Forty-Five Percent of a Parliament, Written by Something That Cannot Be Called to Testify

The number worth sitting with this morning is 68 out of 150. A regional outlet in the Netherlands, working with an academic detection tool, found that sixty-eight members of the Dutch lower house used artificial intelligence to help draft at least one speech or committee intervention. The reporters conceded their figure was almost certainly conservative, because only contributions scoring above fifty percent machine-generated were counted. A separate pass over twenty-eight thousand parliamentary questions earlier in the autumn, by the national broadcaster, had already flagged roughly a quarter as bearing the stylistic fingerprints of a language model.

The Dutch prime minister himself has acknowledged using the technology to compose social media posts, including reflections on the delicate historical file of the Moluccan veterans, defending the practice by noting that around three-quarters of Dutch people also use such tools. The leader of the farmers' party BBB compared the practice to using a calculator. A former member, Esmah Lahlah, explained that she had used a model to compress conversations with constituents into speaking-time constraints. A representative of the far-right PVV confirmed she used it for concision, and insisted she would never say anything in parliament she did not support. A centre-right CDA member said he used it to edit, not to generate.

Take each of those justifications seriously for a moment, because they are not frivolous. Then ask the question the AI Act is supposed to answer, and does not: what is the compliance regime for a parliamentarian who outsources the drafting of her own words?

The answer, read strictly from the text adopted in 2024, is that there is none. The Act's prohibited-use list does not touch legislative drafting. Its high-risk Annex III covers systems used by public authorities to evaluate eligibility for benefits, to assess asylum claims, to allocate emergency services. It does not cover a system used by an elected representative to generate the sentences she will later read into the record. General-purpose model providers face transparency obligations on training data and systemic-risk evaluation; those obligations run upstream, to the lab, not downstream, to the user in the chamber.

My reading is that this gap is not an oversight. It is a political choice made during trilogue, when parliamentary negotiators were unwilling to legislate against tools their own staff were already quietly adopting. The file was written by people who knew exactly what they were leaving out.

Which brings us to the enforcement architecture, and here the picture is unflattering. The AI Office in Brussels is standing up its staff. National competent authorities under Article 70 are, in several member states, still a line in a draft budget. The Dutch data protection authority — the body that would in practice field a complaint about a parliamentarian's use of a chatbot — was reported this week to have taken its own website offline alongside other Dutch public bodies during the PHP vulnerability cascade. The regulator charged with policing the digital conduct of the Dutch state could not keep its own homepage up. That is not a cheap shot; it is the operational baseline against which any AI Act enforcement must be measured. The outage, it should be said, stemmed from a flaw in a widely used programming language, not from any failure specific to the authority itself — which is rather the point about shared digital infrastructure.

The transmission to the Gulf is, I want to flag as projection rather than fact, worth watching. Several GCC consultative bodies have in recent years looked to European legislative-tech practice as a reference point. My expectation — and it is an expectation, not a reported finding — is that this week's Dutch disclosure will be read in the region as permission rather than warning: if nearly half of a northern European chamber is drafting with a model and the Union's flagship regulation has nothing to say about it, the political cover for similar adoption elsewhere has effectively been issued, free of charge, by The Hague. Vendors selling parliamentary drafting assistants into the region will, I would wager, find the citation useful. Readers should treat that as my inference from the European facts, not as reporting from the Gulf.

Two smaller points, because they matter. First, the parties that showed no detectable AI use — the Socialists, the animal rights party and Volt — are among the smaller benches. Draw from that what you will about which groupings have the staff budget to write their own speeches. Second, the detection tool itself is a private product, and its false-positive rate on non-native English and on institutional Dutch has not, as far as the published reporting indicates, been independently audited. A regulatory regime that would depend on proprietary classifiers to establish facts about parliamentary conduct is a regulatory regime with a vendor-lock problem waiting in its second act.

A parliamentarian who delegates drafting to a model has not broken any article of the AI Act. That is the problem, not the exoneration. The Union legislated the risks it could see and chose not to see this one. The next revision — and there will be one, probably sooner than the Commission's work programme currently admits — will be judged on whether it closes the gap or ratifies it.