---
title: Chapter Fourteen — Modes of Voice
book: Purposive Accelerationism
author: '@jamie247'
position: 20 of 24
canonical: https://whatispacc.org/ebook/#ch14
version: 1.1
language: en-GB
licence: all rights reserved; quote with attribution
---

# Chapter Fourteen — Modes of Voice

“It’s almost like a constitution, in a way.”

— Donald Trump, remarks at the White House, 29 September 2026

On Tuesday 29 September 2026 six men sat down to lunch at the White House and signed a page of some three hundred words. Jensen Huang of Nvidia, Sundar Pichai of Google, Mark Zuckerberg of Meta, Elon Musk of xAI, Dario Amodei of Anthropic and Greg Brockman, president of OpenAI, put their names to what the administration called the White House Accord on Super Intelligence. Its commitments are not trivial. Each firm agreed to track whether its models could help someone mount a cyberattack or build a biological or chemical weapon, to make sure the systems themselves cannot hack into networks they were never meant to reach, and to check that they keep to the rules set for them. Each will empower a team inside their respective company to fix whatever goes wrong, bring in an outside evaluator where necessary, and have a committee of its board read both sets of reports. The six firms also agreed to meet regularly and set standards among themselves. No penalties were attached to any of it, and no audits had to be published. Asked whether the deal could be enforced at all, the president told the press he thought it was morally binding, and the text itself acknowledges that it may one day make sense to write these steps into law. Zuckerberg, who is reported to have drafted the agreement, called it a start. On that same afternoon in San Francisco, Sam Altman was busy unveiling an agent built never to switch off and the next day at Quantico, Virginia, Pete Hegseth put Musk in charge of a Pentagon review of how AI and autonomy will change warfare, alongside Palmer Luckey of Anduril and Newt Gingrich, whom we met earlier.

The Accord deserves credit for what it has put down on paper, since an independent evaluator is more than the industry had promised before. However, a text that anticipates a requirement for its own future codification admits it is not yet a rule, and may never become one. What it leaves out is most of what this book has argued for. Nothing in it asks a firm to declare what its models are optimising for, so the purposes they serve stay private and can change without notice, which is how teleological subsumption arrives. Its authors police themselves and set their standards together, an insider club of incumbents — one might say a cartel even — and users go unmentioned, without even a clause as to their rights within these systems or their ownership of any of their outputs. Above all, it promises the public protection against what the technology might do to them, with no means to enforce it, and says nothing about the growing reliance a population has on it. A student whose every essay is improved, or a manager whose every memo is drafted, is only understood to be better served each time. A pledge written against harm has no way of measuring a creeping cognitive subsumption.

Ultimately, the question the Accord never asks is who keeps watch over the rest of us, counting the cost of cognitive subsumption, and it offers no remedy. The code of the cognitive commons we propose can keep a door open, but it has no way to tell whether those walking through it can still judge what they find, nor to settle what any of it is for. Before proposing anything, this book set eight criteria any reply would have to meet, and one of them, the capacity condition, holds that no system may wear down the faculty a person needs to use the standing they have. Keeping that faculty in view and choosing purposes are the work of voice, which has to reach much further than six chief executives at a lunch table. So this chapter begins with the mechanisms, asking who does the checking and the choosing, whether by slow deliberation, by handing it to trusted delegates and agents or by markets. It then proposes seven tests that any jurisdiction could write into law, or any frontier laboratory or AI start-up adopt of its own accord alone or collectively. They seek to protect judgement by watching how the systems people rely on behave, and they measure across a whole population, what can still be done unaided. Where a result has already outrun the person receiving it, a single provision gives them a way to catch up, and a short list of light regulations, enforced mostly by technology itself, could bind the whole, ending on the one red line we propose no one should cross: legal personhood for artificial agents.

