AI and Indigenous Languages: What Microsoft's Lab Is Building

Ask a chatbot a question in English and you will usually get a fluent, confident answer. Ask it in a language with a few thousand speakers and the experience can fall apart. Translations come out garbled, or the tool does not seem to know the language exists. That gap is the subject of this article. It matters most for Indigenous and other small-community languages, where the stakes are cultural survival and not just convenience.

I want to look at what is actually documented right now. Microsoft's AI for Good Lab page and its low-resource languages project page describe a set of programs aimed at this problem. They are interesting, but they are also easy to oversell, so I will be careful about what they do and don't claim.

Why most languages are invisible to AI

Modern language models learn from enormous amounts of text. A language with little text online, few clean datasets and no standard tests simply gives the model very little to learn from. Microsoft's project page puts it plainly: most of the world's languages are underrepresented in today's AI systems because of limited digital content, few high-quality datasets and little benchmark data to measure progress.

Those three missing pieces are worth separating, because each is a different kind of problem:

• Content: there is not enough written or spoken material in the language to train on.
• Datasets: even where material exists, it may not be collected, cleaned or licensed in a usable way.
• Benchmarks: without standard tests, nobody can tell whether a model is actually good in that language or only appears to be.

The third point is easy to overlook. If you cannot measure quality, you cannot improve it, and you cannot warn people when a tool is unreliable. In my view that is the quiet scandal of multilingual AI: a tool can look impressive in a demo and still fail the people who most need it.

What the AI for Good Lab says it is doing

The lab's framing is that language sustains culture, identity and opportunity, and that every language deserves a place in the digital world. Its stated approach is practical and partner-driven. It combines data stewardship, evaluation benchmarks, translation tools and adaptable training workflows that can be reused across languages and contexts. The goal it names is making modern AI usable and safer in low-resource settings.

I like that "safer" is in there. A mistranslation in a medical or legal setting is not just an annoyance. The lab's own page also lists "Low Resource Languages" as one of its initiatives, alongside work on fundamental rights, sustainability and the AI economy. So language is treated as a core part of the lab's mission, not a side project.

The page also highlights Inuktitut, an Indigenous language, as an example of a low-resource language where AI is being used to help preserve and expand access. It says more than 2,000 languages are at risk of disappearing. That is Microsoft's figure, and it links to a separate story about Inuktitut. I could not read that story, so I cannot describe the Inuktitut work in detail here.

The LINGUA open calls: money for data, not just models

The most concrete part of the page is a pair of funding programs. The first is LINGUA: Expanding Europe's Voices in AI. According to Microsoft, it is an open call supporting ethical, open dataset creation for European languages underrepresented in digital spaces and AI systems. The selected projects span 16 languages and dialects across 10 countries. They are led by universities, nonprofits, a government language center and a public broadcaster, and they aim to expand access to speech and text data.

The second is LINGUA Africa. Microsoft's AI for Good Lab, the Masakhane African Languages Hub, the Gates Foundation and Google.org have announced its awardees. The open call supports the development of language models for African languages. It builds on LINGUA Europe but focuses on applications, backing community-led efforts and helping people act on critical information in their own languages.

A caveat is in order. These two programs target European and African languages broadly. The page does not say they are specifically about Indigenous languages, and I do not want to stretch the claim. But the model they follow, funding community-led groups to build open data, is the same one that Indigenous language projects often argue for. I find that encouraging.

BYOL: a recipe for small languages

On the technical side, the page points to a framework called BYOL, short for Bring Your Own Language. It is described as using three steps for low-resource languages: data refinement, synthetic data generation and fine-tuning. The aim is to build language-specific models. Microsoft says these outperform multilingual baselines. A paper and code are linked from the page.

In plain terms, that means three things. First, you clean up whatever scarce data exists. Second, you generate extra training examples artificially. Third, you adapt an existing model to your language instead of hoping a giant general model happens to cover it. The claim about beating multilingual baselines comes from Microsoft's own description, so treat it as a promising result and not an independent verdict. Synthetic data deserves particular caution. If a model generates text in a language it understands poorly, it can bake errors into the very data meant to fix the problem.

What a translation tool cannot do

It is tempting to read all this as "AI will save endangered languages." I would push back on that framing. A language survives because people use it at home, in school and in community life. Software can support that. It can help learners practice, help elders' recordings become searchable, or make a phone work in a mother tongue. It cannot replace the speakers.

There are also questions that no benchmark answers:

• Who owns the data? Community knowledge, stories and recordings are not just raw material. The LINGUA calls stress ethical, open datasets, but "open" and "community-controlled" are not always the same thing.
• Who decides what counts as correct? Languages have dialects, registers and sacred or restricted knowledge. A model trained on whatever text is available may flatten them.
• Who benefits? If the commercial value of a trained model sits with a big company and the community gets a demo, that is not partnership.

The lab's emphasis on partner-driven work suggests it knows this, but a project page cannot show how well it works in practice. That would take independent reporting and the voices of the communities involved.

A correction worth making to the old story

An earlier version of this article made broader claims, including a named Google tool and a Microsoft collaboration with Indigenous communities. I have removed them because I cannot support them from the sources I can point to today. Microsoft's pages describe the programs above, and I have stuck to those. It is a small reminder that this field moves quickly and that enthusiasm is not evidence.

What to watch next

If you want to judge whether this wave of work is real, here is what I would look for over the next year:

• Published benchmarks for specific Indigenous languages, with results that independent researchers can reproduce.
• Datasets released under terms that communities themselves helped write.
• Tools that speakers actually use, like keyboards, voice assistants or learning apps, and not just research papers.
• Honest reporting of failure cases, especially in sensitive areas like health and law.

Funding open calls and releasing code, as Microsoft describes, are real steps. They are the kind of boring infrastructure that tends to matter more than flashy demos.

Your turn to think about it

Most of us never notice which languages our devices understand, because ours are already covered. For hundreds of millions of people, that is not true. So think about it for a moment. What would it mean if your grandmother's language could not be typed on your phone, searched for online or understood by the assistant everyone else uses? And who should get to decide how technology learns it? The answers will shape not only the future of AI but also which parts of human culture make it into that future.