Genetic AI models trained on European data often fail for non-European populations—a problem Google Research is now tackling with surprising findings about when transfer learning helps or hurts. In a new study published on their AI blog, researchers evaluated how to improve disease risk prediction across populations by combining European and Japanese genetic data. The team tested eight clinical traits using over 200,000 Japanese individuals from Biobank Japan and hundreds of thousands of Europeans from UK Biobank. The key finding: while borrowing large European datasets initially boosts accuracy in smaller populations, the benefit reverses dramatically once the target population reaches about 15,000 samples. Beyond that threshold, adding more European data actually worsens predictions for Japanese cohorts—a counterintuitive result that challenges the assumption that more data always improves AI models. The research has immediate implications for healthcare AI fairness. As genomic prediction tools increasingly guide clinical decisions, the ability to accurately predict disease risk across diverse populations is critical. Google's findings provide practitioners with concrete guidelines on how much European data to mix with target population data—and when to rely solely on local genetic information instead. The work underscores both the promise and peril of transfer learning in healthcare AI.