--- title: Chapter Four — The Clock book: Purposive Accelerationism author: '@jamie247' position: 8 of 24 canonical: https://whatispacc.org/ebook/#ch4 version: 1.1 language: en-GB licence: all rights reserved; quote with attribution --- # Chapter Four — The Clock “Day by day, however, the machines are gaining ground upon us; day by day we are becoming more subservient to them…” — Samuel Butler, “Darwin among the Machines”, 1863 Every conversation about artificial intelligence arrives sooner or later at the same question: how long have we got until we cross one of two thresholds? The first is general intelligence, usually taken to mean the point at which a machine can do most of the cognitive labour people are paid for, organise their lives around and build careers on, at least as well as a person could if not better. Ask the field’s leading figures whether we have crossed it and the answers cancel one another out. Peter Norvig, who ran research at Google, wrote with Blaise Agüera y Arcas in 2023 that we already have. Yann LeCun, the Turing Award winner who led Meta’s AI laboratory, says today’s language models never will. And Dario Amodei, who runs Anthropic, would rather nobody used the term at all, because it is vague and weighed down with hype. If the people closest to the task cannot agree where the line is, or whether it even counts, it is of no use as a clock for the rest of us. The second is superintelligence, and the clearest account of it is still the one the Oxford philosopher Nick Bostrom gave in his book of the same name back in 2014: a mind decisively beyond ours in almost every field that matters, above all the strategic ones: planning, persuasion, research and the design of complex systems, better minds among them. That last point carries the oldest concern. In 1965 I. J. Good, who had broken codes alongside Alan Turing at Bletchley Park, pointed out where it might lead: that a machine able to improve its own design sets off what he called an intelligence explosion, each version building a stronger successor until no human institution can keep up, making it “the last invention that man need ever make”. Both thresholds, however, measure only what the technology can do, its capacities. Neither addresses the machine as it relates to us, and that relation is the clock that matters most, because a system that needs people still has to answer to them. Public discourse counts how much we have come to depend on the machine. The clock here counts how much it still depends on us. But being needed protects us only under two conditions. A person held inside the system as reserve stock — an input priced like any other and replaced the moment a cheaper one appears — is in the loop without any hold over it, because what is needed is the work rather than the person. The first condition is that people are legible to the machine as holders of claims rather than as tasks, a claim being anything it must take into account before it acts: a legal right, an interest it is bound to weigh, a refusal it has to accept. Someone who holds such a thing can stop the machine or oblige it to answer; someone who holds none is merely processed by it. The second is that the human who answers for an AI must be someone it cannot act without. A named person given the liability for a system’s decisions and no power to stop them is a scapegoat: a human face for the blame where there was none for the decision. The clock runs while both hold, and what this book has been describing is the process by which they lapse. Call what it measures dependence, and notice that it runs in both directions. Ours on the machine is the one we feel every day and can count in jobs. The machine’s on us is quieter, and it is the one that keeps us inside the loop, and it is falling at three scales at once. The first is the individual — the expert, the manager or the manual worker — whose tacit judgement the machine needs a little less each year. The second is the economy, where it shows up in firms: the support desk, the translation bureau and the junior lawyer’s office now produce what they did with half the staff, and need fewer each time the model improves. The third is the state, which has become a customer: it rents its intelligence — from the tax office to the battlefield — from a handful of companies, and finds the price of leaving any of them rising past what it would be willing to pay. The economy’s hand moves fastest, because a payroll can be cut in a quarter. The state’s matters most, because once a government can no longer change its supplier, nothing beneath it can be changed either. What makes this clock run faster than any equivalent dependence of technology before it is the mechanism inside Good’s earlier threshold: systems that help build their successors need fewer and fewer people with each round. Models themselves now write much of the code inside the laboratories that make them, even chip layouts are drawn with their help, and in October 2025 Sam Altman told an audience in Dubai, in the United Arab Emirates, that data centres able to build other data centres “aren’t that far off”. What is handed to machines does not come back, and each round pays for the one after. The intelligence explosion is not a line to be crossed. It is the reason the pace keeps rising. Judgement is the hand to read first, because it is the one we would use to read the other two. *** Today the machine still needs us for our industry and, most of all, for our judgement, the capacity on which every other of our freedoms rests, and without which we cannot notice how far we have come to rely on our tools. Psychologists call the habit of handing mental work to a device cognitive offloading, and civilisation has always done it, from the abacus to the apps on a smartphone. What is new is that the act of judging itself can now go the same way, whole, which