AI language models learn from data. The data on which they are trained shapes not only their capabilities but also their entire sense of what is normal, important, and worth expressing. Language use on the internet is today heavily concentrated. English alone accounts for roughly half of all websites whose content language can be identified (W3Techs 2026).

The research is unambiguous. Low-resource and marginalised languages receive weaker datasets, poorer benchmarks, and significantly worse model performance than high-resource languages (Grützner-Zahn & Rehm 2024; OECD 2023; Zhong et al. 2024; Alam et al. 2024; Micallef et al. 2025; Nuha et al. 2026). The consequences of this are not limited to specialists and enthusiasts, but affect each and every one of us.

As children, teachers, and ordinary users increasingly turn to AI assistants to explore literature, history, and cultural identity, the heritage they encounter will be the heritage of the majority — the heritage that is well represented in training data. The rest risks becoming harder to find, harder to generate, harder to search — and therefore easier to overlook. This is especially true as AI models increasingly displace traditional search engines such as Google Search.

Languages do not disappear only when people stop speaking them. They disappear when they cease to be useful in the institutions and tools around which our everyday lives and routines are organised. Search, translation, writing assistance, education, administration, and entertainment are increasingly mediated by AI. If these systems do not support a given language, users come to experience that language as unnecessary, deficient, inadequate, and perhaps even old-fashioned.

Yet this does not happen all at once. The experiences accumulate interaction by interaction — when a Sorbian writer finds that language models cannot reliably support her work, or when a Sámi musician realises that his recordings appear in no AI training corpus, or when an Aromanian songwriter cannot get translation tools to handle his lyrics convincingly. These are not edge cases to be tidied up later, but part of a much larger problem for which there is no solution in a market-driven world.

The Parliament’s resolution explicitly links copyright, creativity, and Europe’s cultural diversity (EP 2026a). The Council of Europe’s language framework exists precisely because regional and minority languages require active support to remain living parts of European cultural identity (Council of Europe 1992/1998). And yet a market-centred system for AI licensing and opt-outs will, by its nature, favour those who already have scale — which means it will favour dominant languages, large rightsholders, and well-resourced collecting societies.

The market will not solve this problem. Cultural policy recognised decades ago that markets alone cannot preserve linguistic diversity, because demand for smaller languages is limited. That is precisely why active public support mechanisms exist. The copyright debate around AI has not yet woken up to this. It should.