In short

A rule can draw a line on human conduct and hold it. Written against nature, on an outcome that physics denies or on the configuration a system slides toward by itself, the line stops nothing: it presents the bill, to those who try to comply and to those who will be caught unprepared. This is not a discovery: it has already happened, more than once, and each time it is forgotten.

No parliament has ever required an airliner to take off below a hundred kilometres per hour. No legal code requires a body in free fall to accelerate at less than five metres per second squared. And no official gazette has ever published a ban on entropy increase for isolated systems. Rules like these do not exist because their absurdity is visible to the naked eye: physics does not take orders. Yet rules with the same logical form do exist, are in force, and more are written every year. The absurdity is identical; it is only distributed so as not to be seen.

What follows tells nothing new, and is largely an interpretation. The cases collected here run from 1775 to today, and in each the form is the same: the subject changes, the names change, the bill stays the same. What seems to change little is memory: every technology that arrives finds someone who would stop it as if it were the first, and the bill, already presented, is paid again. The texts of the rules, the dates and the numbers are facts, and are given for what they are; the readings of their effects, and of the direction systems tend toward, are the author's, and the text marks them as such. The examples come mostly from cryptography and from artificial intelligence systems, because that is where the form is easiest to see today; the reflection holds for any regulated technology.

Three kinds of situation look alike and are not the same thing.

The first is the job of rules: physics allows a document to be forged, a well to be poisoned, a whole city to be watched; the rule forbids humans to do it, and the line can be held. The serious questions about this kind of line are different ones, and this site has already met them: which line, decided by whom, at what cost.

The second is the pathology: the rule does not forbid a conduct, it demands an outcome that nature denies, or forbids an outcome that nature guarantees. The phenomenon does not stop, because it cannot; the rule stays in force all the same, and produces effects of its own, which do not touch the phenomenon: whoever enforces it decides case by case where the line is, and the cost falls on those who try to comply.

The third sits in between, and is the most insidious because it is not impossible: a system left to itself slides toward the most stable configurations it has available, like heat flowing back into a refrigerator that has been switched off. Keeping it away from one of them can be done; but it costs, and it costs in a precise way. It is a bill to work out before writing.

Who draws the line

In July 2017 the Australian government announced a law to compel messaging companies to hand the content of encrypted messages to the police. The objection was already known. With today's encryption the key sits only on the phones of the people talking: the company carries the messages but has no way to read them, and to hand them to the police it would have to weaken everyone's encryption, opening a door it could not reserve for the authorised alone. Asked whether a law could impose this on mathematics, Prime Minister Turnbull replied: "The laws of mathematics are very commendable, but the only law that applies in Australia is the law of Australia." It is the ban on entropy increase, spoken aloud.

The law came the following year, and the proclamation was gone. The text requires companies to assist investigations, but forbids asking them for a "systemic weakness": a weakening of protection that hits all users, and not only the target of the investigation. The whole discussion since then has turned on that point. Is a change made to one person's app systemic or not? And if the same change can be repeated on anyone? The text does not say, and criticism of the ambiguity of that word has accompanied the law for as long as it has existed.

The European AI Act makes the same move in writing. It asks that content generated by artificial intelligence systems be marked with solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible". For written text, where the mark is a statistical signal hidden in the choice of words, the measure of the feasible is known and has been looked at closely on this site: under repeated paraphrase, detection drops from 99.3 to 9.7 per cent, because whoever reshuffles the statistics erases the signal. The closing clause carries the whole weight of the rule. It is not an isolated case, and perhaps not even a fault: the regulation had been proposed in 2021 for a different generation of systems, and the chapter on general-purpose models was added mid-course in 2023. A rule written before the thing is understood can only be written on the outcome it must not produce, not on the conduct that produces it; and on the outcome, then, the clause is needed.

Looking through the rules in force, the "ban on entropy increase" written bare is almost nowhere to be found: it lives in speeches, and in the texts it arrives covered by a feasibility clause. One might conclude that the system defends itself. But the clause does not defuse the impossible rule: it moves the drawing of the line from whoever writes to whoever enforces. How far a mark is "technically feasible", when a weakness becomes "systemic", the text does not say: enforcement will decide, case by case, authority by authority. The line is still there; only, nobody knows where it is, and one finds out at one's own expense.

