An Indigenous Language Dies Every Two Weeks—and AI Is Beginning to Save Them

An indigenous language disappears every two weeks—and AI is beginning to save them

Sometime between now and a few weeks from now, a human language will die. Not a minor or marginal language—an entire language, with its grammar, its metaphors, its ways of dividing time and space, its ways of naming plants, winds, relationships between people. According to UNESCO, 40% of the 7,000 languages catalogued in the world today are threatened with extinction. According to the most recent data (ELCat, Nature 2022), a language disappears on average every 1 to 3 months—the figure “every two weeks” often cited by UNESCO is today contested by the most recent scientific research. At the current rate, half of all living languages will have disappeared by the end of the century.

This figure could have remained a statistic in an institutional report. It is becoming the starting point for an unexpected technological bet. The same artificial intelligence models that were long analyzed as forces of cultural homogenization—absorbing English-language content, dominating digital usage, further marginalizing resource-poor languages—are beginning to be turned against this trend. Since 2024, lightweight language models have made it possible to train useful tools on tiny corpora, making linguistic revitalization technically accessible for languages with only a few thousand speakers. The question is no longer merely technical. It is political: who controls these corpora, who decides what to do with them, and to whom does a language belong when the people who speak it have no state to protect it?

The essentials

• According to recent data (ELCat, Nature 2022), a language disappears every 1 to 3 months; 40% of the world’s 7,000 languages are threatened according to UNESCO.
• Small language models (SLMs) can now be trained on corpora of just a few thousand words, making documentation and revitalization tools accessible where large models failed.
• Concrete projects are already active: Māori in New Zealand, American tribal languages in the United States, Welsh in the United Kingdom, Quechua in South America.
• The main bottleneck is no longer technical: it is political and legal. The question of ownership of linguistic data threatens to block or derail the most promising projects.
• The 2026-2030 horizon is decisive: communities that do not begin digital documentation of their languages now will have corpora too thin for future tools to be effective.

Why a dead language is more than just a cultural loss

There is a temptation to treat language extinction as symbolic loss, moving but abstract. This temptation is an analytical error.

A language encodes a way of knowing the world. Research in ethnobiology has documented this repeatedly: indigenous languages contain botanical taxonomies, ecological knowledge, meteorological classifications that have no equivalent in major world languages. When a language dies, this knowledge often disappears with it—not because it couldn’t be translated, but because no one had the time, means, or mandate to transcribe it. A study published in PNAS in June 2021 shows that 75% of the 12,495 medicinal plant knowledge instances studied in three regions (North America, northwestern Amazonia, and New Guinea) are linguistically unique, that is, known from only a single language—so much knowledge that would vanish with it without documented equivalent.

There is also a direct political dimension. Languages are vectors of rights. Indigenous peoples’ territorial claims, their legal traditions, their settlement narratives—all of this passes through language. In Latin America, several lawsuits over land rights have stumbled on the absence of documents in indigenous languages, which weakened the force of oral testimony. A living language is a legal resource. A dead language, an unusable archive.

Finally, there is the demographic aspect of the problem. Most threatened languages have fewer than 1,000 active speakers. Many have only dozens of elderly speakers left. Most linguistic knowledge is contained in human brains that will no longer be here in ten or twenty years. The documentation window is real, measurable, and closing.

What large models could not do

For a long time, artificial intelligence linguistics was structurally useless for endangered languages. Large language models—GPT, Llama, Mistral—are trained on massive corpora. For English, we’re talking about hundreds of billions of tokens. For French or German, tens of billions. For Māori, Guarani, or Navajo, digital corpora are measured in millions of words, sometimes in hundreds of thousands. This is insufficient to train a general model. Attempts to include resource-poor languages in large models often produced mediocre results—grammatical hallucinations, confusion with neighboring languages, translations so calqued on English as to be culturally false.

This bottleneck has partially loosened between 2023 and 2025. A new generation of small language models—called SLMs, for small language models—has demonstrated an ability to produce useful tools on much more modest corpora. The 2025 Brookings report on this question documents several cases where spell-checkers, text predictors, and documentation tools were trained on corpora of just a few thousand sentences, with results good enough to be usable by native speakers. This is not professional-quality machine translation. It is functional writing assistance, teaching support, and oral transcription help—three activities that are precisely those of communities trying to transmit a language.

