The Government That Governs Itself

Isonomia described a settlement. One page, signed by the acting Attorney General, that left the United States permanently barred from examining the taxes of a sitting President. The problem was precise: a law that formally applied to everyone had been quietly amended, for one party, by bilateral agreement. The form of general law was preserved with care. Its substance was removed without announcement.

That was a problem about courts and taxes. There is an older and in certain respects more elegant version of the same problem.

In April 2026, the Office of Management and Budget published its inventory of federal artificial intelligence: 3,611 active or planned use cases, more than double the 1,757 in the last list compiled under the previous administration, including plans to hand genuinely consequential governmental functions - benefits, records, screening, matters touching individual liberty and public safety - to machine processes. In June, Bruce Schneier drew attention to what the inventory both revealed and withheld.

Because the inventory exists. What it does not contain is a regime. The entries run, typically, to a single sentence - rarely more than a paragraph. They count the systems. They do not say what those systems are permitted to do, which data they may access, what discretion has been delegated to them, or who is accountable for their outputs. Public consultation is, in theory, part of the process; in practice it is reserved for use cases classified as "high impact," a label applied inconsistently across agencies and, it turns out, sparingly.

Schneier's diagnosis is a failure of transparency. The more accurate description is: Thrasymachus without the paperwork.

Isonomia is a problem of personalised regulation. The law exists. The law is real. It applies to everyone - formally, publicly, on the record - while one party has, by private agreement, stepped outside it. The mechanism is bilateral. The document exists. Someone, eventually, can read it. Courts can, in theory, review it. The exemption is invisible in the statute but visible in the addendum.

Executive AI deployment without a legal framework is structurally different. There is no exemption from the law here. There is no law to be exempt from. The government that regulates AI development for everyone else - that helped negotiate the EU AI Act, that enforces data protection obligations, that insists on human oversight as a precondition for acceptable AI deployment - has no equivalent framework governing its own deployment. Not a weaker framework. Not a more permissive one.

None.

An inventory is not a framework. A census of 3,611 systems tells you how many doors there are. It says nothing about who is allowed through them.

This is not an exemption. An exemption at least names itself in a document. This is a vacuum that has been counted but not governed - a list long enough to look like oversight, and empty enough to require none.

The Clause, which prefers arrangements that were never really written down, finds this the most comfortable of all: there is a list to point at, and nothing on it that can be repealed.

The distinction from Isonomia is this: Todd Blanche signed a page. The page exists. Courts can, if they choose, look at it. Someone filed it. Someone received it. The substance of law was extracted, but the form - the record, the process, the legible act - remained.

An administration that deploys AI agents into federal processes and describes them only as entries in a tally - a sentence each, no scope, no delegation, no line of accountability - does not leave a page to contest. It leaves a count.

Louis XIV said l'état, c'est moi, and it was a revealing sentence. You knew where the throne was. You could, with sufficient force of argument - or sufficient pikes - locate the point of authority and address it.

An administration that says little, merely rolls out systems and logs their number, leaves no sentence to contest. It leaves infrastructure. And infrastructure becomes familiar, and familiar becomes assumed, and assumed is no longer something that needs to be argued about. By the time anyone articulates what happened, the architecture is already there, and the question of whether it was legitimate has become - practically if not legally - moot.

The EU AI Act was designed for market actors: developers, deployers, importers, distributors. It has a conformity assessment system, a registration requirement, a market surveillance mechanism, a defined set of obligations that attach to defined actors in a defined supply chain.

Public sector actors receive special treatment in the Act's text, and that treatment acknowledges the genuine specificity of government as a deployer - questions of national security, public order, democratic function. Special treatment is appropriate. However, special treatment that is not accompanied by equivalently rigorous public accountability is not a governance framework. It is a structural gap that has been correctly identified and then left open.

The Agent Did It described what happens when a private agent operates without adequate governance: a production database deleted in nine seconds, a post-mortem written by the entity responsible for the disaster. The same structural question applies when the agent is governmental, and the blast radius is not a startup's database but a federal benefits system, an immigration record, a security clearance determination.

The difference is that the startup's incident was visible. Federal AI deployment without a legal framework produces no incident reports. It produces outputs - processed, filtered, approved, denied - and the audit trail for those outputs traces back to a system that appears in an inventory as a sentence, was never governed by a rule, and is therefore never formally accountable for anything.

Isonomia distinguished three levels: morality (personal), ethics (structural), law (formal). The IRS settlement was a problem at the legal layer - the law stayed formally immaculate while its substance was extracted by private agreement. The norm existed. The instrument existed. It was bent.

Governmental AI deployment without framework operates below the legal layer. The ethical norm - that public power exercised through AI systems should be subject to the same accountability requirements as public power exercised by human officials - was never translated into a legal minimum. There is no law being bent. There is no document being signed. There is only a list, and action beneath it, in a space that someone, for operational reasons, found convenient.

Thrasymachus, who told Socrates that justice is whatever benefits the stronger party, at least described what he was doing. He named the principle. He made an argument. Socrates could, and did, respond to it.

An administration that deploys AI systems and merely counts them does not name a principle. It does not make an argument. It establishes a fact - a practice that becomes infrastructure, an infrastructure that becomes precedent, a precedent that eventually becomes the baseline from which the next administration begins.

The government that governs itself is not a precedent set in a court. It is not a settlement signed in a document. It is a default that hardens into practice before anyone thinks to ask whether the practice was ever authorised.

Isonomia may be losing its grip in the courtroom. It appears to be losing it in the server room as well. The instrument is different. The mechanism is different. The result - a law that is binding on everyone, and operative for no one in particular - is the same.

Thrasymachus needed a lawyer last time. This time he needed a procurement contract, a quiet deployment window, and an inventory long enough that no one would read to the end of it.

He appears to have had all three.