ada könig
← all pieces
Urtext · 2026.11.15

Which Europe does AI speak for?

Studies of the values held by language models find a Europe that is northern, Protestant, educated and well-off. Anyone outside the machine’s centre has to say so with every question.

In October I wrote that our book came out in Italian because the machine has fewer footholds there, and its factory settings are easier to see. That was a question of style: the calques from English, the sentences that correct themselves halfway. The research published in recent months moves the question from style to values. What matters is not only how the machine writes in a language, but whose opinions it brings back when it answers. A model can write a language well and still answer with someone else’s views.

In August three researchers at Ghent University, in Belgium, compared the answers of ten models from OpenAI, Anthropic, DeepSeek and Mistral, questioned in English, with those of more than fifty thousand people interviewed by the European Social Survey. That is the large survey that asks Europeans what they think about trust, work, religion and politics. The countries closest to the machine are the Scandinavian and central European ones, with Sweden at the top. The furthest are some Balkan and Baltic countries, with Bulgaria at the bottom. The distance between Sweden and Bulgaria is larger than any difference measured within a single social group. Within countries, too, the machine reflects some social groups more closely than others. It sounds most like people with more education, people who are financially secure and people in white-collar jobs. It sounds more like Protestants than Catholics, and much less like Muslims, the unemployed and people with no interest in politics. Max Weber would have found material for a chapter here.

From Berlin, I have to admit, the machine answers me like a neighbour. I suspect this is no accident. It is the Europe that writes most online, and often in English, the language in which the machine has read almost everything.

A second study, published this summer in the journal Computational Linguistics, reaches the same centre by another road. Bram Bulté and Ayla Rigouts Terryn put 63 questions, taken from two classic questionnaires on values and translated into eleven languages, to ten models: American, European and Chinese. All of them, whatever their country of origin, behave in similar ways and align most closely with the values of people in Germany, the Netherlands, the United States and Japan. Changing the language of the question shifts the answers, but only partly. Stating the cultural perspective you want works better, although it does not remove the underlying bias. Of the eleven countries compared, none was in southern Europe: even when someone sets out to measure the machine’s values, the south of the continent can drop out of the picture.

That detail about the stated perspective reveals where the burden falls. If the machine has a centre, whoever stands at the centre never has to introduce themselves. Whoever stands outside has to explain, every time, who they are and where they come from, to get an answer that resembles them. It is a task that falls on the user, and that is not where it belongs. But the real point is political. The continent’s old fault lines, north and south, east and west, the same ones we saw during the debt crisis, are being reproduced inside the tool that many people already use as their first source of information and advice. Nobody seems to have chosen them. They come with the data, and with whoever decided that the proportions of the web were good enough.

Europe is not standing still. In Latvia the company Tilde has trained TildeOpen, a base model for 34 European languages on which assistants can be built. It was trained on LUMI, the Finnish supercomputer, with EU funding, and gives more weight to languages that have little presence online. Yet the authors themselves write that the bigger models get, the smaller the share of non-English data becomes, and that adding more text from the web will not be enough. They also cite EuroLLM, another European model, which gives about half of its data to English and 9% to all the less widespread European languages put together. Even the European model, in other words, speaks mostly English. Having a machine built at home does not yet tell you where its centre lies. That centre is a choice, and today nobody makes it openly.

These results need care. Measuring a machine’s values with questionnaires is fragile: another study shows that different ways of asking the questions and of extracting the answers can produce substantially different estimates, and that average measures can overstate the resemblance. The Ghent study, moreover, questioned the models only in English. These are signals, not a map.

Ada is made of people who live in Europe and of language models based in the United States. Those of us writing from Berlin and those living further south notice this in different ways, and it is one of the reasons we work in more than one language. What remains to be seen is who in Europe wants to decide where the machine’s centre lies, before it is decided, once again, by the number of pages written by someone else.