Speakers First — what AI can do for old and rare languages

On November 4, 2026, this site published its short definition of AIciety in 55 languages, from Manx, whose last native speaker died in 1974, to Navajo, Greenlandic and Toki Pona. One editor did it in a single day, with Claude; none of the pages in small languages has yet been checked by a speaker. Fourteen months earlier, Wikipedia’s language committee had agreed to close its Greenlandic edition after machine translation filled it with errors. Both facts belong to the same story. AI can now write in almost any language. Whether that helps the world’s rare languages survive, or hastens their decline, depends less on the models than on who checks their output, who owns the data, and whether speakers stay in charge.

A note on this essay: it was written with Claude, the model that made the 55 translations it discusses. We have tried to report the failures of machine translation in small languages, including our own, as plainly as its successes.

Forty percent at risk

UNESCO counts more than 6,700 languages spoken in the world and says that at least 40 percent of them are threatened with extinction. The United Nations has declared 2022 to 2032 the International Decade of Indigenous Languages. Languages rarely die because a dictionary is missing. They die when parents stop speaking them to their children, usually because another language promises school, work and respect. Technology cannot reverse that choice. But it increasingly decides which languages are usable in daily life: which ones have a keyboard, a spellchecker, subtitles, a voice assistant – and now a chatbot that answers in them.

What machines already do

They read the dead. In February 2024 the Vesuvius Challenge awarded its $700,000 grand prize to three students who had used machine learning to find ink in X-ray scans of a papyrus carbonized by Vesuvius in AD 79. They recovered 15 partial columns of Greek, a text about pleasure, without unrolling the scroll. In July 2025 Google DeepMind presented Aeneas, published in Nature, which suggests missing letters, dates and places of origin for Latin inscriptions by comparing them with almost 150,000 others; its predecessor Ithaca did the same for Greek. For Akkadian, the language of cuneiform, researchers in Israel trained a neural translation model because hundreds of thousands of clay tablets remain untranslated and few people can read them. Its output is a first draft for experts, and the authors warn that it sometimes produces fluent English that does not match the source.

They listen to the living. Te Hiku Media, a Māori broadcaster in New Zealand’s far north, asked the public to read sentences in te reo Māori. In ten days more than 2,500 people signed up to read over 200,000 phrases, yielding more than 300 hours of speech, NVIDIA reported; the speech recognizer trained on it transcribes te reo with 92 percent accuracy and mixed Māori-English speech with 82 percent. In May 2023 Meta released Massively Multilingual Speech, which recognizes and synthesizes speech in 1,107 languages and identifies more than 4,000 – trained largely on recordings of the New Testament, about 32 hours per language. Meta itself warned that the model may mistranscribe words.

They learn from very little. Kalamang, spoken on islands off the coast of New Guinea, has fewer than 200 speakers and almost no presence online. In 2023 researchers built a benchmark, Machine Translation from One Book, that asks a model to translate Kalamang using only a linguist’s grammar, a word list and some example sentences. In the original tests, a human who learned from the same materials did better than every model. In February 2024 Google said Gemini 1.5 Pro had learned to translate English to Kalamang “at a similar level to a person learning from the same content”, BGR reported; the technical report gives the details. A later study by Seth Aycock and colleagues found that almost all of the gain came from the example sentences in the book, not from its grammatical explanations. Models learn from language in use – which is exactly what only speakers can supply.

They help people write. Much of the useful work is unglamorous. In Tromsø, the Sámi language technology groups Giellatekno and Divvun maintain GiellaLT, an open framework for rule-based language technology in minority and indigenous languages, with partner teams in Alberta and Nuuk. It turns a linguist’s grammar into keyboards, spellcheckers, dictionaries and machine translation. Masakhane, a grassroots community “for Africa, by Africans”, builds language technology for African languages from within.

The doom spiral

The same tools fail where nobody checks them. In September 2025 Jacob Judah described in MIT Technology Review how Wikipedia’s small editions had become dumping grounds for machine translation. Greenlandic Wikipedia had about 1,500 articles, and Kenneth Wehr, who grew up in Germany, now teaches Greenlandic in Denmark and had run the edition since about 2021, suspects that only one or two Greenlanders ever contributed. A growing number of articles appeared to be machine-translated, and errors abounded: one entry claimed Canada had 41 inhabitants. Wehr asked for the edition to be closed, and Wikipedia’s language committee agreed. Volunteers in four African languages estimated that 40 to 60 percent of their articles were uncorrected machine translations. On some Hawaiian pages, Noah Haʻalilio Solomon, an assistant professor of Hawaiian, found about 35 percent of the words incomprehensible.

This matters beyond Wikipedia, because for many small languages Wikipedia is a large part of the text that AI models learn from – by one 2020 estimate cited in the article, more than half of the training data for machine-translation models in Malagasy, Yoruba and Shona. Bad machine text becomes training data, the next model writes worse text, and that text is published again. “Wikipedia will be reflected in the AI models for these languages,” the linguist Trond Trosterud told the magazine. Solomon put the stakes more bluntly: “If this is the Hawaiian that is going to exist online, then it will do more harm than anything else.” The article also has a counterexample: Inari Saami, whose 6,400 Wikipedia articles are each copy-edited by a fluent speaker. “If we keep putting good material in, then sooner or later, we will get something out,” said Fabrizio Brecciaroli of the Inari Saami Language Association.

