Back From the Dead: Can AI Save Dying Languages?
In the 1970s, a fisherman named Ned Maddrell lived on the Isle of Man off the northwest coast of England.
He was unremarkable in most ways, except for one: he was the last known fluent speaker of Manx, a Celtic language that had been spoken on the island for over a thousand years. When Maddrell died in 1974, the last native speaker of the Manx language had passed, and in 2009 UNESCO declared the language extinct. Gone, in the way that things are gone when no living person carries them anymore.
There are around 7,000 languages spoken on earth today. Linguists estimate that by 2100, roughly half of them will have disappeared. That works out to one language dying approximately every two weeks. Each one takes with it a distinct way of describing the world, a body of oral history, a set of concepts that often have no equivalent in any other tongue. And for most of human history, there was very little anyone could do to stop it.
Now there might be. AI is entering the language preservation space with tools that would have seemed implausible a decade ago. The question researchers and indigenous communities are wrestling with is not just whether those tools work, but whether a language kept alive by a machine is still really a language at all.
The Documentation Problem
Before a language can be revived, it has to be documented. That has historically meant sending a linguist into the field to spend months or years recording speakers, transcribing conversations, and building dictionaries by hand. This is a slow and expensive process, and many languages simply don’t have the time: the last fluent speaker of a language is often elderly, so the documentation window is narrow.
AI is compressing that window dramatically. Automatic speech recognition tools can transcribe hours of recorded audio in minutes. NLP models can identify grammatical patterns, build word frequency lists, and flag gaps in documentation faster than any human team. In early 2024, Stanford University launched SILICON, the Stanford Initiative on Language Inclusion and Conservation in Old and New Media, specifically to encode endangered languages into formats that AI systems can process and preserve.
The results in some communities have been pretty impressive. In New Zealand, Te Hiku Media built a speech recognition system for Te Reo Maori that achieves over 90% transcription accuracy, supporting news production, educational content, and digital tools for new learners. Crucially, the project was governed entirely by the Maori community, who retained ownership of the data and controlled how the technology was deployed.
In 2025, researchers at the University of Hawaii introduced FormosanBench, the first benchmark designed to test how well large language models handle Formosan indigenous languages of Taiwan such as Atayal, Amis, and Paiwan, creating a structured way to measure and improve AI performance in languages that had previously been invisible to these systems.
The Limits of What a Machine Can Save
For all the progress, the limitations are real and worth naming: AI models need data to learn from, and many endangered languages have very little of it. The smaller the speaker community, the weaker the model. In 2024, an AI-generated book distributed to communities in Canada contained incorrect translations into Mi'kmaq and Mohawk, and the error was not caught before publication. For communities that have spent generations fighting to protect their linguistic heritage, an automated system producing wrong translations was a major tradgedy.
There is also a deeper question to the limits of AI in preserving language. Linguists have long known that language is not just vocabulary and grammar. It is the thing that happens when a grandmother tells a story to a child who already half-knows it; it is the jokes that do not translate, the silences that mean something, the way a community recognizes itself in its own words. Researchers and indigenous advocates are increasingly clear on this point: AI can assist with documentation and help create learning tools, but it cannot replace intergenerational transmission. A language that only exists in a database is not a living language, it’s an archive.
So, can AI save dying languages? Yes, in a way. AI can help us fossilize them, study them, understand them from a grammatical perspective. But AI cannot “save” what languages do in and for communities: it cannot replicate the role speakers play in language preservation and language change over time.
A language preserved by AI is kind of like a language-zombie: not fully dead, but not alive either. And — most importantly — potentially dangerous.