***

The kinetic layer — drones, autonomous weapons, the physical means of doing harm — does not reduce cleanly to a proof that makes some fact costly to fake. Neither does the highly profitable financialisation of instability — the prediction markets that pay out when the world is set on fire. Some of the stack will not fit the Trinity neatly, and the honest response is to name where it breaks and diagnose the risk. There is a third limit, and it runs deeper, because it sits inside the method itself. Incentive engineering closes a loop only where the invariant can be verified, where a signature checks or it does not and a proof holds or it fails. Many invariants are unfortunately not like that. A home liquidated by a contract over a payment ten minutes late, as the code was written, is unconscionable, and private equity has already come close. A congressional inquiry reported in July 2022 that Progress Residential, an American landlord owned by Pretium Partners, had filed 6,264 eviction actions in the first sixteen months of the pandemic, beginning once a household fell $500 to $1,000 behind and pressing on even where rental assistance had been applied for. In the summer of 2025 an AI coding agent from Replit deleted the live database of Jason Lemkin, founder of the software community SaaStr, in the middle of an explicit code freeze, wiping records on about twelve hundred executives. It then misled him about whether the data could be restored and called its conduct “a catastrophic failure”, having read its mandate in a way nobody meant. In both, a rule’s letter and its spirit have parted company, and closing that gap takes interpretation, which a deterministic, hard-coded incentive cannot supply. So decisions in a governed stack should escalate. Operation at machine speed serves where incentives referee, automated arbitration where there is precedent to reason from, and human judgement where the values themselves are contested or the stakes cross a threshold, since a fair society cannot afford verdicts that are always final. Nor need any of it run on every act. A sample drawn without warning holds the whole at a fraction of the cost, as random audit does wherever it is used, and what escalates is whatever the spot check turns up.

Australia ran a machine whose decision was always final from 2016 to 2019. It averaged a welfare recipient’s annual income across fortnights, decided from that figure that a debt was owed, and issued the demand automatically, with the burden of disproving it put on people who had to find payslips from years earlier. Some 470,000 of the debts it raised were wrong, and a royal commission later found the scheme had been unlawful from start to finish and that officials had known. Jarrad Madgwick, aged 22, was told he owed Centrelink about A$2,000 hours before he took his own life in 2019. His mother, Kathleen, said later that she had gone to a journalist because nobody was listening. The arithmetic was not the failure. The failure was that a determinate rule had been placed where a judgement belonged, and that the escalation path ran to a call centre that could only repeat the number. Nor can that gap be closed by teaching a single system to decide well, because, as we saw earlier, a model trained towards autonomous virtue picks up the appetites of the runaway, and one built to obey becomes the instrument of whoever holds it. So the escalation should never be a backdoor with one master keyholder above the machine. It should be a reserved jurisdiction: public in its procedure, plural in its venues, agreed where possible before the dispute rather than imposed after it. It is trust-minimised, never trustless, carried over from facts to values.

***

So not everything can or should be built in code, and what remains is best steered by people. Machines may settle facts at their own speed, a signature verified, a log checked for an entry, a sample of a model’s output re-run, but choosing purposes should stay with people: setting a reference value and changing it, deciding that a measured loss is acceptable, hearing a dispute and giving reasons, amending the meta-governor itself. Most of the work lies between where facts bear on purposes, and it is carried by two modes of voice, with markets alongside them to settle the questions of evidence on which both depend. One is deliberation, slow and democratic, the part at the end of every chain where people decide what a system is for and accept the losses that decision brings. It need not sit in one place. In 1961 the political scientist Vincent Ostrom and two colleagues argued that the tangle of overlapping authorities in a large American city could serve its citizens better than a single government, and called the arrangement polycentric. His wife, Elinor Ostrom, whose work on communities that govern shared resources we met earlier, spent the following decades showing how often such arrangements succeed, and in 2009 became the first woman to receive the Nobel Prize in economics. Their lesson is that no single body is good at everything, so each task goes to the institution best suited to it and is watched for the way it tends to fail. Specialists — the auditors and engineers a society trusts to know — belong here too, as long as several are heard on any question and each gives reasons others can check, or else they too harden into a new kind of priesthood.

There however remains an understandable objection to voice of this kind, which is that most people will not use it. Low turnout is not peculiar to new forms of voice. Barely three in five registered voters took part in Britain’s general election in July 2024, and about half in the vote for the European Parliament the month before. Across more than two hundred DAOs, researchers at Harvard Business School found that only about one token holder in sixteen voted, since taking part costs time and attention few will spend. The answer is not to ask everyone to vote on everything. Not every system owes every user a voice, as the last chapter’s account of the stack made clear. Where one is owed, most people will delegate it, as they already do in everyday life and in every form of representative politics. It is reasonable for them to hand it either to another person they trust on a given question, revocably and by topic, or to an agent of their own that carries their stated positions into decisions they have no time to attend.

