AI Systems Lack Training Data in Most African Languages

Most AI systems are trained primarily in English, major European languages and Chinese, while African languages remain severely underrepresented according to a report published in Nature. 9 doctors per 10,000 people.

A 25-year-old woman identified as Falmata, whose first language is Shuwa spoken by just 4% of families in northeastern Nigeria, could not communicate her child's symptoms to Hausa-speaking medical staff at a displacement camp clinic. The child received rehydration salts without instructions on using boiled water, leading to a subsequent health crisis after a neighbor omitted a key safety detail.

Language barriers have contributed to delays in HIV diagnosis, treatment errors in malaria cases that caused an estimated 608,000 deaths in 2022, and reduced adherence to tuberculosis regimens. Initiatives such as African Next Voices are creating multilingual health datasets for languages including isiZulu, Hausa, Yoruba and Dholuo, while Lesan AI targets communication needs in the Horn of Africa.

Africa hosts less than 1% of global data centre capacity and only 5% of African AI researchers have access to sufficient computing power for advanced model training.