A rule that nobody can fully comply with thus turns compliance into discretion. If full compliance is impossible by construction, everyone it applies to is in breach to some degree, and deciding whom to prosecute stops being a finding of fact: it becomes a choice of the enforcer. And there is a selection, worth looking at squarely. The cost falls on those who try to comply: the provider that measures, documents and certifies an effort that cannot reach the end. Whoever ignores the rule pays none of this, and the phenomenon, which depends on neither, goes on unchanged. A rule of this kind selects against honesty: it rewards, in relative terms, the behaviour it wanted to discourage. It is the limiting case of an effect already observed on this site: compliance costs are absorbed by those with broad shoulders, and the ban ends up favouring them. Here not even broad shoulders can buy compliance, because it does not exist. None of this concerns the first situation: a line on conduct, who forges and who does not, can be ascertained, and the rule does its job.

A pump that never switches off

A rule written against a natural tendency is not an impossible rule: it is a pump.

Tendency, here, in an almost literal sense. A system left to itself slides toward the most stable configurations it has available, down a gradient, like heat toward cold. Physics does not forbid keeping it away from one of them: it charges for it. A refrigerator does not violate thermodynamics: it pays for it every hour it runs, and the day the plug is pulled the heat comes back in. This is the load-bearing image of this piece, and it is worth saying what it is: an image, not a law to deduce from.

A rule of this kind works as long as someone feeds it, and the energy bill never stops arriving: checks, sanctions, definitions to rewrite at every version of the technology, and a guard that can never be declared finished.

Then there is a second cost, which appears on no balance sheet and is probably the dearest: the preparation that did not happen. The resources spent guarding the unstable configuration are not spent organising the one the system was sliding toward. When that configuration arrives, and if the tendency was real it arrives, it finds rules written to deny it and none to govern it: no inventories, no deadlines, no responsibilities assigned. The guard has not only failed: it has used up the time in which one could have prepared.

When is a tendency of this kind? A rigorous criterion does not exist, but three clues can be counted. First: every actor, on its own, gains by moving in that direction, with no need to coordinate with anyone. Second: the paths leading to the configuration are many and independent of one another, and closing one does not close the others. Third: holding the tendency still requires a continuous guard, not a one-off act.

A warning, because the step is slippery in both directions: "inevitable", said of a system made of people, is a forecast and not a theorem, and whoever pronounces it in the indicative makes the mirror-image mistake of whoever writes the ban. This does not contradict what was said about bans: a ban is never inevitable, because it is a decision; a tendency of a system can be, and the rule written against it remains a decision, with its bill. This is why the three clues are clues, not a proof: they are worth as much as the real cases they apply to, starting with those in which the tendency lost.

Where the rule wins

The difference between the tendencies that rules have stopped and those on which they have worn themselves out does not seem to lie in the strength of the ban: it lies in a count of actors and paths.

Chlorofluorocarbons were worth producing and using, and the first clue was all there. But they were produced by a few dozen plants in the world, the paths could be counted, and substitutes existed. There was a chokepoint, that is, a point all the paths pass through and that can be counted, and the Montreal Protocol of 1987 was able to tighten there, without denying the convenience but repricing it and organising the replacement, with an act that requires no guard over every single user. In 2026 the ozone layer is recovering.

Nuclear weapons are the most instructive case, because there the forecast of inevitability bore the most authoritative signature of the twentieth century. In March 1963 President Kennedy said he saw the possibility, by the 1970s, of a world with "15 or 20 or 25 nations" holding nuclear weapons. In 2026 there are nine. Building a bomb takes highly enriched uranium or plutonium, and producing them takes plants that are large, slow and visible, which can be counted even from afar. The Non-Proliferation Treaty does not forbid wanting the bomb, an outcome no treaty can prevent: it puts the material and the plants under inspection, kilogram by kilogram, and the countries that have the technology control whom they pass it to. It has not made the bomb impossible, and some of the nine built it while staying outside the treaty or leaving it: it has made it slow, visible and expensive, and the tendency taken as almost certain stopped far earlier than forecast. The count cuts both ways: it refutes bans written where there is no chokepoint, and it refutes prophecies of inevitability pronounced where there is one.