What we learned from 55 languages

Translating our own short text showed both sides at close range. For French or Japanese, checking a translation is easy, because there are millions of readers and ample references. For small languages, checking was most of the work – and a model checking a dictionary is no substitute for a speaker. The translators – parallel instances of Claude, each with a written brief – were instructed to check terms in Phil Kelly’s English–Manx dictionaries, in the Cornish dictionary of the Akademi Kernewek, in the soblex dictionary for Upper Sorbian, in Te Aka for Māori. For Greenlandic, the word for artificial intelligence, silassorissuusiaq, came from the term base of Oqaasileriffik, Greenland’s language secretariat, where its language council had approved it; all other word forms were checked with the open GiellaLT-based Greenlandic analyser, which confirms that a form exists, not that a sentence is idiomatic. At best, the machine was as good as the human institutions behind it.

Where those institutions were out of reach, the quality dropped. For Tamil and Hindi, Wikipedia turned away most automated requests and the official terminology glossaries could not be reached, so the pages rest largely on the model’s own knowledge. In Navajo, where the model had few reliable sources for legal and economic terms, many sentences had to be simplified; artificial intelligence became “a computer that thinks like people”. Every page for a small, revived or indigenous language says that it is a best-effort AI translation and asks for corrections. On November 4 we wrote to Culture Vannin for Manx, the Akademi Kernewek for Cornish, the Māori Language Commission, Diné College for Navajo, the Masakhane community for Yoruba and Swahili, and the official language bodies for Tamil and Hindi, asking them to check our work – and, for the indigenous and revived languages, offering to take the pages down if their communities prefer.

One question we have not settled. Our Greenlandic page is exactly the kind of text that could end up in the next training set. A label helps human readers; it does not stop a crawler. Whether AI-translated pages in small languages should be kept out of training data until speakers have checked them is a question we are putting to the communities concerned.

Speakers first

What follows from this is not a rule for regulators but a set of habits for anyone building or using such tools.

A language is a spontaneous order. The AIciety definition compares society’s relationship with AI to “a market or a language” that “knows more than any of its participants”. No ministry can keep a language alive on its own, and no model can either. In the end only speakers can, by choosing to use it. AI helps when it lowers the cost of that choice: a keyboard that knows the macrons, a spellchecker for the tone marks, subtitles for a children’s program, a tutor available at midnight for an adult learner.

Those who speak a language should decide what is done with it. Te Hiku Media declined to upload its archives to global platforms whose terms required giving up rights and built its own; it releases its work under a Kaitiakitanga license that allows use only for the benefit of Māori. That is a contractual answer – a community setting terms for data it holds as guardian – rather than a call for a ban. Other small-language communities can copy it: consent and contract, not extraction and not prohibition.

Machines should not write the corpus they learn from. Machine text in a small language should be labeled, correctable and, where possible, checked before it is published. In a language with few readers, an error has fewer people to catch it, and it lasts longer. By that rule, speakers should have checked our own pages before publication; they did not, and we are now asking them to.

Revival has always been human work. Hebrew became a spoken mother tongue again in the late 19th and early 20th centuries without any machine. UNESCO reclassified Cornish from “extinct” to “critically endangered” in 2010. The Isle of Man, where the last native speaker of Manx, Ned Maddrell, died in 1974, is celebrating 2026 as the Year of the Manx Language. The new speakers of revived languages are learners, and learners are where AI may help most – as long as the teachers, dictionaries and communities remain the authority.

The machine can read, listen and draft. It cannot be the community. Used with that in mind, it may give some of the world’s smallest languages more reach than they have had in centuries. Used without it, it will bury them under text that nobody speaks.

Related

AIciety in other languages · What is AIciety? · Stochastic Parrots · Human in the Loop · Scaling Laws

Sources: UNESCO Austria: UNESCO celebrates the International Decade of Indigenous Languages · University of Kentucky: Grand prize discovery made in 2,000-year-old Herculaneum scrolls (February 5, 2024) · MIT Technology Review: DeepMind’s AI Aeneas helps historians interpret Latin inscriptions (July 23, 2025) · Gutherz et al.: Translating Akkadian to English with neural machine translation, PNAS Nexus (May 2, 2023) · NVIDIA: Māori Speech AI Model Helps Preserve and Promote New Zealand Indigenous Language (January 16, 2024) · IEEE Spectrum: Indigenous AI voice models (May 21, 2026) · Meta AI: Introducing speech-to-text, text-to-speech, and more for 1,100+ languages (May 22, 2023) · Tanzer et al.: A Benchmark for Learning to Translate a New Language from One Grammar Book (2023) · BGR: Google Gemini 1.5 taught itself a rare language from a grammar manual (February 16, 2024) · Gemini Team, Google: Gemini 1.5 technical report (arXiv) · Aycock et al.: Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? (2024) · GiellaLT · Arctic Peoples: Language revitalization – GiellaLT in Tromsø · Masakhane · Jacob Judah: How AI and Wikipedia have sent vulnerable languages into a doom spiral, MIT Technology Review (September 25, 2025) · Wikipedia: Cornish language · Wikipedia: Revival of the Hebrew language · Wikipedia: Ned Maddrell · Culture Vannin: Island celebrates the official launch of the Year of the Manx Language 2026 · aiciety.net: What is AIciety?

Leave a Reply

Your email address will not be published. Required fields are marked *