Who writes the grammar?

A less visible dependence

We debate the infrastructure of artificial intelligence: who builds it, who hosts it, who finances it. That debate is necessary; it does not cover everything we depend on, for another dependence lodges further upstream, in the very categories with which we judge these systems.

Two earlier positions of this chamber raised the question of the norm embedded in the system that answers, then that of the library that forms its judgement. This one bears on a third plane, harder to see because it resembles nothing that can be owned: the categories in which the object is judged.

An institution deploys a model, complies with it, secures it, audits it. Each of these operations mobilises concepts: risk, compliance, resilience, safety. None of these concepts is neutral. Each comes from a school, a doctrine, a jurisdiction; it carries a history, a language, a way of qualifying what counts. To observe this is description, not accusation.

The categories that judge

Risk, compliance, resilience, safety are not facts one observes, but grids through which one looks, and a grid always comes from somewhere.

Risk management has a history, a language, a place. So does AI safety. So does resilience engineering.

One adopts the tool believing one adopts only a tool. One also adopts, without having written it, the way of judging it.

Yet these categories, far from being merely descriptive, decide in advance what will count as a risk, what will amount to compliance, what will be held resilient, what will pass for safe.

They do not frame the decision. They preform it.

These categories thus organise the arbitration without having themselves been its object. The rules applicable to systems are put to deliberation; the grammar in which they are judged is not.

The incomplete answer of sovereignty

It will be objected that sovereignty has taken hold of this. That the actor is nationalised, that hosting is in Europe, that the applicable European rules are complied with.

The answer is real. But it bears on the object, the machine, the data, the supplier, the infrastructure, and not on the categories by which that object is qualified. One can own the thing and borrow the rule that judges it.

Law and engineering have organised mastery of the instrument: auditability, hosting, reversibility, traceability. Nowhere have they organised mastery of the categories that qualify it.

Owning the tool is not writing the rule that arbitrates it.

The default grammar

Where no grammar of arbitration is produced, arbitration takes place all the same; it takes place by default, in the grammar of whoever wrote it first.

That grammar does not always impose itself as a constraint. It settles in as a convenience, and nothing is harder to take back than what renders a service: one adopts it, and little by little one ceases to know that one could have thought otherwise.

An organisation can then be compliant, national, sovereign on paper, and still reason in a language it did not choose.

The categories in which we judge artificial intelligence constitute, in that respect, an invisible dependence: not because it is weak, but because it takes the very form of the obvious.

Closing

Mastering where the machine runs is a question of infrastructure.

Mastering what it has read is a question of corpus.

Mastering the categories in which it is judged is a question of sovereignty.

And that one has been settled nowhere.

Whoever does not write the grammar of arbitration is subject to its arbitration.

Paris, September 2026

Related positions · Where does the norm come from? · What have they read?