On the other side stand the tendencies with no chokepoint, and cryptography is the example this site knows best. Until the late 1990s the United States treated strong encryption as a weapon: exporting it required a licence, and software with long keys stayed at home. But strong encryption is published mathematics, implemented in programs that sit on every computer and are copied at zero cost: all three clues are there. In June 1999 a study by George Washington University counted 805 encryption products developed in 35 countries outside the United States; in January 2000 the rules were loosened. The ban had not stopped encryption: it had moved development across the border, and whoever did not apply the rule acted as a thermometer, making legible a cost that until then only American producers had been paying. This seems to be how rules of this kind fall: not because someone refutes them, but because the comparison with those who do not apply them makes the bill visible to all. The copying of information is the limiting case of the same family: a file duplicates without being consumed, and published information cannot be withdrawn, because the paths of the copy are all the disks in the world.

The current case is artificial intelligence, and the three clues split it in two. In October 2025 an appeal signed by some of the founders of the discipline, and by hundreds of others, called for a prohibition on the development of "superintelligence" until there is scientific consensus on doing it safely. It is a proclamation, like the Australian prime minister's, and like that one it lives in speeches; but it is worth counting it as if it were a rule. Every laboratory gains on its own by going ahead: the first clue is there. The paths, however, are not all alike. Training the largest models requires processors that come out of very few factories and compute runs that can be measured, so much so that the AI Act reserves its heaviest obligations for models trained above a compute threshold, ten to the power of twenty-five operations: there a chokepoint exists, and a rule has a place to stand.

But a chokepoint is a state of affairs, not a property: it lasts as long as nobody opens a parallel path. Fissile material has been one for eighty years because physics keeps it so; the fastest processors have been one for a few years, through industrial concentration, and the concentration already seems to be loosening. The same holds for the other candidates the debate proposes in turn, the energy required, the size of the models, the best researchers: each can be squeezed for a while, none for long, and a rule built on one of them demands a continuous guard, exactly as the gradient it would stop does. The weights of an already trained model, that is, the result of training, are instead a file, they can be copied like any file, and the methods are published: there the chokepoint does not exist. It was already the difference between the oil barrel and the model, noted when discussing bans: the barrel can be controlled because it cannot be copied, and what is truly scarce sits upstream, in compute; the model, unlike the barrel, is copied. A ban that does not find the chokepoint binds only those who apply it, and those who do not act as a thermometer, like the encryption producers outside the United States in 1999.

A year after the appeal, in September 2026, the proclamation arrived in a bill (in the United States), so far only introduced: it defines superintelligence by outcome, a system that "exceeds human cognitive performance and capabilities across most domains or tasks", entrusts a department with recognising it, and for whoever builds it provides the same penalty as for building a nuclear weapon. The frame is the nuclear one; the chokepoint of fissile material, here, sits in compute and for now in nothing else. What will happen, no theorem says. The count says something smaller: which part of a rule can hold, and which part is a pump.

The rule wins where there is a chokepoint that can be counted, and where it reprices instead of denying. Where the paths cannot be counted, the pump remains. The chokepoint says where a rule can hold, not whether it should: that remains the question from before, which line, decided by whom, at what cost.

Refusing the questions, preparing for the probable

The alternative to the pump has existed since before all the rules discussed so far, and it has two faces: one answers the impossible outcome, the other the tendency.

The first face is refusing impossible questions. In 1775 the Académie royale des sciences in Paris resolved that it would no longer examine any solution of the squaring of the circle, the trisection of the angle and the duplication of the cube, nor any machine announced as perpetual motion: a single act, for mathematics and physics together, with the reason on record, the talent and the fortunes those researches consumed in vain. Two and a half centuries later the United States patent office does the same with a single rule: it almost never asks for a working model, "with the exception of cases involving perpetual motion". The well-made rule does not forbid entropy to grow: it refuses the questions that demand it decrease, and puts the burden of proof on whoever asserts the impossible, the only place where that burden stands up.

The second face is preparing for the configuration that is approaching. A quantum computer, if and when it reaches the necessary scale, would be able to derive, from the keys with which digital signatures are verified today, the secret keys that produce them. Nobody proposes banning the algorithm. The work under way is different: in 2024 NIST, the United States standards institute, published the first signature schemes resistant to quantum computation, and around those standards people are drawing up inventories of what is exposed and migration plans, including the one already told here for Bitcoin. A standard is not a ban: it is a point to converge on, and whoever publishes it guides without imposing. It is the same logical form as the Académie's, turned forward: not denying the probable configuration, arriving at it with the questions already in hand.