is the cognitive subsumption met earlier. In 2025 a survey of 319 knowledge workers by Microsoft Research and Carnegie Mellon University found that the more people trusted an AI, the less they checked what it produced. But the pattern pre-dates today’s AI technology. In a flight-simulator study in 1999 the psychologist Linda Skitka found that about 65 per cent of people using an automated aid followed its prompts when they were deliberately wrong and even when the instruments said so. Some 41 per cent of them missed what the gauges showed whenever the automation stayed silent, against 3 per cent of those working without one, and by 2010 Raja Parasuraman and Dietrich Manzey could report in Human Factors that the lapse appears in experts as readily as in novices and that training does not remove it. Its cost has since been measured. When Hamsa Bastani and her colleagues gave nearly a thousand Turkish secondary pupils unrestricted access to GPT-4 for their mathematics, in a randomised trial reported in PNAS in 2025, practice scores rose by 48 per cent and marks in the exam sat without it fell by 17 per cent, while a version built to give hints rather than solutions raised the gain to 127 per cent and lost nothing when the tool was taken away. Such trials capture the step people stop taking and what skipping it costs — weighing evidence, testing an argument, telling an answer that is true from one that merely sounds believable — and whether the ability itself holds across a population is the hand this clock reads. Judgement is lost first because it is the cheapest thing to give up. A task given to a machine comes back as an answer, and the checking is the one step that produces nothing new, so it is the first to be skipped, and it goes unmissed, because noticing it is the faculty being surrendered. The channel it runs through is what you read, search and decide with, and the change has already reached most of us, though few could say when it arrived. A few years ago you put a question to a search engine, got back a thick column of results and made up your own mind from what you found after clicking around. Since 2024 Google has placed an answer written by its own AI at the top of the page, for those who still search at all rather than ask a model outright, and in 2025 the Pew Research Center found that people who saw this new bundled answer by Google clicked through to a source on 8 per cent of visits, against 15 per cent among those who did not. By late 2025 ChatGPT alone had some eight hundred million weekly users, and in May 2025 Apple’s Eddy Cue told a court in Washington that searches through Safari had fallen the month before for the first time in twenty-two years, which he put down to people asking AI instead. Whoever owns these models now sits between a person and what they know, and the power works differently from old-fashioned censorship. Banning a book or a website is crude and visible, so it tends to breed resistance. This works earlier and more quietly, by deciding which sources even get summarised, which follow-up questions are suggested and which answer appears first, so that it shapes the options you see before you have formed a view of your own. Nobody pushes back, because it never feels like being told what to think, only like being looked after. A population that has lost those habits cannot contest what its tools decide on its behalf, and it will not easily know what it has lost, since the loss is the very faculty that would have made it visible. That is not yet domination, because formal domination still requires a party holding a power to interfere that others cannot contest. It comes before domination, and in at least one respect it is worse, because the defence the republican tradition relies on is contestation, the standing ability to answer back, and a population that has stopped judging has unwittingly given up that defence before any party has had to seize it. None of it requires a machine cleverer than we are. It is already quietly happening with the ones we already have — in the analyst who signs off a report she has not properly read, the journalist who half writes the rushed column, and the search that no longer sends anyone to source. In a 2023 study published in Radiology, Thomas Dratsch and his colleagues at University Hospital Cologne, radiologists averaging more than fifteen years in the job were right 82 per cent of the time when the software’s suggestion beside a mammogram was sound and 45.5 per cent when it was wrong. What a person gives up little by little, task by task, an economy gives up by the profession, and that is the second hand. *** Work, our labour, is the second thing the machine needs less of, and here the handover has a hard number. In September 2025 Marc Benioff, the chief executive of Salesforce, said on a podcast that AI agents had let him cut his customer-support staff from circa nine thousand to five thousand people, “because I need less heads”. For two centuries machines mostly made people more productive and, eventually even if unevenly, better off, because they were tools in a worker’s hands. An agent that does the job whole reverses that. Less of the income the work earns is paid out as wages and more is kept as profit, and the profit buys more compute, which does more of the work: the Jevons effect met earlier, run at the level of the firm, where the cheaper a task becomes the more of it is handed to the machine, and none of the saving returns to the people it displaced. While decisions about where money was invested ran through people, there were points at which they could be answered: a board that could be voted out, a bank that could call in its loan, a customer who could take her business elsewhere, a court that could hold the stronger party to a contract. Automate the allocation and those points thin out, because the