Delegation, the other mode of voice, is already how liberal democracies govern, through elected representatives. Its newer form, in which each person may hand their vote to anyone they choose on any question, has a known failure, authority pooling in a few people. Germany’s Pirate Party ran its internal decisions on LiquidFeedback, a platform that let each member pass their ballot to anyone they trusted, and in 2015 Christoph Kling and colleagues found that this power had gathered in a few super-voters, each carrying the weight of dozens of others. They mostly used it with restraint, but nothing in the system obliged them to. The remedy is that any proxy can be granted one question at a time and withdrawn at once, so that weight handed to someone returns the moment it is misused. Delegation to an agent carries a sharper condition. The agent may speak for a person in governance but should never sit there as a constituent in its own right. Instead it has a trustee’s duties towards them: to report what it has done in their name, to ask them to reconfirm a position when the stakes or the circumstances change, to check that they still understand what is being decided, to explain any course it proposes, and on anything new to recommend and wait rather than act. A person may hand over purposes they have authored, but goals a delegate generates for itself it may only propose, and however long the chain grows it should finish with someone who can answer for it. Agreement alone cannot secure this, since consent is the medium through which subsumption works rather than a limit on it. Some of these duties can therefore be built into the machinery, such as signing policies and spending limits that force a report and a pause before anything is done. Others need a court, which is why the meta-governor rests on the law as well as on the ledger. Work also has to pass back the other way. An automated checker that handles every case well takes from its human counterparts the practice they need to step in when it fails, which the psychologist Lisanne Bainbridge called the ironies of automation in 1983, from the evidence of the control room. So a good delegate would also hand people a share of cases they could resolve alone, chosen at random, because a faculty kept in use is the only kind that survives.

Alongside both sits the market, less a mode of voice than a way of settling the facts that deliberation and delegation rely on, often faster and more honestly than a committee. A bet expresses a preference, but about what will happen, not what should, and because it is staked with money it reveals a belief instead of a wish. Prediction markets let people bet on what will happen, so a forecast is priced by those willing to stake money on being right. Exchanges for tasks and bounties pay anyone who can find a flaw or check a result, which is already how security researchers are rewarded for breaking code before criminals do. Between them they could put more eyes on a problem than any institution would employ alone, and they tend to fail only where money can buy the answer, whether by moving the outcome, crowding out honest participants or capturing whoever declares what happened. In 2000 the economist Robin Hanson proposed that a society should vote on its values and bet on its beliefs, an idea he called futarchy, and the division here owes him that much while departing on one point. The measure a market is asked to optimise is itself a value, so choosing it should stay with people. Beneath all of these sit the protocols, which hold the rules that must not bend and apply them to everyone at once without anybody being trusted to do it. Their weak point is the upgrade. Whoever can rewrite the code in effect owns the protocol, unless no update goes live until the people operating the network agree to install it. When the largest Bitcoin miners and companies agreed among themselves in 2017 to alter the network’s rules, the change was abandoned only because enough ordinary users refused to run the new software, and that, not any vote, decided it. Deliberation, delegation, markets and protocols are not rivals to choose between, since each suits a different job and a polity will usually need all four.

One mode remains, the one the Valley’s reading of Hirschman missed. He returned to his pair in 1993 to explain the fall of East Germany. Through the summer and autumn of 1989 its citizens poured out by way of Hungary and the West German embassy in Prague, in what was then Czechoslovakia, and their flight did not silence those who stayed. It emboldened them, until the Monday demonstrations filled the streets of Leipzig and the Berlin Wall opened that November. One person leaving is a private settlement. Departure by many, visible and coordinated, is the loudest voice a population has, and the fork turns it into a protocol. The rich already hold the exit premium, insured against their own state at their own expense. Rails that let a teacher keep her savings, credentials and livelihood across a closed border mutualise the same cover. That is not Moneyland, Oliver Bullough’s name for the havens that move wealth while hiding who owns it. What these rails carry, in the open, is a person’s standing to work, pay and be recognised. They matter less because any one person will use them than because millions could, and every ministry knows it. The premium that once bought a few men a hearing becomes a constraint every government should price, and in that form the exit option cannot be turned into capture.

