Artificial intelligence is learning Somali from a narrow, accidental fragment of the internet—and the weaknesses are already visible. Somali institutions should treat written Somali as public infrastructure and require AI systems to prove themselves in the language before trusting them with health care, education or employment.
In August 1974, Somalia dispatched tens of thousands of students and teachers to rural areas. They joined nomadic communities and taught their members. Nearly 20 years later, a UNICEF review recalled that the educators had been issued “nothing except a portable blackboard and a box of chalk”. Their other essential tool was a language that had become an official written language less than two years earlier, on 21 October 1972, along with a guiding slogan: “Haddaad taqaan bar, haddaadan aqoon baro.” If you know, teach. If you do not know, learn.
Written Somali was the product of a deliberate national effort. Its script had been debated for years, endorsed by committees and carried across the country by people travelling on foot. John William Johnson, the folklorist whose career has focused on Somali oral poetry, has likened the campaign to Atatürk’s literacy reforms in Turkey.
Half a century later, another defining decision about written Somali is taking place—but without a national plan behind it. AI systems that respond in Somali have learned primarily from whatever Somali-language material happened to be available online. Increasingly, the language people encounter in translated leaflets, homework assistance and other services will be produced by those systems. Without intervention from Somali institutions, machines will shape Somali by default, rather than under the direction of accountable authorities.
The scale of the problem becomes clear in the data. Common Crawl, a free archive of web pages, supplies raw material for much of today’s AI. A Mozilla Foundation study found that almost two-thirds of the 47 large language models it examined had been trained on the archive. In its latest crawl, Somali accounted for just 0.0056 per cent of pages—roughly one in every 18,000. Dutch, despite having a broadly comparable number of speakers, represented 1.7 per cent, about 300 times more. English alone accounted for approximately 42 per cent.
The material within that small share is also uneven. This year, Somali researcher Khalid Yusuf Dahir released SomaliWeb, a cleaned Somali corpus, and recorded the problems he encountered. In the “cleaned” Somali portion of one major public dataset, more than half of the documents contained corrupted characters, while roughly one in six was an exact duplicate. News websites and blogs made up about 60 per cent of the text he collected. Encyclopaedic material—careful, explanatory writing that AI needs to answer questions about medicine or law—accounted for less than 1 per cent. Dahir’s conclusion was that Somali online content is heavily tilted towards the diaspora, news, politics and religion.
The consequences can be measured. In a separate study called SomaliBench, Dahir submitted 100 harmful requests—the sort of prompts a responsible AI system should refuse—to four open models in both English and Somali. One model rejected all 100 requests in English but only seven in Somali. The strongest Somali result was 60. Most failures did not produce fluent yet dangerous responses; instead, the systems became confused. They answered in the wrong language, generated broken sentences or responded to another question entirely. The models often failed to understand what they were being asked in Somali.
Dahir acknowledges the study’s limitations: he translated and reviewed the prompts himself, and the four models were small rather than the commercial chatbots most people use. The findings are therefore a warning, not a final judgment. Yet they echo research from Brown University in 2023 involving GPT-4. Harmful requests that slipped through less than 1 per cent of the time in English succeeded about 79 per cent of the time after being translated into Zulu, Scots Gaelic, Hmong and Guarani. Safety developed in English does not automatically transfer to other languages.
Confusion itself can be dangerous. A system unable to identify a harmful Somali request cannot reliably recognise a distressed user or a question about a child’s medicine. An incorrect answer in English may be obviously wrong. For a Somali speaker who turns to a machine because no alternative help is available, the same response may appear credible. Any aid agency, ministry, school or employer that places such a tool before Somali speakers is taking a risk.
Dahir’s figures reveal another important reality. If news and blogs constitute most machine-readable Somali, then Somali journalism—including outlets such as this one, which has published since 1999—may have taught AI more Somali than any ministry or university. That influence emerged without a decision or oversight process. Somali media are already serving as teachers to machines; no one has yet asked them what lessons they are imparting.
Written Somali should therefore be treated as public infrastructure, much like the script itself: developed intentionally, maintained over time and governed according to terms set by the community. Somali is not the only language facing this challenge. The tool used by Common Crawl to classify pages can identify roughly 160 languages, while more than 7,000 languages are spoken worldwide. Somali is at least visible in the system. Most languages are not.
Some communities are building alternatives. In Aotearoa New Zealand, the Māori media organisation Te Hiku Media invited speakers to record themselves reading sentences. More than 2,500 people generated over 300 hours of speech in just 10 days. The resulting data is governed by a licence that permits its use only for the benefit of Māori people. In South Africa, Lelapa AI and its partners have developed a licence for African-language data that requires non-African companies to pay, with the proceeds financing the collection of additional data.
Somalis have also experienced what happens when technology spreads ahead of regulation. Following the collapse of the state in 1991, the country’s telecommunications sector expanded rapidly despite the absence of a recognised government, and mobile money later followed. By 2018, the World Bank reported that more than 70 per cent of adults used mobile money regularly. It described Somalia as one of the world’s most active mobile-money markets, while noting the lack of robust consumer protection. Technology arrived before the rules. That sequence should not be repeated with systems that will speak to Somali children in their own language.
The next step is straightforward. No AI tool intended for Somali speakers should be trusted with health guidance, education or employment until it has been tested in Somali by Somali speakers, with the results made public. Institutions that hold high-quality Somali text—broadcasters, universities, publishers and news organisations among them—should also agree on how that material can be shared, rather than surrendering the decision to the next web crawler.
I have a personal interest to disclose. I lead a company that develops AI assistants for education, health and employment programmes, including Somali-language services. The standard outlined here must apply to companies such as mine before anyone else.
Machines do not yet understand Somali well, and someone will determine how they learn it. Half a century ago, Somalia answered that question deliberately, with a slogan and a box of chalk. Haddaad taqaan bar. Those who know the language—and the institutions that preserve it—now have 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.
