Algorithmic bias and cultural imprinting

Language models are not neutral information systems. They reflect the perspectives, priorities and blind spots of their training data. Because that data comes mostly from English-language sources, the English-speaking world is structurally overrepresented in almost all mainstream models, not by intent but as a consequence of the data.

This has visible effects. Users notice that the same AI answers more fluently, in more detail and with more context in English than in German, Spanish or French. Step by step they draw a conclusion: what works better in English comes from a world that is further ahead.

Content matters too. Ask ChatGPT and DeepSeek the same politically sensitive question, for instance about the status of Taiwan or about the Tiananmen massacre, and you receive systematically different answers. Both are competent and polite, and both transmit a specific worldview. People who use one system daily get used to its framing: which countries count as modern, which conflicts as legitimate, which values as universal.

This is not a reproach against the manufacturers. It is a structural property of all large language models trained on a real world of unequal distribution. The question is not whether the bias exists. It is whether people are given the tools to recognise it.

For millions of workers outside the English-speaking world, this bias also decides who benefits from AI at work. Training in their own language, with AI used critically, is how we close that gap.