Dec 16, 2025 5 minutes read # How a Community Effort is Teaching AI to See Africa’s Richness AfriAya, a vision-language dataset, addresses Africa’s underrepresentation in AI. Created by Ugandan engineers, it builds on Aya improving AI’s recognition of African cultures, covering 13 languages, with plans to expand. Part of Cohere Labs’ inclusion efforts, it aims for cultural robustness in AI. Research doesn’t always begin in a university lab or with an official title. Sometimes it starts with a question that refuses to leave you alone. For two engineers in Uganda, and Leads within the Cohere Labs [Open Science Community](/content/research/open-science/index.html), that question was simple: _What would it take for AI systems to truly see Africa — its foods, its places, its cultures — instead of describing them generically?_ That question became AfriAya, a community-built vision-language dataset grounded in African contexts. And while AfriAya is still early in its journey, its story already reveals what’s possible when open science, local expertise, and global collaboration meet. ##### **The Roots: Two Engineers, One Community, and a Missing Piece in AI** Kato Steven Mubiru first joined the Cohere Labs Open Science Community as the Regional Lead for Africa. His role centered on amplifying local research and creating space for people across the continent to collaborate on ML research. His posts routinely surfaced the same underlying truth: Africa’s linguistic and cultural richness was dramatically underrepresented in AI research; however, the potential to make an impact here was abundant. One of the people who noticed was Bronson Bakunga. He had been working on machine translation for East African languages and, like Steven, felt the gaps in existing datasets every day. Seeing Steven’s community updates, he joined the Cohere Labs community to find collaborators, mentorship, and direction. When Cohere Labs launched [Expedition Aya](https://sites.google.com/cohere.com/expedition-aya-2025/home), a 6-week global incubator for new, community-led research projects, Steven and Bronson saw their opening. 14 months since release, [Aya](/content/research/aya/index.html) itself had already become a landmark multilingual initiative: an open, community-trained foundation model spanning over 100 languages, many of which had little to no representation in previous LLMs. Subsequent models added increased performance in 23 languages, and the additional [Aya Vision](http://google.com/search?q=aya+vision+blog&oq=aya+vision+blog&gs_lcrp=EgZjaHJvbWUyBggAEEUYOTIKCAEQABiABBiiBDIHCAIQABjvBTIHCAMQABjvBTIHCAQQABjvBTIHCAUQABjvBdIBCDI2OTZqMGo0qAIAsAIA&sourceid=chrome&ie=UTF-8?ref=cohere.com/blog) introduced Cohere Lab’s first vision-language model. Expedition Aya invited researchers around the world to connect with new collaborators build on top of that work. Steven and Bronson asked: _Could we do for African images what Aya did for African languages?_ That’s where AfriAya began. ##### **Why This Matters: AI Still Doesn’t “See” Africa** In early project meetings, Bronson described a moment that crystallized the problem. He showed a model an image of ugali, a staple dish across East Africa, and asked what it saw. The model responded with: _“A starchy meal.”_ To a machine without cultural context, it wasn’t wrong, but it also wasn’t right. Africa appeared in models as a set of categories, rarely as itself. Steven, Bronson, and their collaborators wanted to push past that. A model should recognize _chapati_, _matooke_, or _kakalika_; it should distinguish the differences in landscapes of Kenya, Uganda, and Rwanda; it should caption people, environments, and objects with specificity rather than stereotype. They envisioned a dataset rooted in African perspectives, not just African content. ##### **From Two People to a Continent-Spanning Collaboration** Building a dataset is slow, meticulous work. Building one across 23 African languages where little organized data already exists is even harder. “Some of the challenges we faced were not technical,” Steven reflected. “For example, enthusiasm in the community.” The team was spread across time zones, skill levels, and internet access realities. People joined after work or between other responsibilities. Many were new to research. And the scale of the project, collecting, annotating, and verifying images across dozens of languages, felt intimidating. The open science culture in the Cohere Labs community proved critical. Mentors stepped in with annotation strategy sessions and dataset pipeline reviews. Aya contributors offered language expertise and guidance. Most importantly, fellow community members showed up with encouragement. “This was the first time I was leading something that was this big,” Bronson said. “It really comes down to the basic things, like celebrating the wins, even if it's the small ones. Even if it's one language over the line.” Progress