Rachel Thomas, co-founder of fast.ai and AI researcher with a Masters in Microbiology-Immunology, discusses the risks and limitations of AI in life sciences. Key concerns include overreliance on existing data rather than investing in new assays and biomarkers, the difficulty of catching errors without deep domain expertise (illustrated by a flawed enzyme classification paper in Nature Communications), self-reinforcing feedback loops from incorrect predictions entering training data, and the misleading authority of scale when data collection is fundamentally flawed. She highlights the Zoe COVID app as a case where scaling bad data design made things worse, and argues for slower, more rigorous collaboration between domain experts, patients, and ML practitioners throughout the entire pipeline. She also advocates for continued investment in bench science and causal mechanism research, since current AI systems interpolate within existing data rather than discovering truly novel paradigms.
