The Writ and the Machine: Cognitive Sovereignty in the Age of AI

A woman applies for housing. Six weeks later a letter arrives saying the application was unsuccessful. It gives no reason, because no reason is required. What the letter does not say is that a model scored her, that the score turned on a pattern in her rental history that a human reviewer would have read differently, and that a person with the authority to reverse the outcome never saw the file.

She has not been wronged in any way the law currently recognizes. She has simply never been told that anything happened.

This is the ordinary shape of the problem, and it is worth noticing how old the problem is and how new the shape.

What every civilization built

Societies that have lasted tend to share one feature. Each constructed a moment at which power had to stop and answer to a particular person.

Athens required every departing official to submit to the euthyna, an accounting before citizens who could question the record. Rome developed procedures by which a magistrate’s act could be challenged by the individual it fell upon. England produced the writ of habeas corpus in 1679, which is often described as a protection against unlawful imprisonment but is more precisely a demand for an explanation: produce the body, and state the cause. The Islamic world established mazalim courts where a subject could bring a grievance against the administration itself. Chinese courtrooms hung the phrase 明镜高悬, the bright mirror suspended high, above the magistrate’s bench, a standing reminder that judgment was supposed to reflect the matter truly rather than the interests of the powerful.

These institutions differ enormously in politics, in theology, in what they assumed about human nature. They converge on one thing. Each identified a moment when authority is applied to a person, and each insisted that at that moment the person is owed something: a reason, an accounting, a face.

Call this the encounter. It is the smallest unit of civilizational dignity we have ever managed to build, and nearly every tradition arrived at some version of it independently.

What inference removed

Artificial intelligence has not attacked the encounter. It has done something more difficult to see. It has removed the moment at which the encounter would occur.

When a model produces a determination about a person, there is no magistrate, no bench, no departing official to question. The decision may be assembled from a score generated in one place, an eligibility rule applied in another, and a policy set by a vendor in a third. Each participant can say honestly that they made no decision. The system as a whole made one. The person it was made about was never present, and in most cases never learns that a system was involved at all.

This is not malice. It is architecture. And it is why the ordinary responses have not worked. We have asked whether models are accurate, whether they are biased, whether they are safe. These are good questions about systems. They are not the question the euthyna asked, or the writ, or the mirror. That question was always about a person: at the moment this fell on you, what were you owed?

The convergence, and the direction

Here is the part that surprises most audiences.

Both halves of the world have now built substantial machinery for making artificial intelligence transparent. In China, providers of consequential services complete security assessments and file their algorithms with the state, and a national standard specifies how synthetic content must be marked so that its provenance can be traced. In the West, we have produced model cards, system cards, evaluation reports, and an expanding literature of technical documentation.

The two systems are usually described as opposites. On this question they are twins. Both point the disclosure upward. China’s filings travel to the regulator. Our documentation travels to researchers, auditors, and procurement officers. In both cases the information about how a person will be judged reaches everyone except the person.

I do not think this happened for ideological reasons. Transparency was built by institutions for institutions because institutions were the parties who knew to ask. But the effect is a world in which a great deal is known about these systems and almost none of it reaches the human being standing in front of one.

That gap is what I have come to call cognitive sovereignty: the capacity of a person to know that they are being assessed, by what, toward what end, and to contest the result. It is not a right to refuse technology. It is the old right of the encounter, restated for a moment that no longer has a room.

A small law with a large idea

On August 4, Representative Erika Uyterhoeven filed a bill in the Massachusetts legislature called the Digital Commonwealth Act. It is a state measure in one American commonwealth, and it may take years to move. I raise it because of what it attempts rather than what it is likely to achieve.

The bill relocates the entire question. Its obligations attach at inference, meaning the moment a deployed system is applied to a real person. Not at training. Not at market entry. Its central sentence is almost plain enough to miss: the right begins at inference, and no claim about training data, design, or authorship diminishes it.

From that single move, the rest follows. A person subject to a consequential determination is entitled to know that a system was used, what it was for, what it inferred about them, whether their information was retained or used to improve the model, and how to reach a human being with the authority to look again. The bill states that a disclosure the average person cannot understand does not count as a disclosure, and that one they had no reasonable occasion to encounter was never made. It requires nothing about how models are built, and it grants no power to prohibit anything. Its own phrase for this is a right to visibility, not a right to veto.

It also does something unusual with scale. It gives municipalities the authority to decide for themselves whether they adopt these systems in their own schools and departments, and on what terms. A town may require that AI be visible. A town may not require that AI be absent. The reasoning is modest and, I think, correct: nobody knows what these systems will look like in five years, and a community that cannot choose cannot adapt.

Whether this particular bill passes matters less than whether the idea travels. Its findings say that when data generated by a community is used to improve the systems deployed upon it, the value taken from that community should be visible to the community itself. That is not an American thought. It is the euthyna, applied to a machine.

Why this is the achievable thing

There is a version of the AI governance conversation that has become paralyzed, because it asks whether these systems should exist and whether anyone can control what they become. Those questions are real and they are hard.

The encounter is not hard. It is a solved problem, solved repeatedly, by societies with far less capacity than ours. We know how to build a moment where power stops and explains itself, because we have been building it for two and a half thousand years in Athens, in Baghdad, in London, in Beijing. What has changed is only that the moment now happens inside a system rather than across a bench.

Europe has demonstrated something useful here without meaning to. When the European Union postponed the heaviest parts of its AI Act this summer, the obligations that slipped were those requiring elaborate machinery to be constructed first. The duties owed to a person at the point of use held their dates. The layer closest to the human being turned out to be the layer that could actually be delivered.

That is the encouraging finding buried in a delay. The most humane thing we can do is also the most practical.

I would not describe any of this as regulating artificial intelligence. What is being proposed is far older and far more modest. A person should be able to know when a decision is being made about their life, and by what, and should be able to ask someone to look again.

Every civilization we admire figured this out. We are simply being asked to do it once more, for a kind of authority that does not sit behind a bench.

Russ Jude Wilcox writes on artificial intelligence, institutions, and the future of AI and humanity. He is the founder of ArtifexAI, Policy Chairman of the American Society for AI, and AI Council Chair at United World Leaders.