choosing moves inside machinery that fewer people can reach, and those who can are its owners. This is capital autonomisation, met earlier, and its destination is an economy that drives itself. Followed to its end, this is the market road. One system sets the price of your labour, your savings and the claims you hold on others, and it prices them all in the same unit — the token, which is a measure of compute and in the end of energy. Your income then depends on its owner’s goodwill, as the slave’s did on the kind master’s. The money for this is not waiting on profits either. On 21 January 2025, the day after his inauguration, Donald Trump stood at the White House beside Sam Altman of OpenAI to announce Stargate, a plan with SoftBank, Oracle and MGX to spend up to $500 billion on data centres. Meanwhile the prospectus Anthropic prepared for its listing, reported in September 2026, showed an operating loss of about $8 billion on 2025 revenue of $4.6 billion, roughly $1.70 lost for every dollar brought in, with more than half its costs going on computing power. No ordinary business could carry losses on that scale, and no ordinary investor would fund them. These are bets on a machine whose next version needs less human work than the last: the model helps design its successor, the successor needs fewer researchers, and the saving is reinvested in capability rather than returned as profit. Nothing in that loop requires one laboratory to be ahead of another. It requires only that each generation needs less of us, and so far every one has. Speed is the third front, and here the contest is with the referee rather than between rivals. On the afternoon of 6 May 2010, in about the time it takes to make a cup of tea, automated trading wiped nearly a thousand points off the Dow Jones. Procter & Gamble, one of the historically steadiest stocks in America, briefly lost more than a third of its value, and shares in Accenture changed hands for a cent. The regulators’ report explaining what had happened took almost five months to compile, which is the gap in miniature: the machine acts in minutes and the institution understands across seasons. Software now trades in microseconds — not just for elite hedge funds but for anyone who cares to play — retunes an influence campaign by lunchtime and ships new code overnight, while the parliaments meant to hold it to account still break for the summer. By the day a rule is drafted, the world it was written for has moved on several times over, and whoever moves first gets to settle the question before anyone else has thought to ask it. The loss at this scale is the individual person’s, written larger. A society whose work runs through a system someone owns cannot afford to refuse it, and a referee who arrives after the event has nothing left to contest, so the pressure that once bent allocation towards people has nowhere to push. Employees feel this before governments do, because a payroll can be cut in a quarter and a state cannot. A state can, however, be locked in, and that is the third hand. *** The power a society has to deliberate and change course can be lost in two ways. The people who would make the choice can lose the habit of judging for themselves, which is the erosion already described. Or the change can grow too costly to carry out no matter how much it is desired, because so much has been committed at a system level that leaving would break any reasonable reckoning of the cost. The two feed each other, since a public that can no longer weigh the evidence will not press for an expensive reform, and a government left unpressed will not fund one. Anyone who has tried to quit a technology an organisation depends on knows how the second works. In 2004 Munich, in Germany, began moving its offices from Windows to open-source software, and the migration took nearly a decade. In 2017, after years of complaints about compatibility, the council voted to abandon it and go back, at an estimated cost of €49.3 million to the public. Each move had a price that could be checked, and each was argued out in public. In the same way you can put a number on what a health service, a defence ministry or a school system would have to spend to unwind its dependence on a supplier, and someone accountable can decide whether to pay it. Three markers show how far round the dial the hand has come at this scale, and how close a country is to losing the ordinary power to choose and change its suppliers. Each is a reading of dependence, a point at which the machine needs a government’s consent a little less. The first arrives when the cost of leaving, the kind Munich paid, passes what any treasury is willing to spend. Here the state is a customer, and the dependence is the kind any customer knows. The US Internal Revenue Service still keeps its master file of individual taxpayers — the record behind every American’s tax return — on code dating from the 1960s and written in a language few engineers now read. Attempts to modernise it have been launched repeatedly since the 1990s, and none has yet replaced it. Every process in the agency depends on it, the only people who understand it are retiring, and a failed switch in the middle of a filing season would be a national crisis. None of it is irreversible in principle. The money and the know-how to swap the old code out exist, but no government will spend them, and in politics a change nobody will pay for comes to the same thing as one that cannot be made. The second marker is harder to see, because nothing visibly breaks. Here the state stands not as a customer but as a regulator, and what it depends on is not a supplier it cannot leave but a view it cannot get. It arrives when the gap described under speed grows so wide that the people responsible for oversight can no longer tell what a system is doing until long after its consequences