it arrives quickly and privately, and the harm arrives slowly and publicly, spread across people who did not agree to anything. Climate is the closest precedent and the comparison is not encouraging. The third condition, blindness to your own position, is met more completely than in any earlier negotiation, because nobody writing these instruments knows which of today’s schoolchildren will be standing where.

The question of who pays deserves its own paragraph, because it is where practice is furthest from rhetoric. When a system causes harm today, the liability tends to land on the operator rather than the maker: on the hospital, the airline, the employer, the local authority that deployed the thing. So the party with the least technical knowledge and the smallest capacity to assess the risk carries the cost, while the party that designed, trained, and released the system is a step removed. The tribunal case in part two is the exception that proves the rule in both directions: the airline could not disclaim its own website’s advice, and the maker of the chatbot was not in the room at all.

§The disanalogy, the verdict, the markers, and the answer

The disanalogy is about what the rule is for, and it completes the arc of the series. The garden’s constraint was about a relationship and about mortality. The handbook’s was about safety. Ours is about capability and liability, which neither of the older stories contains at all. And there is a second difference that cuts the other way. A rule written by the people it binds can be amended by those people, and in part two we watched exactly that happen when four compliance dates moved. That is a defect if you want a wall. It is a feature if you accept that the rule-writer’s knowledge is incomplete, because a rule that cannot be revised is a rule that will be broken rather than corrected. The Hebrew text understood the danger as well as the advantage, since the first recorded amendment in that story is made by the constrained party, and it makes the rule stricter rather than looser.

The climate record is the closest precedent for the cost problem. The nineteen ninety-seven agreement set binding targets for a group of industrialised countries and did not cover the largest emitters of the following decade, and the country that mattered most declined to ratify it. The agreement that replaced it in twenty fifteen inverted the architecture: instead of targets handed down, each country submits its own plan, and the rule is a requirement to submit, publish, and return every five years with a stronger plan rather than a requirement to hit a number. Participation became nearly universal and the enforcement almost entirely reputational. That is a self-binding regime designed by people who had learned that a target nobody accepts binds nobody, and it moved the rule from the unenforceable category to the expressive one on purpose.

The verdict is narrower than three episodes of reading would suggest. A people can bind itself, but only under conditions it cannot manufacture. It needs an agreed account of the facts, which is why the unglamorous report matters more than the declarations. It needs a quantity that can be measured by somebody who does not trust the maker, which is why computation has become the instrument of choice. It needs a cost that lands on whoever breaks the rule, which is where every regime of this kind is weakest: the breakers are the strongest parties, and the harm is deferred onto people who were not in the room. And it needs a moment of writing that precedes the writer’s knowledge of their own position. Where all four hold, a self-binding rule holds, and the ozone case is what that looks like. Where two hold, the rule becomes expressive, which is what a summit declaration is. Where none hold, you have a handbook of robotics: three elegant sentences, no measurements, no enforcement, and an inexhaustible supply of stories about how it went wrong.

Which leaves the series owing the listener more than a verdict, so the conclusion should be able to fail. Three markers to watch. First, whether the European high risk regime is applied to a real system by the end of twenty twenty-seven, because on that deadline hangs whether a conformity assessment is an instrument or a filing exercise. Second, whether any jurisdiction ever blocks or withdraws a model on the basis of a computed threshold, because that would be the first time a quantity-based rule did something a press release could not. Third, whether a laboratory’s own published threshold is ever triggered in the direction that costs the laboratory money, because a self-written commitment tested against its author’s interest is the only kind that has been tested at all. If all three happen, self-binding has earned more confidence. If none does, what we have is a handbook.

And the listener, who sent in one comparison and one question three episodes ago and has waited through the lexicon, the commentaries, the five science fiction stories and the treaty record, is owed an answer. He said the garden and the three laws look like the same move twice, and that the constraint is the problem. His proposition survives, in a stronger form than he sent. Part one found that the Hebrew text does not withhold knowledge from a mind that is not ready: what it withholds, after the eating, is a second tree, and the enforcement clause is about death. What the tradition offers him is the Irenaean strand, Irenaeus on the infant who could not yet take solid food, Hick on a distance in the dimension of knowledge, and Keats on the world as a place of soul-making. Part two found that the prohibition is not what constrains real systems either; the undefined term is, the unbuilt testing regime, and the date that moved because the standards were late. Part three finds what all three stories share and only we have to face: the constraint does not withhold knowledge. It withholds position.

The garden is the place where the position of every party was visible, and that is precisely why the story has to send the man out of it. A rule that binds a people can only be written by people who cannot yet see which of them it will catch, and the moment they can see it, the rule has already stopped working. The tree did not take an ability from anyone; it moved them out of a place. The handbook did not make a machine stupid; it put a rule between the machine and the world. The listener said the constraint holds a mind back from knowledge. The better formulation is that a constraint tells a mind it is not the one holding the pen, and that the only version of that statement a people can be trusted to write is the version written before it knows which of its own children the rule will catch.

This has been a deep dive into what a people can bind about a technology it is still using.

Sources: this examination names its sources inline — the DeepDives three-part series on the garden, the handbook and the statute book; Jon Elster on precommitment, published in nineteen seventy-nine and revisited in two thousand; John Rawls on the veil of ignorance; the Montreal Protocol of nineteen eighty-seven and its secretariat, compliance committee, trade measures and fund; the non-proliferation inspection regime and the test ban opened for signature in nineteen ninety-six; the nineteen seventy-four national academy halt and the nineteen seventy-five conference in California, and the twenty eighteen germline case in China; the principles of twenty seventeen, the open letter of March twenty twenty-three, and the summit declarations of Bletchley Park, Seoul and Paris; the international safety report and its twenty twenty-six second edition; the training-computation thresholds of the European high risk regime; and the compliance dates moved in part two. Every source is weighed with the same skepticism regardless of politics or business model. The audio version of this article, read by DeepDives, is available on Wavlake.