The Tongue Commons

The problem

The language models that now draft contracts, diagnose from symptoms and tutor children were trained on the text of the internet, and the internet is written in about a dozen languages. Africa speaks 2,144 of the world's living languages, and for nearly all of them the training data does not exist: no digitised books, no transcribed radio, no parallel text against English, no labelled speech. A model asked a question in Hausa, spoken by 80 million people, answers worse than it does in Icelandic, spoken by 400,000, because Iceland put its literature online and Nigeria's Hausa press did not exist in a form a machine could read.

The work of fixing this has been done, so far, by volunteers. Masakhane, founded in 2019 as a collective of African researchers with no institutional budget, built the first baseline translation models for more than thirty African languages and the largest human-annotated datasets for named entities and news classification in twenty of them. Meta's No Language Left Behind project released a model covering 200 languages, including many African ones, in 2022, and the Lacuna Fund has financed open text and speech datasets for several. It is a beginning, and it is a beginning built on evenings and weekends and a few million dollars of philanthropy, for the languages of a billion people.

The stakes are not academic. № 029 argued that intelligence should be infrastructure rather than a subscription, and proposed a Model Commons for the United States: open models and public compute, as the land-grant universities once made knowledge public. № 014 and № 027 described what happens to the people the machines can serve. The people the machines cannot understand are excluded from all of it: the farmer who cannot ask the model about the blight in Wolof, the clinic that cannot use the diagnostic in Oromo, the child who cannot be tutored in Shona. The century's largest population will speak to the century's most powerful tool in languages the tool has never heard.

The proposal

The Tongue Commons: a ten-year American programme to build open datasets, speech corpora and open models for every African language with more than a million speakers, about 150 of them, under open licences owned by the language communities and the universities that build them. The programme funds African researchers, the Masakhane model at scale, with compute from the Model Commons of № 029. The Forecast Gift, the Malaria Contract and the Hospital Fleet deliver their services in the languages the Commons makes possible.

How it would work

1. The languages. Every African language with more than a million speakers, about 150, prioritised by speakers and by the absence of existing resources. The long tail of smaller languages follows as the tools built for the first 150 make each additional language cheaper.

1. The data. For each language: a text corpus assembled from existing print, broadcast and web sources with the rights cleared and paid; a parallel corpus against English and French built by paid translators; a speech corpus recorded from native speakers across dialects and paid for; and the annotations that let a model learn the language's structure. The data is released under an open licence, with the language community and the building institution named as owners.

1. The builders. African universities and research collectives, funded directly, at African rates that let a linguistics department in Ibadan or Addis Ababa employ fifty people to build its language's corpus. The programme pays for the work at the wages of the place where it is done, and the Masakhane community, which knows how, is funded to lead rather than to volunteer.

1. The compute. Training runs on the public compute of № 029's Model Commons, allocated to the programme; the models are released open, so that any developer, company or government can build on them without asking.

1. The uses. The Forecast Gift of № 114 delivers weather warnings in every language the Commons covers; the Malaria Contract of № 113 runs its phone-reported surveillance in them; the Hospital Fleet of № 059 uses them in its clinics; the English Hour of № 104 teaches from them. A language the machine understands is a language in which every other proposal in this block can be delivered.

1. The licence. Open for use, with attribution to the community, and with a clause that revenue from commercial models trained substantially on Commons data returns a royalty to the language's institution, on the model of the Data Dividend of № 016. The people who spoke the words into the microphone own a share of what the words become.

The numbers

The languages. About 150 with more than a million speakers; about 800 million people among them.

The cost. Building a full set of text, parallel and speech resources for a language costs on the order of $2 to $5 million at African wages; 150 languages is on the order of $500 million over ten years, plus compute and the smaller languages after. Call it $100 million a year. The philanthropic and volunteer effort to date is a few million a year.

The comparison. A single frontier model training run costs more than the whole programme. The Commons is cheap because its output is data and open models, not a proprietary product, and because the work is done where the languages are spoken.

The return. Uncountable directly and enormous by any reasoning: the productivity of the tools this catalog describes, delivered to 800 million people in the languages they think in. The royalty in item 6 is the countable part.

The honest objections

"The market will build this when there is money in it." It will build it for Swahili and Hausa, eventually, as proprietary products, with the data owned by the company that gathered it. It will not build it for the 140 languages below them, and it will not release what it builds. The Commons exists for the same reason the land-grant universities did: the market builds what pays, and a public good is what does not.

"Open data will be scraped by Chinese and American companies who give nothing back." It will be used by them, which is the point of open data, and the royalty clause in item 6 is the something back. A licence that forbade commercial use would forbid the local company in Lagos that the Commons most wants to exist.

"Paying African universities at African rates is exploitation." Paying them at American rates would fund a tenth as many people and bring the work to Boston. The programme funds the departments that exist, in the currency and at the wages of the place, with the ownership of the result held there. That is not exploitation; it is how a commons is built by the people it belongs to.

"Language models are a fad and the money should go to schools." The tutor that teaches a child in Shona is a school, and the Commons is what lets it exist. The money is a hundred million a year; the school budgets of the continent are hundreds of billions; the Commons is the piece of the second that the first unlocks.

"Who decides which dialect is the language?" The community, through the institution that builds the corpus, and the corpus records the dialects rather than choosing one. A speech corpus with a thousand speakers from every region is the answer to the question the objection asks, and it is the reason the corpus is expensive.

Sources

• Masakhane, Participatory Research for Low-resourced Machine Translation: A Case Study in African Languages (2020); MasakhaNER 2.0 and MasakhaNEWS dataset papers
• Meta AI, No Language Left Behind: Scaling Human-Centered Machine Translation (2022)
• Lacuna Fund, portfolio of African language datasets, lacunafund.org
• Ethnologue, Languages of the World, count of living languages in Africa
• AfricaNLP workshops, 2023 and 2025, proceedings and community reports
• Proposals № 014 (The Fractional Republic); № 016 (The Data Dividend); № 027 (The Right to a Human Decision); № 029 (The Model Commons); № 059 (The Hospital Fleet); № 104 (The English Hour); № 113 (The Malaria Contract); № 114 (The Forecast Gift)