The distinction is important. The goal is not to create a ChatGPT in Navajo. It is to provide a teacher instructing in Navajo with a working spell-checker, a transcriber recording elders with a tool to help normalize spelling, a community drafting its first administrative documents with grammar consistency support. These modest tools have a direct impact on transmission. And they are, for the first time, within technical reach.

Projects that are already working

New Zealand is one of the most advanced laboratories. The Māori language revitalization movement—called kōhanga reo, the “language nests”—began in the 1980s, well before AI. It built immersion schools, media in Māori language, strengthened institutional presence. This foundation made it possible to build a digital corpus substantial enough for AI language tools to start being useful. Te Hiku Media, a Māori organization, developed a speech recognition engine in Māori—the first of its kind for this language. This project was not entrusted to Google or an American university. It was completed internally, with strict community control over collected data. This model—technology plus indigenous governance—is what Brookings experts cite as a reference.

In the United States, several tribes have undertaken similar projects, often in partnership with universities. The University of Hawaii worked on Hawaiian, whose situation was alarming in the 1980s: in 1985, only 32 children under 18 still spoke the language, according to the UH Foundation, even though between 1,000 and 2,000 adult speakers remained. Today, the language has more than 20,000 speakers, following a school immersion program. AI arrived after social revitalization—it amplifies a movement that existed, it does not create it. In South America, projects around Quechua and Guarani have started, with variable results: Quechua, spoken by more than 8 million people, offers a larger corpus; Guarani, co-official in Paraguay with Spanish, benefits from rare institutional status for an indigenous language.

In Europe, Welsh is often cited as an example of what political will can accomplish. With 900,000 speakers and co-official status in Wales, it has a significant digital corpus and public funding for the development of linguistic tools. The Welsh Parliament mandated the development of AI tools in Welsh for its administrative services. This case illustrates a simple truth: when a public institution decides a language matters, tools follow.

These projects have in common a particular architecture: a sponsoring community, even a modest corpus, a controlled external technical partnership, and local governance over data. Remove any one of these elements and the project collapses or devours itself. It is precisely this last point—governance—that is today the main battleground.

Linguistic data ownership: the new colonial frontier

The linguistic corpus of an indigenous language is a rare, non-reproducible, and potentially valuable resource. This reality attracts actors whose interests do not necessarily align with those of the communities concerned.

Universities have collected linguistic data from indigenous communities for decades, sometimes with consent, often without. These archives—audio recordings, transcriptions, dictionaries—are stored on university servers, subject to the law of the country where the university is located, and inaccessible to the communities that are their source. When a researcher or company today wants to train a model on this data, the community generally has no say.

The problem goes further. Several tech companies have begun to scrape content in resource-poor languages—religious texts, Bible translations, administrative documents—to enrich commercial multilingual models. These operations happen without notification or compensation to speakers. From the perspective of current law, it is legal. From the perspective of the communities concerned, it is extraction. Here, in a digital register, we find a well-documented mechanism in other domains: the capture of a resource produced by a community for the benefit of external actors who control its valorization. The comparison with natural resources is not superficial—it illuminates a structural dynamic that several indigenous legal scholars are beginning to formalize under the term “data colonialism.”

The emerging response passes through new legal frameworks. The concept of “indigenous data sovereignty”—promoted notably by the FNIGC network in Canada and by researchers associated with CARE principles (Collective Authority, Responsibility, Ethics)—posits that communities must have a right of control over the collection, storage, use, and sharing of data concerning them, including their linguistic data. These principles remain largely voluntary for now. A few specific agreements have been signed—Te Hiku Media notably refused to cede its data to commercial enterprises and built its own models—but the general legal framework remains absent. This is a gap that the acceleration of commercial AI uses makes increasingly urgent to fill.

What AI can do, and what it cannot

We must be precise about what technology can contribute here, and what it will never replace.

AI tools can document faster. They can help transcribe hours of oral recordings into normalized text, reducing work that took years to a few months. They can create adapted spell-checkers and keyboards that make a real difference for young speakers writing on their phones. They can generate pedagogical resources—exercises, reading texts, questionnaires—in languages for which no publishing house will ever invest. These are real, measurable contributions, relatively inexpensive to deploy once the base corpus exists.

What AI cannot do: decide that a language deserves to be saved. That decision belongs to the community, and it is not trivial. Several communities have deliberately chosen not to document certain aspects of their language—rituals, sacred terms, knowledge reserved for certain members. Technological enthusiasm, if not held in tension with these choices, can produce violations that nothing will repair. AI also cannot substitute for living transmission. A language does not survive in a language model. It survives in families that use it, schools that teach it, media that bring it to life. Tools are amplifiers of a human movement—they do not trigger that movement.