Two contrasts keep this face in its place. Preparing is not surrendering: a migration is made of deadlines, inventories and responsibilities, that is, of more government, not less; whoever guards in order to deny gives up precisely this work. And stable does not mean desirable: preparing for a configuration does not mean blessing it, it means refusing to pay for it twice, first in the guard that does not stop it and then in the unpreparedness when it arrives.

Where the line runs

Rules can do much, even against tendencies. A dam does not forbid water to flow downhill: it channels it, and in doing so draws energy from it. Channelling and repricing is the job of rules, and the cases in which they won, from the Montreal Protocol to the nuclear treaty, are cases of this job done well, on tendencies that had a chokepoint to tighten on.

What rules cannot do is what this piece started from. The line drawn on human conduct holds; written against nature, on the impossible outcome or on the almost inevitable tendency, it does not stop nature. It presents the bill: discretion in place of compliance, selection against those who try to comply, the guard to feed forever, the unpreparedness at the moment the stable configuration is reached. The bill was presented in 1775, when the Académie put on record the talent and the fortunes consumed behind perpetual motion, in 1999, when the encryption products born outside a ban were counted, and many times in between; each time it seems to arrive for the first time.

In a world where whoever does not apply a rule is one border away, or one link, the thermometer is probably closer than yesterday, and the bill arrives sooner. The most useful rules, on these tendencies, will probably be those that worked out the bill in time: which configuration is more stable than the present one, what it costs to guard it, what is needed to govern it when it arrives. For artificial intelligence that configuration does not yet seem readable, and for this reason it does not seem useful to try to guess it. The ban on entropy increase will remain absurd in every official gazette. The entropy bill, that one, can always be worked out beforehand.

References

  • Transcript of the press conference of Prime Minister Turnbull and Attorney-General Brandis, Sydney, 14 July 2017: pmtranscripts.pmc.gov.au
  • Telecommunications and Other Legislation Amendment (Assistance and Access) Act 2018, No. 148, section 317ZG: legislation.gov.au
  • Regulation (EU) 2024/1689 (AI Act), Articles 50 and 51: eur-lex.europa.eu
  • Future of Life Institute, Statement on Superintelligence, 22 October 2025: superintelligence-statement.org
  • Ban Artificial Superintelligence Act of 2026, S. 5493, 119th Congress, introduced 23 September 2026 (House text: H.R. 10538): govinfo.gov
  • V. S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, S. Feizi, Can AI-Generated Text be Reliably Detected?, TMLR 2025: arxiv.org/abs/2303.11156
  • UNEP, Montreal Protocol on Substances that Deplete the Ozone Layer, 1987: ozone.unep.org
  • J. F. Kennedy, News Conference 52, 21 March 1963: jfklibrary.org
  • SIPRI, Increasing focus on nuclear weapons amid heightened escalation risks: new SIPRI Yearbook out now, press release, 8 June 2026 (the nine nuclear-armed states): sipri.org
  • L. J. Hoffman, D. M. Balenson, K. A. Metivier-Carreiro, A. Kim, M. G. Mundy, Growing Development of Foreign Encryption Products in the Face of U.S. Export Regulations, Cyberspace Policy Institute, George Washington University, report GWU-CPI-1999-02, 10 June 1999: cspri.engineering.gwu.edu
  • U.S. Department of Commerce, Bureau of Export Administration, Revisions to Encryption Items, Federal Register vol. 65 no. 10, 14 January 2000, pp. 2492-2502: govinfo.gov
  • Histoire de l'Académie royale des sciences, année 1775, Paris, 1778, pp. 61-66 (the resolution on the three classical problems and on perpetual motion machines, with the reasons): archive.org
  • M. Jacob, Interdire la quadrature du cercle à l'Académie : une décision autoritaire des Lumières ?, Revue d'histoire des mathématiques 11 (2005), pp. 89-139 (the session of 3 May 1775 in the minute books): doi.org/10.24033/rhm.26
  • USPTO, Manual of Patent Examining Procedure, section 608.03: uspto.gov
  • NIST, NIST Releases First 3 Finalized Post-Quantum Encryption Standards (FIPS 203, 204, 205), 13 August 2024: nist.gov