AI Understands Only 10–20% of Hausa Despite 94 Million Speakers, Builder Says

Hausa, spoken by nearly 94 million people in Nigeria and tens of millions more across the Sahel, remains severely underrepresented in AI systems, with large language models correctly understanding only 10–20% of Hausa sentences. The gap stems not simply from a lack of data but from structural issues: informal online Hausa omits phonemic characters, two distinct writing scripts exist, and heavy code-switching with English and Arabic makes clean data collection difficult. A developer spent three years addressing this problem, founding AI Bauchi in October 2022 at Abubakar Tafawa Balewa University, which has since grown to over 500 community members in Bauchi State. Rather than rushing to build a full generative model, the community took a deliberate, incremental approach — starting with a Hausa text-to-speech model to establish a clean data foundation. The effort highlights how standard web-scraped corpora produce a degraded, flattened version of the language, which then propagates errors through tokenizers and downstream AI models.