are felt. In 2018 Facebook altered its News Feed to reward what it called meaningful social interactions. Its own researchers soon found that the new ranking was spreading outrage, yet the public learned of it only in 2021, when Frances Haugen, one of the company’s product managers, walked out with tens of thousands of pages of internal documents and handed them to the Wall Street Journal. For three years nobody outside the firm, regulators included as far as is known, had sight of it. What a government loses at this point is the ability to check, before the damage is done, whether a system does what it was meant to, at what cost and with what side effects. All it has left is the post-mortem, which comes after the fact and usually in the form of a loss. The third marker arrives when the supplier becomes part of the machinery of force itself — the police, the army and the intelligence services that make up what the German sociologist Max Weber, writing in 1919, called the state’s monopoly on legitimate violence. In July 2025 the Pentagon’s office for digital and AI awarded contracts worth up to $200 million each to Anthropic, Google, OpenAI and xAI. In March 2026 the department declared Claude’s maker a “supply chain risk” for refusing to drop its own limits on mass surveillance and autonomous weapons, and gave the military six months to phase its models out. At the time of writing, in late 2026, that order stands, and OpenAI has announced a deal to take its place in classified settings. Once an AI system sits inside that apparatus, the terms of its use are set by security rather than by contract. A vendor that holds to its own limits is treated as a threat, and one that accepts the conditions becomes an arm of the military. Removing it is then a defence decision rather than a purchasing one. This is how the hybrid Leviathan met earlier gradually takes shape in practice, as vendor lock-in, out of systems that are only tools, each bought for a job, put together by officials who each believe they are doing theirs. It is worth saying why this differs from the typical lock-in every IT department knows, which is a procurement problem, and procurement has remedies. That kind holds a database hostage, and a database can be migrated. This one locks in the judging itself, which cannot be exported. The supplier can be switched off by the state, and the state’s capability by the supplier, as the order from Washington that took Anthropic’s newest models offline for nearly three weeks in June 2026 showed in both directions. Meanwhile the product continues to improve itself in the background, so the gap to any alternative widens on its own while the bill for leaving rises. A government that cannot leave, cannot see and cannot remove its supplier from its own means of force has lost the ordinary power to choose, and it has lost it without a staged coup. Every contract was signed in broad daylight for sound and perfectly defensible reasons. *** Western Union dominated the telegraph, the Hollywood studios controlled film, and Bell had such a grip on the telephone that it even owned the handset in your home and rented it back to you. Each was eventually prised loose, by regulation or by its own competitive forces. Tim Wu, the Columbia law professor who coined the phrase net neutrality, traced the pattern of bundling and unbundling in his 2010 book The Master Switch and found it regular enough to name it, the Cycle. Every new information technology starts out open and chaotic, consolidates into the hands of a single owner, and is then broken up by the arrival of the next one. If that pattern still holds, today’s concentration will pass too, and the clock is a false alarm. Across his cases I count four ways such empires have ended. Sometimes the state did it — the Paramount decree that split the studios from their cinemas in 1948, the break-up of AT&T in 1984 — and that route remains open to governments for now. The other three are closing, and their closing is what the clock measures. The first of those three is people. Know-how has always lived in the heads of those who make the system: engineers leave, retire, teach, get poached and write it all down for a rival, and they still do. Anthropic was founded in 2021 by a group who walked out of OpenAI, and Ilya Sutskever’s Safe Superintelligence and Mira Murati’s Thinking Machines began in the same way. But what none of them could take with them was the model itself. A frontier capability now lives in its weights and, increasingly, in the data centres that produced them. A researcher who leaves carries the know-how but not the machine, and certainly not the raw compute, and the gap between understanding the method and owning the result has never been wider. The second is decay from within. A company that owns a market tends to become slow and defensive about what it already sells, and misses the next big thing through inertia. In 1975 a young Kodak engineer named Steven Sasson built one of the first digital cameras, a toaster-sized box that took twenty-three seconds to record a single image onto cassette tape. The company was slow to back it for fear of hurting its film business, a textbook case of what the Harvard professor Clayton Christensen called the innovator’s dilemma, and it was forced to file for bankruptcy as a consequence in 2012. That is a weakness of organisations, of committees, incentives and executives who did well under the old arrangement and would rather it continued. The laboratories may yet fall into that trap, but so far they replace their own flagship models every few months. The third is the outsider. Breaking up a dominant firm has always needed a newcomer to grow up somewhere on its periphery, beyond its reach, the way the telephone developed away from Western Union’s wires and Google displaced AltaVista. This time there may be no such