Cohere Labs announced the release of AfriAya on December 16, 2025. Developed by Ugandan engineers Kato Steven Mubiru and Bronson Bakunga, this vision-language dataset aims to improve AI's representation of African cultures and contexts, featuring image-caption pairs in 13 African languages. The initiative addresses the historical lack of African culture in AI training data.
The AfriAya project is part of a collaborative effort within the Cohere Labs Open Science Community. It aims to capture Africa's cultural richness and rectify common AI training issues where models misidentify local foods and clothing. By using local images and culturally relevant captions, AfriAya enhances AI's understanding of African identities. The team plans to expand the dataset to 25 languages.
AfriAya tackles previous AI models' limitations in recognizing African cultures. For example, during discussions, a model misidentified ugali as "a starchy meal." AfriAya ensures AI accurately identifies cultural artifacts, fostering "visual sovereignty" for African communities to narrate their stories. The dataset has attracted interest, with presentations at Masakhane and institutions like the Swiss Federal Institute of Technology.
Despite its goals, issues remain such as undisclosed dataset size and benchmarks. Some skepticism exists about community involvement and the effectiveness of collaborations. The expansion of the dataset and its impacts with increased language coverage are yet to be seen. Participation from AI experts will be vital.
Cohere Labs envisions AfriAya as a long-term investment in cultural diversity in AI. This aligns with organizations like African Next Voices, which launched a speech dataset to enhance African language representation in AI. As detailed in the Slator article and Down to Earth report, AfriAya and similar efforts are changing how AI understands diverse cultures.
