Researchers argue that AI in genomics risks deepening global inequities unless African datasets, leadership, and community trust are built from the ground up.
A single number captures the problem: African populations represent roughly 3% of participants in global genomic datasets, according to estimates cited in research published in the American Journal of Human Genetics. AI systems trained on that imbalance do not generalise. They carry the bias forward.
Writing in the journal, Gaye, Bonham, and Mersha argue that trustworthy AI in African genomics cannot be retrofitted after the fact. It has to be built differently from the start.
The architecture of trust
The researchers identify five interlocking requirements. The first is data: representative African genomic datasets, built at scale, without which any model will replicate existing gaps rather than close them.
The second is transparency. The authors argue that AI models applied to African genomic data must be validated openly — their methods legible, their limitations named. Opacity, they write, compounds mistrust in communities that have historical reasons to regard outside research with caution.
The third requirement is governance. Gaye, Bonham, and Mersha are direct: African researchers and institutions must lead, not consult. The distinction matters. Advisory roles do not transfer decision-making power. Leadership does.
Local capacity is the fourth pillar the authors identify. Training, infrastructure, and sustained institutional investment within African research environments are prerequisites, not afterthoughts. Without them, data leaves the continent and so do the benefits.
The fifth requirement is community engagement — and the authors are careful about what they mean. Generic outreach is not sufficient. Engagement must be culturally grounded, attentive to specific histories of mistrust, and designed to return value to participants rather than extract from them.
Why this reaches the albinism community
For people with albinism across Sub-Saharan Africa, the stakes of genomic research are not abstract. Albinism is caused by variants in genes including OCA2 and TYRP1 — variants that occur at notably higher frequencies in certain African populations. Research into those variants affects diagnosis, genetic counselling, and the development of any future treatments.
If the AI systems analysing genomic data are trained predominantly on European ancestry datasets, findings relevant to African populations with albinism will be incomplete or missed entirely. Gaye, Bonham, and Mersha's framework is, in this sense, directly relevant to the community this publication serves.
The researchers describe the current moment as one of both promise and risk. AI has the capacity to accelerate genomic discovery across populations historically excluded from its benefits — but only if the structural conditions they outline are met. Without them, the authors argue, AI is more likely to reinforce global genomic inequities than reduce them.
The paper does not read as a warning against AI. It reads as a precise set of conditions under which AI becomes worth trusting.
Keywords
Core topics and entities mentioned in this summary.