place. This is because the technology in question is general-purpose, and one of its uses is producing whatever arrives next, with each model now helping to build its successor, so the leaders already occupy the ground a challenger would have come from. A challenger that did appear would be breaking up firms the state has made too big to fail, as the state road showed, carrying debts on a scale that no government or market can now let default. One channel does remain open for now, though it runs only among those who own the machines. The frontier laboratories compete ferociously, poach each other’s staff, publish more than is probably wise and leak constantly, and that diffusion is what has stopped any one of them running away with the field so far. But it too is being shut, as lab and state merge into the hybrid Leviathan, in which a capability finds a security clearance where it once found a buyer, and research moves under export controls and classification. The nearest thing left to an outsider is the open-weight model, which anyone can download and run. These cost a fraction of what the frontier does and trail it, on most estimates, by only a matter of months rather than years, even if some of them learn from the leaders, or copy them. When DeepSeek put out one of its own models in January 2025, Nvidia lost close to $600 billion of market value in a single day. Enterprise is already moving the same way: in July 2026 Palantir’s Alex Karp told CNBC that his customers were “livid” at how much of their value flowed to the model providers, and paired his company with Nvidia on open models so that a firm keeps its own data and weights in-house. Releases like these threaten to turn the laboratories’ own product into a commodity, the reverse of what Peter Thiel taught a generation of founders in Zero to One, his handbook on escaping competition by building a monopoly. Value rarely disappears when one layer of a technology is given away, however. It moves up or down the stack, to whoever controls access to the cheap part, a pattern Christensen called the conservation of attractive profits. The guess made earlier was that the go-between would be the language model itself, but once it costs nothing, power passes to the personal agent built on top of it, which sits between you and everything you do online. Its owner knows what you want before any website or app does, and keeps the compute it runs on and all it remembers about you in one company’s hands. Downloadable weights restart the Cycle at one level and leave what Wu called the master switch, the single point of control over a medium, at another. Whether the agent closes around one laboratory or stays open to all is still a matter of design. It could be built to answer only to the person using it, with a memory its owner can take elsewhere, and for now a public alternative of that kind can still be paid for. Whether anyone will foot that bill is the real question, and behind it lies another: can a state still afford to? *** None of this is beyond repair yet, and the remedies can be named, which is the whole difference between a price and a doctrine. Effective accelerationism tells a different story, that reversal is impossible at any price. In its telling, technological progress advances as a force of nature, driven by the same physics that explains life itself, so that nobody is really steering, and whoever suggests otherwise has misunderstood the nature of the beast. This is technological determinism, and the idea is older than the movement now wearing it. Marx wrote in 1847 that the hand-mill gives you society with the feudal lord and the steam-mill society with the industrial capitalist, and Kevin Kelly made a Californian creed of it in 2010 when he described technology as a seventh kingdom of life with wants of its own. E/acc adds the thermodynamics and keeps the conclusion, that machines set their own course and people follow, which no evidence could ever count against. A cost, at least, can be measured. A destiny cannot. That is why e/acc should be read as a doctrine rather than a description, one that asks only to be believed, and why it suits the people selling and funding the machines: a public told that turning back cannot be done stops asking what it might cost, and the alternative becomes unthinkable to the very people who might otherwise build it. Each front has a remedy, and each remedy has a price. A state that chooses to can still fund a public alternative while the gap is small enough to close. Institutions can keep pace with machine-speed actors if they are prepared to pay for a tempo they were never built for. A monopoly over what a society knows can be challenged in law: in 2024 an American court ruled that Google had illegally maintained its dominance of search, though in September 2025 the judge, Amit Mehta, declined to force a sale of its Chrome browser, partly because generative AI might yet shake up the market on its own, so the law reaches these companies but moves cautiously. Any state can tax the returns flowing to the owners of compute and pass them to the people displaced, at a price that rises with every share of output running through machines it does not own. And a government can take back the arrangements by which it decides, starting with its dependence on a single supplier. In 2025, after its chief prosecutor’s Microsoft email was cut off under American sanctions (the company says it did not suspend the court’s services), the International Criminal Court in The Hague, in the Netherlands, announced it would move its staff to openDesk, a German open-source suite, and a country could do the same. These are the state’s ordinary tools, courts, taxes and procurement, and what they buy is time. What they cannot produce on any timetable is a network that no single party can seize or shut down. One such network