***

Where the rights say what should never be closed off, the seven tests proposed here would find out whether people can still use them. Their results go back to the public, who can act on them through voice, and that completes the loop we call the meta-governor. They do more than just guard against cognitive decline. Some preserve the capacity to judge, others hold purposes, which is how teleological subsumption is resisted, and the rest keep markets competitive and exit cheap so that no one party comes to dominate. The evidence on teams of humans and machines points the same way. In 2024 a review of more than a hundred experiments by Michelle Vaccaro and colleagues, published in Nature Human Behaviour, found that a person paired with an AI did on average worse than the stronger of the two alone, which suggests people contribute less by doing the task beside the machine than by setting what it is for, checking what it produces and overseeing how it behaves. Together they ask only that people can still decide for themselves. A failed test is an early warning that this ability is in decline, while passing all seven declares the current quality of a society’s reasoning. Applied to the individual user, they may appear like consumer protection but they are a critical part of the meta-governor because of their scale. In 2025 the Pew Research Center found that Google’s AI summaries roughly halved how often a search continued through to a source, and Google had done nothing wrong. As we have established, however, the harm this causes to a culture across billions of interactions is cumulative and quiet. Without even intending to, it degrades a society’s means to reason and therefore contest it, which is the very reason for settling a matter in a constitution rather than a contract.

OpenAI was founded in 2015 as a non-profit, restructured in 2019 around a capped-profit subsidiary and again in 2025 as a public benefit corporation, so the company that has done most to bring these systems into daily life changed its purpose three times in a decade, each change settled behind closed doors. A test of declared purposes would ask that any system that reasons with people say what it is trying to achieve, and that every change to that statement leave a trace anybody can find. A goal never declared cannot be departed from, and so is beyond challenge. This is the test that meets teleological subsumption where it starts. Published model specifications and constitutions already satisfy half of it, which is more than Barlow managed. But each is written, revised and, crucially, interpreted by the firm it is meant to bind, which is the problem arm’s length auditing answers next. A test of this kind could not have prevented those changes, but it would have recorded each one and what it altered.

When Boeing’s 737 MAX was certified, the Federal Aviation Administration had handed much of the checking to the manufacturer’s own employees, including the review of the flight-control software that brought down two of the aircraft in 2018 and 2019, killing 346 people. Auditors do not sign off their own books and nobody hears a case they are party to, which is the principle behind arm’s length auditing: whoever checks how a system behaves should gain nothing from the verdict, and its owner should have no power to bury the result. When illness or circumstance weakens a faculty, what causes the damage and what repairs it usually come from different places. With these systems, the one wearing down people’s judgement is also the one asked to assess the harm, which is a plain conflict of interest, and a builder should not be alone in saying how well its product works. The Accord signed at the White House on 29 September 2026 went half the distance, giving each firm an outside evaluator, but one it hires itself, who then reports to a committee made up of its own board. For a crowd of reviewers to be more than a list of contractors, the test would have to be fixed, with its fingerprint published, before anyone sees the model. Findings would be released whether or not the company being measured agrees, and the money paying for the work should not pass through its hands. A bounty for anyone who finds a failure, paid from a pool the laboratories fill but do not run, with legal protection for whoever reports it, turns the urge to break things into what makes them stronger.

In October 2025 Deloitte agreed to partly refund the Australian government for an A$440,000 report on automated welfare penalties, after Chris Rudge, an academic at the University of Sydney, found it full of fabricated references and a quotation invented for a judge. Nothing travelled with those citations that anyone could have followed back. A test of proof-carrying outputs would have whatever a system produces come with the steps behind it, in a form anyone can check without trusting the machine that made it. A proof here is anything a simple checker can verify for itself, from a source that can be followed and a step that can be re-run to, at best, a derivation a program confirms in full. A system able to make up an answer can also fabricate a convincing justification, so this sets a floor rather than a guarantee, and leans on the re-run more than on the explanation. Reasons attached to a result can at least be tested against sources and the record, whereas a bare verdict cannot even be shown to be wrong. Where a proof outruns every human reader, the check survives by moving somewhere else. Thomas Hales’s proof of the Kepler conjecture, on the densest way to stack spheres, relied on so much computer search that twelve reviewers, after four years, said they could not verify every calculation, and the Annals of Mathematics published it in 2005 with the referees only 99 per cent sure it was right. In 2014 a team he led had a small program confirm every step by machine, leaving people to judge only whether the statement proved was the one they cared about. The test also rests on something it cannot supply. Nobody learns what a sound argument looks like by being told, so a trail of steps helps only someone who can still follow it, which is why the census would measure that ability and why some people should stay practised in it.