continued, one contributor at a time. ##### **The Momentum: AfriAya’s Release and Where It’s Going Next** After months of steady contributions supported by Cohere Labs researchers, community annotators, and local language experts, the team released AfriAya v1, covering 13 languages with image–caption pairs grounded in African contexts. The release sparked immediate interest. Steven and Bronson were invited to present the project at Masakhane, the continent’s leading NLP research community, and again at the Swiss Federal Institute of Technology. Today, AfriAya is moving toward v2, which aims to expand to 25 languages and support fine-tuning of Aya Vision for African use cases. It’s early, still growing, and still imperfect. But that’s exactly the point: open science allows research to begin before it’s “finished,” grow as people join, and improve as the community evolves. ##### **The Broader Thread: Cohere Labs and Low-Resource Languages** AfriAya isn’t an isolated project; it’s part of Cohere Labs’ long-term investment in linguistic and cultural inclusion in AI, and support of open science. Through Aya, community members across 100+ languages built evaluation sets, fine-tuning data, and local benchmarks. Contributors from regions with historically limited AI infrastructure, South Asia, Latin America, the Middle East, and across Africa, played central roles. These collaborations have already led to: - [INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge](/content/research/papers/include-evaluating-multilingual-language-understanding-with-regional-knowledge-2024-11-29/index.html) – a comprehensive knowledge- and reasoning-centric benchmark across 44 languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed. - [NeoBabel: A Multilingual Open Tower for Visual Generation](/content/research/papers/neobabel-a-multilingual-open-tower-for-visual-generation-2025-07-09/index.html) – a multilingual image generation framework that sets a new Pareto frontier in performance, efficiency, and inclusivity, supporting six languages: English, Chinese, Dutch, French, Hindi, and Persian. - [Global MMLU](/content/research/papers/global-mmlu-2024-12-05/index.html) – a multilingual evaluation set spanning 42 languages, including English. This dataset combines professional translations with post-edits (14 languages), crowdsourced translations (11 languages), and machine translations (16 languages). - [Kaleidoscope: Exams for Multilingual Vision Evaluation](/content/research/papers/kaleidoscope-exams-for-multilingual-vision-evaluation-2025-04-10/index.html) – a multilingual, multimodal benchmark designed to test AI systems with real-world, in-language image+text Q/A across 18 languages, including Hindi, Telugu, and Nepali. - [M-RewardBench: Evaluating Reward Models in Multilingual Settings](/content/research/papers/m-rewardbench-evaluating-reward-models-in-multilingual-settings-2024-11-05/index.html) – a benchmark for 23 typologically diverse languages. M-RewardBench contains prompt-chosen-rejected preference triples obtained by curating and translating chat, safety, and reasoning instances AfriAya builds on this foundation. It extends the multilingual vision of Aya into multimodal territory, inviting people across Africa to define what their world looks like through AI, not simply respond to outsiders’ definitions of it. As we look toward what’s next at Cohere Labs, we are setting our sights on a bold frontier: building AI that genuinely understands the richness of human cultures. Our upcoming research pushes beyond conventional multilingual coverage, aiming to ensure AI systems perform well across diverse languages, contexts, and use cases, with an emphasis on cultural robustness. We're excited to share more as our research unfolds. Stay tuned! ##### **Advice for Beginners: Start Before You Feel Ready** When asked what they would tell someone who wants to get involved in research but doesn’t know where to begin, Steven’s answer was straightforward: “Join a community. Even if you’re quiet, read the messages. Reach out. Try hard things. Be humble when you succeed, grateful when people help.” Bronson added, “Work on problems you are personally attached to, because half of it is motivation. The other half is having the right people. Once those two things come together, it leads you in the right direction.” Their story is proof. AfriAya didn’t begin with funding or institutional backing. It began with curiosity, commitment, and the belief that people working together can reshape the future of AI. If there’s a research question you’re passionate about, we’d love to hear from you, either via our [open science community](/content/research/open-science/index.html) or [newsletter](/content/research/newsletter/index.html). We are planning new ways to foster collaboration, and are always looking to connect with new researchers who share our mission of changing where, how, and by whom breakthroughs happen.