There are roughly 7,000 living languages, and linguists estimate that a significant share of them could disappear by the end of the century. AI models work very well in a few dozen, reasonably in about a hundred, and poorly or not at all in all the rest. This asymmetry is not neutral: it will weigh on what survives.
We showed in our article on the cost of French that even a major European language suffers a technical disadvantage compared with English. What happens to a language spoken by fifty thousand people?
Why models are bad in most languages
The reason is mechanical. A model learns from what is written and available. Yet digitised written output is extraordinarily concentrated: a handful of languages accounts for the vast majority of text accessible online.
A language that is mostly oral, or whose written form exists mainly on paper in non-digitised archives, is almost absent from corpora. The model cannot learn what it has not read.
There is also a technical problem we have already described: tokenisation is optimised for dominant languages. A language with rich morphology or a poorly represented script ends up fragmented in absurd ways, degrading both cost and quality.
What AI actually brings
The picture is not entirely bleak, and some uses are promising.
Accelerated transcription. Documenting an oral language used to take years of manual linguistic work. Tools now make it possible to speed up the transcription of audio archives considerably, provided an initial corpus exists to adapt them. For languages whose last speakers are elderly, time is decisive.
Translation with little data. Techniques now allow a translation system to be built with far fewer examples than before, by drawing on related languages. It is not perfect, but it opens up possibilities where there was nothing.
Tools for communities. Spell-checkers, adapted keyboards, teaching materials. These are mundane tools in a major language, and their absence is a factor in language abandonment in a small one. Making them accessible changes something very concrete for transmission.
A point is regularly raised by the communities concerned, and it deserves to be heard. Digitising a language also means making it exploitable by outside actors. Some indigenous communities oppose their corpora feeding commercial models, for reasons tied to cultural ownership and sometimes to the sacred nature of certain content. The logic of the hunt for clean data makes these corpora attractive, precisely because they are rare and human. The consent of those who speak the language should be the first question, and it is often the last.
What threatens
The main risk is not that AI destroys languages. It is that it reinforces a mechanism that already existed.
If the most useful digital tools only work well in a few languages, the incentive to switch to those languages increases. A parent who sees that their child cannot use an assistant, a spell-checker or a translator in their mother tongue draws practical conclusions.
It is the same phenomenon observed with computing in general, but amplified: where a keyboard could be adapted, the quality of a model cannot be decreed. It depends on data that does not exist.
There is also a subtler risk: mediocre but massively distributed machine translation can impoverish a living language, by imposing turns of phrase calqued on the dominant language. A language survives through use, and degraded use is not neutral.
What to take away
This technology confronts minority languages with a choice they did not ask for: stay outside and gradually become unusable in digital life, or enter the system and accept dependence on actors who have no obligation towards them.
A few public and community initiatives are building open corpora, managed by the speakers themselves, which is probably the only satisfactory path. It echoes the sovereignty argument we made about open models: when a tool becomes a condition of access, relying entirely on someone else to obtain it is a risk.
A language is not just a means of communication. It is a way of carving up the world, with distinctions that other languages do not make. Each disappearance removes a way of thinking. In that respect, the question of which languages these systems will learn is not a technical question; it is a question of what we are willing to lose.