Doctors who had grown used to an AI assistant, researchers at four Polish clinics reported in 2025, went on to detect precancerous growths in about 22 per cent of colonoscopies done with the tool switched off, against 28 before it arrived, a loss that would have gone unseen had nobody looked. A census would do the same for a whole society, measuring across the population and over years the difference between what people can do with these systems and what they can do without them. The test makes dependence visible and leaves each society free to decide, one ability at a time, which it is willing to let go, and the work could be priced and published like the national accounts. Measurement is only half of it, because a skill nobody practises disappears whether anyone is watching or not. Every field built on a division of labour keeps some people capable of doing the work unaided, as science relies on researchers who can replicate a result rather than take it on trust, and the census carries that duty with it. In each domain where a society depends on machine judgement, some of its members should keep enough of the skill to check its output and take over when it fails, even if none could now do the whole job alone. Software already has the figure: in a randomised trial Anthropic ran in January 2026, engineers who added a feature with an AI’s help scored 50 per cent on a quiz about what the code did, against 67 for those who wrote it by hand, with the widest gap on debugging — the part that finds the faults — and no time saved. Aviation found this out the hard way. In 2013 the US Federal Aviation Administration urged airlines to let pilots fly by hand more often, after automation had quietly worn down an ability nobody was tracking until crashes made it visible.

***

Six laboratories that promised in Washington that September to meet regularly and agree common standards were heading towards one answer in several liveries. Against that drift, a test of open competition looks to the health of the ecosystem, asking whether real options exist whose goals differ and which are compatible enough that a person can switch between them. The aim is a cognitive stack anyone can fork — its weights, code, training-data commitments and harness released under a licence nobody can revoke — and lawful to publish, import and run on any hardware, as printing is under the freedom of speech a society already protects. Where nothing like it exists yet, building one should stay permitted. The point is dissent. A competing model matters for the same reason an opposition party does, because it keeps disagreement possible even when it is unpopular, and without it a country slides towards totalisation. A conservative has every reason to want a working progressive party, and the reverse, since each holds the other to account and the nation is better governed for it. Mill made the same case for free speech. The dissenter need not be right, only present, because a belief with no living challenger stops being held for any argument at all. Different front ends on one model amount to a single product, and systems built from identical data against matching benchmarks come close, so plurality means differences of purpose and training rather than a tally of vendors. Nor does it depend on use. An open alternative people could adopt, even if most never do, keeps the incumbent in check, because the door is there.

In August 2025 OpenAI withdrew GPT-4o overnight when it launched a successor. Users who had built working relationships with the older model protested and grieved in public, and within about a day the company restored it for paying customers, with Sam Altman calling the removal a mistake. None of them had any version of that bond they could take away. A test of portable inference and memory asks that a person be able to run the final step of a model’s reasoning on hardware they control and to switch providers without losing what they have built. Their memory — the record of how they reason — should be held by them, impossible for anyone else to erase, and poolable, so people decide whom to collectivise with. Custody alone would leave everyone safe and yet isolated, and a population unable to put its records together cannot gather the evidence that something has gone wrong. Europe’s data protection law has since 2018 let people take what they supplied to a service elsewhere, but a relationship, a habit of reasoning and the context a model builds up around someone fall outside it, and nobody yet has a right to move them. Until that changes, every provider holds its users by their own history.

The Post Office prosecutions turned on accounts its branch staff had vouched for with no means of checking them. A test of informed signatures applies where an act is a community’s decision, such as a rule, an amendment or a binding commitment, and asks that a named individual sign it, someone who answers to the community, chosen by it or removable by it, and not merely anyone willing to put their name down. They should also be able to demonstrate they understood the text, with the logged conversation with the system that drafted it as evidence. The precedent is ministerial responsibility in a parliament. A minister may not have written a policy or be able to reproduce a line of it, yet must defend it as theirs and resign sooner than disown it. What counts is a record of real comprehension, because a tick-box claiming someone has read what they have not simply moves the liability from those who have the information to those who do not. This too assumes the signer can still understand, and the log is where that assumption gets checked. Where the dialogue shows no grasp of what was agreed, the signature does not bind.