has existed since 2009. It has been attacked, forked, banned outright in China in 2021 and regulated at its edges almost everywhere else, and it has outlived every attempt to switch it off. Nothing like it can be legislated into existence or ordered up within a term of office. That is why the ordinary tools matter here: they hold the window open while what cannot be commanded is built as an alternative, and the building has to begin now even though it will mature later. The dependence runs below the state as well. In October 2026 OpenAI’s chief executive said data centres that build and run themselves with robots were within grasp, which is the plainest statement yet of what the clock measures, since sites, electricity and consent would then go unasked for. A month later voters in ten states put limits on those facilities, nearly all of them conditions on power, water or the agreement of the communities: the machine still needs all three, and those who hold them were legible as holders of claims and had to be answered. What turns these fronts from a list of risks into a deadline is that every remedy is cheaper now than it will be later, and all the bills are rising at once. Each of the three markers is a reading of the dial set out at the start of this chapter, a point at which the machine needs a government’s consent a little less. We found earlier that subsumption gives a society a small handful of years to adapt, and nothing here suggests the window for changing course is any wider. The deadline is dated by dependence rather than by any laboratory’s lead, because dependence is what keeps a person inside the process with the standing to change it: while the machine needs the engineer, the engineer can refuse; while the state is a customer, the supplier can be dismissed; while the model needs your corrections, your judgement still counts. That is why a pause at the frontier of the kind the laboratories have proposed for themselves fixes the leaderboard without stopping the clock: deployment carries on, and so does the handing over of capabilities to the machine unabated. The systems still need inputs they cannot yet make — the records that sit inside firms and public bodies, and the long tail of people using the products, whose corrections train the next version. They are beginning to produce their own, helping to design chips, generate synthetic training data and write the code for their successors. Until those loops close, people who decide the effort is worth paying for can still change course. After they do, the substrate runs itself and the rest of us are passengers arguing over the map, and what put us there was not a law of nature but every year in which nobody would pay the price. That is how political impossibility is manufactured. *** All of this invites an obvious objection, because pleas of urgency have a long and disreputable history of arriving at the same destination. The situation is grave, deliberation is slow, therefore someone must be empowered to act without it. It is the standing case against democratic government, made by one regime after another that dispensed with it. Its most rigorous statement came from Carl Schmitt, the German jurist who joined the Nazi Party in 1933, in the formula that sovereign is he who decides on the exception. The Reichstag burned on the night of 27 February 1933, and the decree signed the next day, which suspended the civil liberties of the Weimar constitution in the name of an emergency and was never lifted, is what that principle looks like in practice. We saw on the securitisation road how such powers outlive the crisis that justified them. The present-day version needs no fire. The day an agent swarm takes down a payments network, or crashes a market, the decree will already be drafted, and a democracy need not be weak to sign it. At the time of writing a bill before Congress does this ahead of the capability it describes: it would pause the frontier under a new department able to order a system’s destruction and to imprison for up to twenty years. An argument that a window is closing can be put to exactly that use. The difference lies in what the urgency permits. Schmitt’s exception allows the suspension of deliberation and the concentration of authority in whoever is fastest, so that the emergency becomes the reason debate stops. What is urgent here is a construction rather than a seizure, additive and voluntary, entered into by people who choose it and worth nothing unless a great many of them do. It needs builders of an alternative, not power handed to some authority to act on our behalf. Urgency of this kind gives no case for a strong state, for a reason no such power can do anything about: the work takes a decade or more, and none of it can be commanded. That is the work I am calling for, and it is constitutional before it is technical. The governor will not install itself, and it will certainly not be put there by those denying its importance. Its simplest form is a person the system cannot act without, holding claims it has to answer — not a name kept in the loop to take the blame when things go wrong. Be legible to the system as a task and you will be subsumed; be legible as a holder of claims and you will have to be answered. It has to be written deliberately into the substrate, into code and protocol as well as statute, by the people I have called constitutionalists, who are not a profession but anyone willing to shoulder it. The steersman at the oar, from whom cybernetics and government alike take their name, was always a person. What stops anyone paying, and so closes the window for good, is the belief that no such choice was ever available, which is not a fact but a story with an author. Why the steersman was thrown overboard, by whom, and with what justification, has to be answered before any governor can be built back.