AI companies like the phrase “democratic AI”. Here is what it looks like in practice.
Brussels wrote rules for its own market. Anthropic now weaves an invisible mark into text produced by Claude, worldwide, including text a person wrote themselves and only asked the model to translate or tidy up. Nobody in Nairobi, Jakarta or São Paulo had any say in it.
The mark is not the problem. Knowing whether a model touched a text is worth something, though it answers a narrower question than people will ask of it: it cannot say whether the model supplied the argument or only the English.
Nobody had to lobby for this. A company serving one strict market and several unregulated ones would otherwise build and maintain two pipelines, and one rule applied everywhere is cheaper than two. So the strictest regulator sets the default for everybody, and it does so through a company’s cost arithmetic. Nothing in that arithmetic has to consider the people at the far end of it.
The honest case for it is real. A country too small to legislate the behaviour of a model gets protections it could never have compelled, and a world where every tool behaves differently by jurisdiction would be unaffordable for exactly those countries. The objection is not to the rule. It is that the rule’s reach and its accountability now end in different places: it binds where it was passed, and the artifact carries it everywhere else, with no register, no reasons and nobody answerable outside Europe.
I have started moving a few projects to another model, hesitantly, because it does not mark my text this way.
Readers of the original version of this argument named the pattern faster than I did. Paternalism was one suggestion. Autocratic governance was the other. Either way the practical demand is smaller than representation, which nobody can offer a global public: it is a route to contest the classification. It is the same thing Indonesia’s welfare applicants gained when their appeals channel opened, for a different file.
Assorted links
-
Vietnam did not ban foreign AI from government, it published a scoring sheet. The criteria mark a platform out of 100 and admit it at 70, with at least half of the compliance block, and two domestic systems are currently authorised for state legal work (ministry criteria and Letter 4920 of 7 July, with the enforcement directive reported here). Here the rules were written for the buyer’s own market, in public, by the government that will apply them.
-
Urs Gasser, Viktor Mayer-Schönberger and Fabienne Marco argue that AI oversight built on benchmarks decays as it works: systems learn what evaluators test for and get better at the test rather than at the job. Their proposal includes sunset clauses that void a benchmark unless somebody revalidates it, which is an unusual thing for a governance paper to ask of its own instruments.
-
Eight people, among them municipal staff and ombudsmen, mapped what a Dutch municipality’s welfare eligibility algorithm actually does and found hazards missing from its public register, including no route to contest a wrong decision. The register records that the system exists, not what it can do to the person it decides about.