The paper behind the essay

Where it stands

A first draft exists. It reviews how AI governance has been defined since 2017, starting from a search logged in advance and running well beyond the 38 documents in our history. Before any search was run, the protocol wrote down what the site claimed, and what evidence would show those claims were wrong.

Some of it did. The review found earlier work that had already reached parts of our definition: Domínguez Hernández and colleagues on incentives, institutions and power; UNESCO on anticipation; Aslam on the authority to correct or stop an AI system. So we narrowed what we claim, and the history now credits them. What we have not found is the combination our definition offers: the four objects, a purpose and timing held in one sentence. That is now the claim, and it is a smaller one.

The screening and coding in the draft were done by an AI research assistant, under the author’s direction. We say so here, as the paper does, because it is the reason the next step matters.

What is left before it can be submitted

1. 1A human second coderEvery screening and coding decision in the draft was made by an AI research assistant. A person needs to code a sample independently, so that agreement can be measured. Until then, the draft’s figures are not findings, and we do not publish them.
2. 2The official texts still unreadFive official texts in our record were not read at source for the review, among them the EU AI Act and the Council of Europe convention. Each could change what the review reports about official framings.
3. 3A real caseThe eleven questions have not yet been applied to a documented decision. The paper will write down its expectations first, then work through the public record, and report where the questions helped and where they did not.
4. 4Wider voicesThe search did not yet reach civil-society, labour and affected-community sources. Adding them is the obvious next step, and the gap the history already names.

What it claims, and what it does not

The paper does not claim that each idea in the definition is new. Fields of inquiry and practice, the objects that regulate behaviour, and the problem of timing all have long histories, and the paper credits them. What it proposes is the combination: one sentence that holds what governance acts on, what it is for and when to act, with each part turned into a question people can use.

Whether that amounts to a new way to understand AI governance is still to be shown. It would take real cases in which the questions reveal something other frames miss, a review a second person has confirmed, and people who use the questions and see a decision differently.

The argument in brief