What have
they read?

A question never asked

We have never debated what artificial intelligence models have read.

Which archives were opened to them and which remained closed, in which languages, in what proportion, which traditions of thought weighed and which were treated as incidental. These choices were made upstream, by teams, according to criteria that were never submitted to public discussion. This is not an accusation. It is a description.

An earlier position of this chamber raised the question of the norm embedded in the system that answers. This one concerns what precedes it.

What the law has organised

One will object that the law has taken hold of the matter. The objection is serious, and it deserves to be followed through.

Regulation (EU) 2024/1689 requires providers of general-purpose AI models to make publicly available a sufficiently detailed summary of the content used for training, following a template provided by the AI Office. The template is mandatory. It requires the names of the crawlers used, the collection period, a description of the content harvested and the list of the most represented domains.

The obligation is real. One must, however, read what it is.

It makes public a summary intended, among other purposes, to enable parties with a legitimate interest to exercise their rights. The degree of detail required was calibrated source by source, in order to preserve the provider's trade secrets and confidential business information. The summary is due at the latest when the model is placed on the market, and, for models already on the market before 2 August 2025, it is not required until 2 August 2027.

The law has therefore organised the description of the library, once it is constituted. It has nowhere organised its deliberation.

A publication is not a debate.

The response of states

States noticed late. They responded with sovereignty.

Two years ago, some forty public sovereign AI projects were counted across some thirty countries. Today there are more than one hundred and thirty, in more than fifty countries. Nearly six in ten concern computing infrastructure.

The word sovereign there designates, first of all, a building.

This is not a reproach. A computing centre can be built, financed, located. It falls within what a public authority knows how to do, and the rest does not. Sovereignty went where it could go.

The floors that were never discussed

Take the most favourable scenario. A model designed in Europe, trained in Europe, hosted in Europe, subject to the applicable European rules. It remains dependent on floors that no one has ever discussed.

It will have learned the world from the web archives that Common Crawl, an American non-profit foundation, has been collecting since 2008 and distributes free of charge, the bedrock of many open pre-training corpora.

It will run on chips of which a single supplier, NVIDIA, equips fifty-two per cent of the sovereign infrastructure projects on record.

And if the operator hosting it falls under American jurisdiction, the 2018 law known as the CLOUD Act allows the production of data under its control to be compelled, wherever it is stored.

At none of these floors was sovereignty ceded. There was no treaty, no concession, no renunciation. What settled there did not impose itself as a constraint: it offered itself as a convenience. That is why it is so difficult to take back. One resists what constrains. There is nothing to resist when a thing gives itself.

No instituted procedure today allows an antecedent to be deliberated. Our assemblies decide what will be done. They do not decide what has already been learned.

Closing

A model has no opinions. It has a library.

Knowing where the machine runs is a question of infrastructure. Knowing what it has read, and who chose it, is a political question. It has been settled nowhere.

A sovereignty that does not bear on what forms judgement bears only on the walls.