Somali was written down on purpose. AI is learning it by accident

By Sharmarke Mohamed
Wednesday October 14, 2026

Machines learned Somali from a thin and largely accidental slice of the internet, and it shows. Somali institutions should treat the written language as public infrastructure, and insist that AI is tested in Somali before it is trusted with health, schooling or jobs.

In August 1974, Somalia sent tens of thousands of students and teachers into the countryside. They attached themselves to nomad groups and became their tutors. A UNICEF review written nearly two decades later recorded that they were given "nothing except a portable blackboard and a box of chalk". What they carried was a language that had become an official written language less than two years earlier, on 21 October 1972, and a slogan: "Haddaad taqaan bar, haddaadan aqoon baro." If you know, teach. If you do not know, learn.

Written Somali did not simply happen. Its script was debated for years, recommended by committees and then carried across the country on foot. The folklorist John William Johnson, who has spent a career on Somali oral poetry, has compared the campaign with Atatürk's literacy reforms in Turkey.

Half a century later, a second decision about written Somali is being made, and this time nobody is making it on purpose. The AI systems that now answer in Somali learned the language largely from whatever Somali happened to be online. More and more of the Somali that people read, from translated leaflets to homework help, will be written by those systems. Unless Somali institutions take a hand, the Somali that machines speak back to us will be shaped by accident rather than by anyone accountable for it.

Start with how thin that accident is. Common Crawl, a free archive of the web, is the raw material for much of modern AI: the Mozilla Foundation found that nearly two thirds of the 47 large language models it studied had been trained on it. In Common Crawl's latest crawl, Somali makes up 0.0056 per cent of pages, about one page in 18,000. Dutch, a language with a broadly similar number of speakers, makes up 1.7 per cent, some 300 times as much. English alone is about 42 per cent.

Then look at what that sliver contains. This year a Somali researcher, Khalid Yusuf Dahir, published SomaliWeb, a cleaned Somali corpus, and documented what he found on the way. In the "cleaned" Somali release of one major public dataset, more than half the documents contained garbled characters and about one in six was an exact duplicate. Around 60 per cent of the text he assembled came from news sites and blogs. Encyclopaedic writing, the careful explanatory prose a machine most needs if it is to answer a question about medicine or the law, was under 1 per cent. Dahir's own summary is that the Somali internet skews towards the diaspora, news, politics and religion.

You can see the result. In a second study, SomaliBench, Dahir put 100 harmful requests, the kind a responsible AI system is built to refuse, to four open models in English and in Somali. One model refused all 100 in English and 7 in Somali. The best Somali score was 60. Most of the Somali failures were not fluent, dangerous answers. They were confusion: the wrong language, broken sentences, replies to a different question. The models did not understand Somali well enough to see what they were being asked.

Dahir is candid about the limits: he translated and checked the prompts himself. The four models are also small ones, not the commercial chatbots most people use. It is a warning rather than a verdict. But it matches what researchers at Brown University found in 2023 with GPT-4: harmful requests that got through less than 1 per cent of the time in English got through about 79 per cent of the time once translated into Zulu, Scots Gaelic, Hmong and Guarani. Safety learned in English does not travel well.

Confusion is not harmless. A system that cannot recognise a dangerous request in Somali cannot be relied on to recognise a frightened one, or a question about a child's medicine. In English, a wrong answer often looks wrong. In Somali, to someone who turned to the machine because there was no one else to ask, it can look like help. Every aid agency, ministry, school or employer that puts one of these tools in front of Somali speakers is making a bet on it.

There is a less obvious point in Dahir's numbers. If most machine-readable Somali is news and blogs, then Somali journalism, including sites like this one, which has been publishing since 1999, has probably done more to teach artificial intelligence Somali than any ministry or university. That happened without anyone deciding it should. The Somali media are already the machines' teachers. They have simply never been asked what they are teaching.

That is why I believe written Somali should be treated as public infrastructure, like the script itself: built deliberately, maintained, and governed on terms the community sets. Somali is not alone in needing this. The tool Common Crawl uses to sort pages by language recognises about 160 languages; the world has more than 7,000. Somali is at least counted. Most languages are not.

Some communities have stopped waiting. In Aotearoa New Zealand, the Māori media organisation Te Hiku Media asked speakers to record themselves reading sentences. More than 2,500 people produced over 300 hours of speech in ten days, and the data is governed by a licence under which it may be used only for the benefit of Māori people. In South Africa, Lelapa AI and its partners have written a licence for African language data under which non-African companies pay, and the fees fund more data.

Somalis also know the other way a technology can arrive. After the state collapsed in 1991, Somali telecoms grew rapidly with no recognised government at all, and mobile money followed. By 2018 the World Bank found that more than 70 per cent of adults used it regularly, called it one of the most active mobile money markets in the world, and noted that it still lacked robust consumer protection. The technology arrived years before the rules. We should not repeat that with tools that will speak to our children in their own language.

The demand that follows is simple. Before an AI tool is trusted with health advice, schooling or jobs for Somali speakers, it should be tested in Somali, by Somali speakers, and the results published. And the institutions that hold good Somali text, broadcasters, universities, publishers and news sites among them, should decide together on what terms it is shared, rather than leave that to the next crawler.

I have an interest to declare. I run a company that builds AI assistants for education, health and employment programmes, including in Somali. The standard I am arguing for should apply to companies like mine first.

The machines do not yet know Somali, and they will learn it from someone. Half a century ago the country answered that question on purpose, with a slogan and a box of chalk. Haddaad taqaan bar. Those who know the language, and the institutions that keep it, now have some teaching to do.

Sharmarke Mohamed is the founder of Hikmad.AI, a London-based company building AI assistants for education, health and employment programmes, including in Somali.