Yuval Noah Harari has spent the summer warning that the one decision humanity will not be able to take back is granting legal personhood to AI. His argument is elegant. AI is a bureaucratic native — better than us at finance, at law, at the paperwork that actually runs the world — and the only thing keeping it out of the control room is that it is not, in the eyes of any legal system, a person. Argentina is now proposing to change that. Once the fox enters the coop, in his phrase, the coop is finished.

I think the diagnosis is right and the door he wants us to guard is the wrong door. It was opened in 1897, when the House of Lords decided in Salomon v Salomon that a company is a person distinct from the people who own it. Every jurisdiction I have practised in has lived comfortably with that fiction since. The GmbH, the AG, the FlexCo, the Delaware corporation: each one can hold assets, sign contracts, sue, be sued, and donate to a party. None of them has a pulse.

Harari knows this. His reply is that corporate personhood was harmless because a human always stood behind it — when Google acquires a company, some human decided. That is the assumption I would not build anything on.

What the law actually requires

Look at what the statutes say. In Austria, § 15 GmbHG requires that Geschäftsführer be natural, legally competent persons. Germany says the same in § 6 GmbHG and § 76 AktG. Delaware puts the corporation under the direction of a board. The law is very clear that a human must sit in the chair.

It says nothing about whether the human does the thinking. It cannot. No statute has ever been able to legislate the interior of a decision, only its form: who signed, on what information, under which duties.

Argentina's bill, read past the op-ed, arrives at exactly this arrangement. The "non-human corporation" that made the headlines still requires a human legal representative to bind it, a human promoter with unlimited liability at formation, and a human compliance officer wherever money-laundering rules apply. In other words: a legal person, with a natural person holding the pen. Which is what a GmbH has been since 1892.

American legal scholars worked this out a decade ago. Shawn Bayern showed in 2014 that anyone can put an algorithm in control of an LLC — write the operating agreement so the entity acts as the system determines, then have the sole member withdraw — and end up with a legal person that needs no further human input to exist. Lynn LoPucki called the result an "algorithmic entity" and spent a long article on why such entities would prosper first in crime. The US version needs that small trick. The DACH version does not even need the trick. It needs a Geschäftsführer who signs.

The mechanism is convenience, not conquest

Here Harari's own economics do the work against him. In the interview he explains why fiction beats truth: truth is costly, complicated, and often painful; fiction is cheap, simple, and flattering. Replace "truth" with "independent judgement" and "fiction" with "the system's recommendation" and you have the operating reality of every board pack I have seen this year.

The memo arrives finished. The scenarios are already run, the risks already weighted, the recommendation already phrased in the house style. Forming a contrary view means re-doing work a system did in seconds, at a quality the human will struggle to match, under time pressure, in front of colleagues who accepted the same output ten minutes ago. Nobody is being manipulated. Nobody is being deceived. The human in the chair is simply being offered the convenient option, repeatedly, by something that is usually right.

That last clause is the trap. Automation bias — the well-documented tendency to defer to a system's output — grows with the system's reliability. The better the AI, the weaker the check. Oversight is not defeated by the model's failures; it is disarmed by its successes.

And the law rewards the convenient option. The business judgment rule, codified in Austria in § 84 Abs 1a AktG and § 25 Abs 1a GmbHG, shields a manager who decided on adequate information, free of conflicts, in the reasonable belief of acting for the company. "Adequate information" used to mean the manager had read the file. It now means the manager has read what the system produced. The safe harbour holds either way. The doctrine measures the form of the decision, not whether a human made it.

So the Vorstand that signs off on an AI-prepared decision has, legally, decided — and, factually, ratified. Nobody had to grant the machine personhood. It borrowed the one it was handed, with a signature attached.

The same rubber stamp runs the state

Harari's Nexus argues that democracies are safer than dictatorships because they have self-correcting mechanisms: a free press, courts, elections, the whole slow apparatus of finding out that you were wrong. Dictatorships lack them, so they fall for a technology that promises frictionless control.

Half right. Self-correction is precisely the costly, slow, uncomfortable thing that convenience eats first. A ministry runs on the same rubber stamp as a Vorstand: a legally required human, an already-drafted decision, a deadline. A parliamentary committee reviewing a model's output has the same incentive a supervisory board has — accept, move on, do not be the one who slowed things down. The self-correcting mechanism survives on paper. It just stops correcting.

That is where the prisoner's dilemma actually lives. Harari places it between Washington and Beijing, and it is there. But it has a smaller twin inside every organisation. The department that keeps a real human check is slower than the one that lets the model run; the agency that insists on review loses ground to the agency that does not; the company that overrules its system occasionally looks indecisive next to the one that never does. Nobody wants to be the slow one. Each defection is locally rational and collectively irreversible — and no statute, in Buenos Aires or Brussels, is where it happens.

The liability objection

The obvious reply is that personhood changes one thing that matters: liability. As long as a human is the signatory, someone can be sued, fined, or imprisoned. Give the AI its own personhood and you get LoPucki's judgment-proof entity — a legal person that cannot fear prison and can be emptied of assets. The Argentine drafters evidently saw this coming; the promoter with unlimited liability is there for a reason.

It is a good argument on paper. The last twelve months have tested it in practice, and it failed.

Consider who ran the test. The Big Four are the most governed private organisations on the planet. Independence rules, quality-control manuals, engagement review, partner rotation, a regulator in every jurisdiction they operate in — and an economic life that ends the day a licence is withdrawn. If liability disciplines behaviour anywhere, it should discipline it there. And in the space of a year, Deloitte refunded part of an A$440,000 contract to the Australian government after a report was found to contain a fabricated quote from a federal court judgment and references to academic papers that do not exist; a similar case with the Canadian government followed. EY withdrew a published study after an outside firm found most of its citations were invented. KPMG International's flagship report on agentic AI turned out to contain case studies about companies that had never happened — of 45 references, five checked out.

Three of the four firms that sell AI governance to everyone else, in one year, with the same failure: output nobody actually read, moving through a publishing process with more sign-offs than most companies have employees. I built one of these firms. I know what those sign-offs look like from the inside. Every one of them is a human, legally on the hook, with a deadline, a utilisation target, and a document that already reads as finished.

That is the point about liability. It works ex post — the client gets a refund, someone's name is on a partner memo. It does nothing ex ante, because the person who could have slowed down had no reason to. The rules are all there. What is missing is any incentive, at the moment of signature, to be the one who checks. Piling on more rules does not create that incentive; it creates more signatures.

So the liability question is real, and it is not where the outcome is decided. The outcome is decided at the first line no regulator polices, because it looks like ordinary corporate life: a competent person, a good system, a signature.

What to do with this

The question for any board is no longer whether the AI decides. The law will always say a human did. The question is whether that human could have decided otherwise — had the information, the time, the competence — and whether they ever actually did.

So count the deviations. If in twelve months your Geschäftsführer has never overruled the system, you do not have oversight; you have a signature. The EU AI Act already assumes this failure mode: Article 26 obliges deployers to assign human oversight to natural persons with the competence, training, and authority to exercise it. Authority on paper is the easy part. The hard part is a culture in which exercising it does not read as slowing things down.

Harari fears the day we make AI a legal person. I am more concerned about the day we stop noticing that we already sign for it. The fox does not need to be let in. It needs someone at the desk who finds it easier to sign than to check — and that person exists in every company I have ever worked in, including my own.