AI speaks a hundred languages. Yours may not be one — and that can be fixed
Machine translators cover a few dozen of the world's seven thousand languages. Nahuatl, Maya and hundreds of languages of the Americas are left out or work poorly. The problem is not technical but data: AI learns from text and voice, and little of these languages exists online. The good news: communities like the Māori have already proven it can be reversed — by recording, writing and organizing, under one non-negotiable condition: the data stays with the community.
Why your language is not in the translator
An AI model learns languages the way it learns everything: by reading. For English there are trillions of words available; for Spanish, a great many; for Nahuatl or Yucatec Maya, a tiny fraction — and for hundreds of languages of the Americas, almost nothing digital. Without texts and recordings, there is nothing to learn from. The result is a vicious circle: because the language is not in the technology, it is used less digitally; because it is used less, there is less data; because there is less data, it never reaches the technology.
What is lost if we do nothing
It is not just convenience. When a language does not exist for machines, its speakers are left out of services already taken for granted: dictating a message, asking an assistant for information, subtitling a video, completing a digital government form. And there is a worse risk: that AI “learns” the language from the outside, from data taken without permission, and returns an impoverished or wrong version that ends up being taught as if it were the real one.
Proof that it can be done
The Māori people of New Zealand proved it. Their community radio network, Te Hiku Media, had spent decades recording elders speaking te reo Māori. When AI arrived, the community organized recording marathons: thousands of people donated hours of voice reading sentences. With that they trained their own speech recognizer — today it transcribes Māori better than the systems of the big companies. And they set a condition that made history: the data is neither sold nor given away; it is stewarded under their own license, in which benefit flows back to the source. African languages have their network (Masakhane), the languages of the Americas have allied researchers (AmericasNLP) — the path exists.
What it would take here
Three things, in this order. Voice: recordings of speakers of all ages — read sentences, conversation, stories. Text: parallel translations (the same sentence in Spanish and in the language), which is exactly what a machine translator needs in order to learn. Governance: clear agreements, written in plain language, about who keeps the data, what it is used for and how the benefit returns to the community — before recording the first word, not after.
That is the CIUDADAN-IA project: this outlet informs in your language, and the next step is built with you. If you speak Nahuatl, Maya or another original language, your voice is literally what is missing.