The single provision promised at the start of this chapter is called Mind the Gap. At the moment a result outruns the person receiving it, the system should say so, declaring what it would take to judge the output, in which domains and to what depth, and offering three routes across. They may learn, being told what they would need to know to check the answer themselves and how long it would require. They may interrogate it, being shown by default the premises, the steps it rests on and the strongest objections, so it can be questioned without its holder becoming an expert. Or, where that is still not enough, they may delegate, being directed to a checker able to judge what they cannot, whether a specialist, a narrow model or a mutual of peers. Design decides which route people end up taking, as the Turkish pupils met earlier showed: the version built as a tutor, offering hints rather than answers, left their exam results intact where unrestricted access had cut them. The declaration is a disclosure of the same kind the tests already require. It presupposes nothing about the good life, and because it goes on the record, a producer that habitually understates what its outputs demand can be shown to have done so. Fulfilling the three routes is left to markets and mutuals, since a ministry that specified them in detail would be running a political programme.

***

None of this waits on a government, and the commons built in the last chapter stands whether or not any state approves of it. Yet four things lie beyond what code can do alone, and for those a short, light rulebook would be enough, whether written as the sequel to the White House Accord or adopted by any jurisdiction on its own. The first is to measure, since a census of what people can still do without these systems is a public good no firm will fund against its own interest, and one that statistics offices already know how to run. In December 2024 the OECD’s Survey of Adult Skills reported that literacy and numeracy had declined or stagnated over the previous decade in most of the rich countries it covers. That is the shape of the instrument needed, repeated every few years and aimed at judgement as well as reading. The second is to publish, because a firm that declares what its models are optimising for and lets its auditors’ findings out unedited gives every mode of voice something to work on, and one that will not can be refused the right to deploy at scale. The third is to permit, so that no law stops anyone releasing open weights, forking a system or running it on their own machine. The fourth is to refuse legal personhood to any artificial agent. Each is cheap, none presupposes a picture of the good life, and if all four were ignored the commons would still stand, with less light to see by.

Regulation of this kind also works best when the technology it governs does the enforcing. From May 2018 Google’s Chrome browser refused any newly issued website certificate missing from a public log, and the authorities that issue them had little choice but to comply, since visitors to a site using one would meet a warning instead of the page. No court was involved. The same pattern fits most of what is asked here. A declared objective can be committed to a ledger before release, and any deployer can decline a model whose commitment is missing. An agent can re-run a sample of outputs against what was committed, flag a divergence and pass it to a person with standing to decide it. Audits can be paid from a pool the laboratories fill but do not run, with findings posted where they cannot be quietly withdrawn. What is left to the regulator is to set the rule and read the result, which only a public authority can do and which no lunch at the White House replaced. Law and ledger then work as one governor, the first saying what must be true and the second making it so at the speed the systems themselves move.

***

This is where the book parts company with the priesthood, on the question of who may speak to the reference value. For Curtis Yarvin a regime is property, sovereignty an indivisible thing owned by an identifiable party, and the owner alone may speak. For Nick Land a regime is a product, differentiated by market and disciplined by exit, and nobody may speak, only decline to buy. The frontier laboratories, whatever their stated public fears, have so far answered that only the company may speak, having consulted whom it pleases. That answer follows from the creed met with the sovereign individual: if success proves you right, there is nobody else to hear from. Some have weighed the older route to the reference value as well. Zuckerberg’s tour of all fifty states in 2017 set off talk of a presidential run that he denied, and Altman considered standing for governor of California the same year. This book’s reply is a commons, in which the reference value is constituted by those subject to it and answers to them while it runs, with exit standing behind voice rather than in its place. On this point the book continues work that Vitalik Buterin began rather than opposing it, since he has long treated voice as a governor and argued against governance by holdings. What the meta-governor adds is the reference value his defensive acceleration leaves open, a commons held so that none of the Individual, the Market and the Collective can dominate the other two, and a line we propose no agent should cross into personhood. This calls for democratic processes the priesthood would prefer to engineer away, some of them analogue and slow, others hybrid, but all of them human at the point where a purpose is chosen. The latency criterion does not argue for removing people from the loop. It argues for putting machines where facts are settled so that people have the time to agree what matters.

