What LLM Support for Indigenous American Languages Looks Like Today

Ask a large model to write in Navajo and it will write something. That something is very often not Navajo, and the model will not tell you. The failure mode here is substitution rather than refusal, and it is the reason this ground needs its own page.

The landscape, language by language

“Indigenous American languages” covers many hundreds of languages across two continents, in dozens of unrelated families, with wildly different situations. Four points on that range:

• Quechua — several million speakers across Peru, Bolivia and Ecuador, official or co-official in places, and a dialect continuum usually treated as a language family rather than one language. Ayacucho Quechua appears in FLORES-200, so there is at least a translation test set (NLLB, 2022).
• Guaraní — co-official in Paraguay and spoken by a large share of the population, including many non-Indigenous speakers. The clearest case that official status does not produce a corpus.
• Navajo (Diné bizaad) — the most spoken Indigenous language in the United States, on the order of a hundred and fifty thousand speakers, with an orthography using several diacritics and a high-tone mark, and a verb morphology of considerable complexity.
• Cherokee — written in a syllabary invented in the nineteenth century, with a speaker population now numbered in the low thousands and active revitalisation programmes.

No frontier lab publishes a corpus share for any of these. The public proxies are stark: check any of them against the list of Wikipedias or Common Crawl’s language statistics, and most of them do not appear in the reported tail at all. In the taxonomy of Joshi et al. (ACL 2020) the great majority sit in the bottom class, with essentially no labelled or unlabelled resources (The State and Fate of Linguistic Diversity).

The substitution failure

A model with no coverage of a language still has to emit tokens. It cannot emit nothing, and nothing in its training taught it that this particular language is one it does not know — there is no “languages I know” list inside the weights. So it produces the highest-probability continuation given a prompt that says “Navajo”, and the nearest region of its representation space contains whatever it saw near that word.

What comes out takes a few recognisable forms. Sometimes it is the dominant regional language of the area — Spanish-structured text for an Andean or Mesoamerican language, occasionally with a scattering of real loanwords. Sometimes it is a small stock of genuinely known words, mostly greetings and place names that appear in English-language text about the language, padded out with invented morphology. Sometimes it is a different Indigenous language entirely, because the model has learned the category more strongly than the members. And sometimes it is orthographically plausible nonsense: the right diacritics, the right letter frequencies, no words.

The common thread is that none of these is accompanied by a hedge. The model does not say it is unsure, because the calibrated “I do not know” behaviour was trained in English on English questions — the argument in why low-resource languages hallucinate more applies here in its strongest form. For a revitalisation context this is genuinely dangerous: a learner cannot distinguish invented Navajo from Navajo, and material generated this way can enter teaching use.

How to see it yourself

You do not need to speak the language to detect the failure, which is the useful part:

1. Ask for a short passage in the target language, then ask the model, in a fresh conversation with no context, to translate that passage back into English. Compare with what you asked for. Round-tripping failure with no other symptom is the signature of generated non-language.
2. Ask for the same passage three times at a nonzero temperature and compare the outputs. Real coverage produces variations on the same vocabulary; no coverage produces three unrelated word-stocks.
3. Ask for a gloss: request the passage with a word-by-word morphological breakdown. Invented text produces confident glosses whose morphemes do not recur consistently across the passage.
4. Check any concrete word against a community dictionary where one is published. One lookup usually settles it.

The part that is not about data volume

Everywhere else in this cluster the conclusion is “collect more data”. Here that conclusion is contested by the people whose data it is, and anyone building in this area needs to know that before starting.

Many Indigenous communities hold that their languages are not a public resource to be scraped, that language material carries obligations about who may use it and for what, and that the history of extractive research is a reason for caution rather than an old grievance. This is an articulated position with instruments behind it: the CARE Principles for Indigenous Data Governance — Collective benefit, Authority to control, Responsibility, Ethics — published by the Global Indigenous Data Alliance (GIDA), and licences built for the case where a community wants distribution under continuing community control, such as the Kaitiakitanga License developed by Te Hiku Media for Māori language data (Te Hiku Media).

The practical upshot is that “is this language supported?” is not always the right question. For some languages the answer the community wants is that a general-purpose model should not be able to generate it, and a small community-controlled tool should. A plan that treats consent as a compliance step to be cleared has misunderstood the situation.

If you are building something here

Start with the community, not the model. If there is a language authority, a tribal college, or an existing revitalisation programme, they own the orthography decision, the variety decision and the question of whether the project should exist. That is not a courtesy; they hold the only reference data there is.

Then build for assistance rather than autonomy. Tools that help a speaker — transcription review, dictionary lookup, search over existing recordings, keyboard and spellcheck support — are valuable now and degrade safely. Generation of new text in the language is the least reliable thing a model can do here and the most likely to contaminate a small corpus with material nobody can trace. The collection methodology, if the community wants collection, is in crowdsourcing training data for a low-resource language.