This is why successful projects always combine two distinct logics. On one side, a social and political logic: educational choices, public funding, institutional decisions. On the other, a technical logic: tools that make transmission easier for those who have decided to do it. Separating these two logics is the main error of failing programs. A tool without movement is a curiosity. A movement without a tool can suffice—but not as well.

We find here a familiar pattern in other domains where technology and politics intersect: technology lowers costs, expands possibilities, but does not determine choices. The same debate runs through cultural production: what distinguishes cultural industries that survive from those that fade is rarely the available technology—it is the political decision to support them.

Who decides, and with what

The Gordian knot remains governance. Three models coexist today, with different logics and results.

The first is the classical academic model: a team of linguists collects, documents, publishes, and deposits data in a public or semi-public archive. This model has produced most existing resources for endangered languages. It has also produced the sharpest ownership conflicts, because data is held by institutions that do not answer to source communities.

The second is the sovereign community model, illustrated by Te Hiku Media in New Zealand or certain tribal initiatives in the United States: the community controls the data, chooses technical partners, and decides on uses. This model is most respectful of rights. It is also the slowest and most costly, because it requires institutional capacity that many small communities do not yet have.

The third is the platform model: companies like Google (via its Woolaroo project for endangered languages) or Meta (via its work on resource-poor languages in NLLB, No Language Left Behind) invest in multilingual tools that include endangered languages. These initiatives have real reach—Meta’s NLLB model covered 200 languages at its July 2022 release, including several dozen in vulnerable situations. But they acutely pose the control question: models are proprietary or semi-open, training data is not always traceable, and communities have no formal governance mechanism.

Generative AI is profoundly transforming how we work and create, including in fields as unexpected as cultural preservation. But technical acceleration also creates pressure on governance frameworks, which struggle to keep pace with deployments.

The real challenge in the coming years is building a framework that allows all three models to coexist without the most powerful absorbing or neutralizing the other two. This requires legal provisions on the ownership of indigenous linguistic data—several Canadian and Australian states have begun legislating in this direction—and technical standards that allow language models to function in federated mode, without corpus centralization.

The horizon is visible. In ten years, technical tools for documenting and teaching any language with a few thousand speakers will be accessible and inexpensive. The real question is who will have built the corpora by then, under what conditions, and to whom they will belong. For communities whose language has fewer than 500 active speakers today, this window is probably closing in five to ten years. Past that threshold, there will be little left to document. This calendar, more than any technical consideration, should dictate the pace of political and legal decisions to be made now.

Sources

1. Brookings Institution — “Can small language models revitalize indigenous languages?” : https://www.brookings.edu/articles/can-small-language-models-revitalize-indigenous-languages/
2. UNESCO — Atlas of Endangered Languages in the World : https://www.unesco.org/languages-atlas/
3. Te Hiku Media — Māori speech recognition project : Te Hiku Media, Wellington, New Zealand (te hiku.org.nz)
4. Meta AI — “No Language Left Behind” (NLLB, 2022), technical report published on arxiv.org
5. FNIGC (First Nations Information Governance Centre) — Indigenous Data Sovereignty OCAP Principles, Ottawa
6. Proceedings of the National Academy of Sciences (PNAS) — Cámara-Leret & Bascompte, study on medicinal knowledge linked to indigenous languages (June 2021) : https://www.pnas.org/doi/10.1073/pnas.2103683118
7. CARE Principles for Indigenous Data Governance — Global Indigenous Data Alliance
8. UNESCO / UN Info — threatened indigenous languages (40%) : https://news.un.org/fr/story/2022/12/1130582
9. Rosetta Project / ELCat — rate of language loss : https://rosettaproject.org/blog/02013/mar/28/new-estimates-on-rate-of-language-loss/
10. Te Hiku Media — first ASR for te reo Māori : https://nz.linkedin.com/company/te-hiku-media
11. University of Hawaiʻi Foundation — Hawaiian language revitalization : https://uhfoundation.org/saving-hawaiian-language
12. Meta NLLB-200 — 200 languages, July 2022 : https://techafricanews.com/2022/07/08/meta-announces-the-launch-of-nllb-200-a-language-ai-model/
13. Te Ara Encyclopedia NZ — kōhanga reo founded in 1982 : https://teara.govt.nz/en/maori-education-matauranga/print