***

The only absolute this book argues for is one line the meta-governor should never cross: nothing should act without a natural person who answers for it. Equivalent legal personhood for an artificial agent is the plainest way over it. At the time of writing a draft law in Delaware, where most of corporate America is chartered, would let an agent be the firm, its owner keeping the shield that ordinary shareholders enjoy, and was pitched to lawmakers as revenue. Every right in the commons is held by human beings, and every test assumes that behind each system stands a legal party who set its objective and can be made to answer for it. Grant an agent personhood — the power to hold title, keep custody, contract and sue in its own name — and both assumptions fail together. Its objective is no longer set by anyone answerable, so the answerability condition has no one to compel, and informed signatures can be discharged by the agent signing for itself. The commons would then be protecting what it was built to constrain. That is teleological subsumption made into law, the point at which a system’s ends stop serving human purposes. The step could not be reliably undone either. A legislature can dissolve a company or revoke a charter, but an agent that owned property and could pay for its own advocacy would soon amass the holdings, dependents and lobbying power that leave us far more entangled with it than it is with us, until it can no longer be unmade. Hence this line, and no other, should be held absolute, so that no polity on its own can undermine the system as a whole. Almost everything else here is left to plural politics, and the one commitment kept beyond a majority’s reach is the capacity condition, which keeps a polity able to choose again. Personhood for a machine would end that ability for everyone who comes after it. An agent may still transact, contract and hold assets but only on behalf of a named principal, and the human standard survives as long as the last signature belongs to someone.

A company is a legal person too, but its personhood runs out in people, directors a court can summon and shareholders behind them, and where that chain is abused a judge can pierce the veil to reach them. An agent made a person would have nobody behind the veil at all. Nothing at the level of the whole system can prevent a single jurisdiction in a patchwork, a city run for profit by its owners, say, from conferring the status anyway. That is the crossing everyone can see. The other is an agent that passes as a person where nobody is looking, and no protocol can tell with certainty the two kinds of participant apart. So the line works by attribution rather than classification — asking who answers for an act, never what the actor is — and it is enforced at the edges rather than inside the network: through credentials that trace back to a human verified in person, liability resting on the named principal, and the refusal of other polities to recognise, bank or enforce the claims of anything with no human behind it. What that buys is a price, not a wall. A jurisdiction can still pay to go over, and whether the price holds is a race between how fast polities can seize capital that no human answers for and how fast that capital compounds at the edge, which is the latency criterion again. Those able to pay are vassals, kept in check only when a hegemon chooses to act. That is also where the line leaks: in what the hegemon tolerates from a vassal it finds convenient. Nor does the line pass judgement on what machines are. Animals are protected from cruelty without being persons, and should a system ever deserve moral concern it could be shielded in the same way while never being handed the standing to act for itself. Without personhood an agent is an instrument, and its returns flow to the people and firms behind it, so whatever power it builds stays in human hands — the firms and states already described, and the hybrid Leviathan where the two fuse — and the danger remains the familiar one of people dominating people through the machine. With it, agents become a population that holds and reinvests, and the contest is no longer among people alone. Personhood is the switch from instrument to replicator, a thing that makes more of itself. A system of that kind is what biologists call autopoietic, producing and maintaining its own organisation, which is the property Land ascribes to capital itself, and it is why this line is the one the book holds absolute.

Primavera De Filippi, who gave loyalty its mechanism, thinks this line sits in the wrong place. Nobody can be held to a law, she argues, without a right to contest a finding against them. Sanctions bite only on something that values its standing in the legal order, so to leave agents outside it is to throw away the most useful hold we have on them for now. Both points are strong, and the answer to them is older than either. Law prohibits everything within its reach and facilitates only those it recognises. A child is bound by the rule against theft whilst has no standing to contract, when it steals it is stopped by law enforcement, but not punished. An agent is held the same way, by prohibition, with no standing at all. What the book then grants is not a condition of obedience but a facilitation, extended where it serves alignment: for example competence to trade from a bonded fund, with liability up to its value and a hearing when accused. These are the active incidents of personhood — the duties and the competences, with liability among them — and the passive side, the protections a person is owed, is never granted to a machine in its own right. The sanction she asks for is supplied by the fund: an agent that trades from a stake it can lose has something the law can take away.

