Compounding Neglect: How AI Widens the Equity Gap in Psychiatric Genomics, and How We Close It¶
Manuel Corpas, IDEA plenary keynote, 2026 World Congress of Psychiatric Genetics (ISPG), Glasgow, 1 October 2026.
Everything an AI system knows, it learned from a research record that represents some people far better than others. This talk follows that neglect as it compounds, from which diseases get studied and whose genomes wrote the dosing rules, to the literature at the edges of knowledge and the AI models and agents that inherit all of it. It then shows what our pharmacogenomics benchmark found when the answer key itself was wrong, and closes on three things to do next.
Three ideas to remember¶
- Neglect compounds, and AI inherits it. Africa carries 24% of the world's disease burden; its six biobanks among 70 produced 0.7% of the papers. Among diseases with the same amount of literature, those that weigh most on Africa are served less well by language models.
- Determinism is not truth. Expert-written rules raised agreement with our reference from 55% to 63% to 97%. But when our answer key was wrong for one RYR1 case, every rule-following configuration reproduced the error and was scored right, while every model that read the guideline was scored wrong.
- The safest part of the system reaches least far where knowledge is thinnest. Our expert rules cover 16% to 22% of the distinct pharmacogenomic states found in a Spanish family, Iberian samples and Peru, and 6.6% in Uganda.
Three things to do next¶
- Measure the silence. Count who receives no answer at all, population by population.
- Audit the answer key. Ask whose truth a 97% was measured against.
- Write the missing rules, by and for the populations the current rules miss.
About this recording¶
The video is a full recorded run-through made on the day of the talk, not the live session. It runs 38 minutes 50 seconds; the deck has 19 slides. Corrections: at 15:07 the 13 million figure is PubMed search results, which count an abstract once per matching disease; at 25:35, 97% is agreement with our reference answers, and the spread between models narrowed from 21 points to 6, not to zero; at 28:38 the Peruvian cohort is 736 people; at 30:11 the framing test used European, Latin American and East African descriptions.
Cohort differences on real genomes are not attributed to ancestry: the cohorts differ in size, in the genetic states they contain and in assay.
Papers¶
- Corpas, Guio and Fatumo. Agentic genomics: from pipeline automation to autonomous validation. Cell Genomics, 2026.
- Corpas, Iacoangeli, Bourdenx et al. Pharmacogenomics benchmark, Cell Genomics (accepted). Code and data.
Read the documentation | Follow the events calendar | Join the ClawBio Slack channel