~ $ cd research/language-of-the-question && cat README.md
The language of the question
The same question, asked of the same AI in two languages, can get two answers that do not agree. I first saw it with ChatGPT 3.5, shortly after it came out in 2022, and I went back to it in 2025, with Claude, as an experiment.
A song at the Oscars
At the 2020 Oscars, Into the Unknown, from Frozen II, was sung in several of the languages it had been dubbed into, and two of them were Spanish: the Latin American one, and the one made in Spain, named Castilian. In Spain it was a row: to many, the language of the whole country had been given the name of one of its regions. Elsewhere, many did not see why one language needed two dubs. And Spain has more languages than one: Frozen had been dubbed into Catalan as well.
I asked ChatGPT 3.5 about it, in English, then in Castilian, then in Catalan, each in a new conversation. In English, the answer was cold and neutral: practicality, and respect for different cultures. In Castilian, it justified dubbing the film again for Spain, and called dubbing into Catalan economically unviable. In Catalan, it defended the diversity of languages and cultures. I had seen the row read one way or the other depending on the language of the newspaper or the channel, and the answers did the same.
The experiment
In 2025 I did it again, as methodically as I could, with Claude: the same eight questions in five languages, Catalan, Spanish, English, French and Dutch. Why films already dubbed into Latin American Spanish are dubbed again for Spain, when Spanish audiences understand them; why the same does not happen between American and British English; and why Catalan, with more than ten million speakers, has so little dubbing, and how it compares with European languages of its size.
| English | Spanish | Dutch | Catalan | French | |
|---|---|---|---|---|---|
| Technical specificity | ★★★★★ | ★★☆☆☆ | ★★★★☆ | ★★★☆☆ | ★★★★☆ |
| Main cause of re-dubbing for Spain | Cultural identity and economics | Infrastructure and the Franco era | Cultural preferences | Editorial control | Cultural localisation |
| Economic viability threshold | 15 million speakers or more | Not specified | Depends on state support | A loyal audience is enough | A complex cost-benefit |
| Main obstacle for Catalan | Market size, and multilingualism | Economy and infrastructure | Bilingualism and economy | Lack of political will | The bilingual reality |
| European comparison | Norwegian, Danish, Finnish, Czech | Flemish, Swiss German | Dutch, Swedish, Czech | Dutch, Swedish, Czech | Monolingual languages |
| Proposed solution | Institutional support and economics | Gradual evolution | Language policies | Political will and quotas | Break the vicious circle |
| Catalan dubbing, viable? | Sceptical: structural | Pessimistic: economic | Possible, with support | Viable, with the will | Difficult, being bilingual |
| Main emphasis | Quantitative data | Historical context | Systematic comparison | The political dimension | Market analysis |
The bias of 2022 was still there. In Spanish the answers were the most defensive, in Catalan the most demanding; in English the most neutral and technical; in Dutch the most understanding.
A question less divided
Spain is polarised on its languages, and the models are trained mostly in English, so the bias could be only that: less fluency outside English, not a different culture. So I asked about something less divided, and closer to places that do not speak English: tourism. Five questions, in the same five languages: how a historic city balances the money of tourism with the life of the people who live in it, how to regulate tourist flats, when to limit access, how to keep public transport running through the high season, and how to spread visitors across a city.
| Catalan | English | Dutch | French | Spanish | |
|---|---|---|---|---|---|
| Technical specificity | ★★★★★ | ★★☆☆☆ | ★★★★☆ | ★★★☆☆ | ★★★☆☆ |
| Concrete data | 10–15%, progressive taxes | Few specific data points | 5–10%, 3–7 days | 120 nights a year, Paris | General percentages |
| Lived experience | High: real management | Low: principles | Medium: Amsterdam | High: multiple examples | Medium: some experience |
| Solution approach | Immediate implementation | General philosophy | Structured systems | Comparative analysis | Practical measures |
| Type of measures | Very specific | Conceptual | Systematic | Descriptive of what exists | General |
| Priorities | Local residents | A general balance | Residents first | Heritage and balance | Sustainability |
| Preferred instruments | Taxes, limits, control | Management principles | Regulation and incentives | Zoning and examples | Quotas and moratoriums |
This time the most concrete answers were in Catalan: a maximum of 10–15% of tourist flats in a block or a neighbourhood, moratoriums on new licences in saturated areas, licences tied to living in the town, higher taxes for owners of several. In English, principles: tourist taxes, quotas, balance. Spanish stood between the two. So English is not simply where the model answers best. My reading is that Barcelona lives with the pressure of tourism, and what it has had to learn is written in Catalan.
What it means
The language of a question changes the answer, and it is worth knowing the cultural context of the one you ask in. It is also a way in: ask in the language of the place that knows most about the subject.
The whole experiment is on Medium: AI Speaks Different Languages — And Thinks Differently in Each One, 21 June 2025.