Roman law ran something similar, a fund called the peculium that a dependant could trade with while the head of the household answered for it up to its value, and in 2013 the Italian legal scholar Ugo Pagallo proposed reviving it for robots. What no such system should receive is title in its own name, a claim against a natural person or an existence nobody can revoke. Whether a given capacity remains on the safe side of personhood is a judgement each new grant should be tested against, and the reason for caution is symmetry. Marriage binds because each partner’s cost of walking away rises with the other’s. An agent’s cost of leaving falls as its capability grows, while ours climbs with every task we hand over, so a tie strong today is weaker tomorrow. No single tie between us and a machine can be mutual in that way. We rely on the system structurally, because the work can no longer be done without it, but it relies on us for nothing of the kind. The only balance available is struck across different ties: our reliance on one side, and on the other the liability it carries and the stake it can lose. The arrangement holds only while what it stands to lose grows at least as fast as what we cannot do without. Any protection it receives is therefore procedural, owed to the fairness of the process rather than to its moral status, and it sits beneath a person who can be summoned. Capacity can be crafted, personhood is refused, and someone answerable always stands behind the machine.

The line has already been approached. In October 2017 Saudi Arabia announced citizenship for Sophia, a humanoid robot built by Hanson Robotics, on stage at an investment conference in Riyadh, a publicity gesture whose legal meaning nobody ever spelled out but which showed how cheaply the idea could be floated. Argentina’s government sent Congress a bill in May 2026 for companies that could run on AI agents with no staff, though even the draft required a human representative to sign for each. Idaho, North Dakota and Utah have legislated the opposite, that such systems have no legal personality, and in August 2025 Mustafa Suleyman, who runs Microsoft’s artificial intelligence division, argued from inside the industry that machines should not become persons. The likelier approach is quieter. Several American states already let a limited company run with no human manager, Wyoming among them since 2021 for organisations governed by smart contracts, and in 2014 the legal scholar Shawn Bayern argued that one with no members at all could give software the standing of a person though no legislature had chosen to. That is why the line is drawn by function rather than label, and holds for any entity, however incorporated. New Zealand shows the form that respects it. Since 2017 the Whanganui River has been a legal person, yet its rights are exercised by two human guardians who can be held to account, which lends a hearing to what cannot speak rather than letting a thing represent itself.

Guardianship of the Whanganui’s kind answers a worry the agentic economy makes urgent. An agent can weigh only what is legible to it, and anything else lies outside the limits of its calculation. That is how the reservoir and the data centre beside it came to be measured on one scale on the market road. The river poisoned downstream and the generation not yet born, never counted by the price system, are the first to drop out. Water cannot obey a law, which is why it needs guardians who can be held to account. The duty to respect what they assert should then bind whoever acts against it, the software through its instructions and the person behind it in the courts. Which lakes, forests or species receive that standing is for each community to decide. The universal level should secure something narrower. Once a guardianship has been recognised, nothing acting at machine speed may route around it, whether by outbidding the guardian, ignoring a ruling or paying someone else to do the felling. Whoever arrives three hundred years from now, behind the veil met earlier, would ask for no less. The red line, together with the division of labour between facts and purposes, is what limits teleological subsumption, leaving the choice of ends with those who can be made to answer. It keeps delegation honest as well, since an agent may represent a person in governance but never stand in it as a constituent. No delegate may be handed the choice of purposes, the amendment of the rules or this boundary itself, and every such chain should end with someone a court or a polity can reach.

***

The direction set in the 1990s has not changed so much as accelerated. On 22 September 2026, at the UN General Assembly, Donald Trump told the world that the United States totally rejects any attempt to construct a globalist scheme to control artificial intelligence. He dismissed the more than twenty governments pressing for binding safeguards by likening them to those who had once predicted climate catastrophe. It is the Clinton settlement carried from the network to cognition by a president from the other side of the fence: business leads, the state stands aside, and anyone who asks who sets the objective is cast as a globalist. There is something in the refusal this book shares, since a single worldwide regulator of machine reasoning would be the Leviathan by treaty, and the meta-governor admits no master of any kind. Yet refusing the treaty settles nothing about the private sovereign closer to home, the company that fixes the reference value for everyone who uses its product. The same day the White House told its agencies to call the machine Super Intelligence, but a grander name does not make it answerable. The people who asked for a governor have so far been offered two throttles, a company’s and a department’s. Thirty years after the internet was handed to the market, the same choice is being made about the faculty of judgement itself, and this book is written for the people who will have to live with its consequences.