The medicine section of the Stanford report shows that a model with 111 million parameters outperforms the best previous methods, and that a genomic model with 200 million parameters outperforms a competitor with 40 billion.
The chapter on medicine in the AI Index 2026 report, published by Stanford’s Human-Centered AI Institute, challenges the notion that, in biology, performance automatically scales with model size. MSAPairformer, a protein language model with 111 million parameters, outperforms the best previous methods on benchmark tests. In genomics, a model with 200 million parameters outperforms a competitor eighty times its size.
In clinical practice, the most tangible benefit relates to report writing: doctors using automated report-generation tools report spending up to 83 per cent less time on this task, with a corresponding reduction in burnout. In terms of diagnosis, multi-agent systems achieve 85.5 per cent accuracy on complex cases, compared with 20 per cent for doctors deprived of their usual tools — a comparison whose methodology is worth bearing in mind.
The report also highlights two areas of concern. The FDA authorised 258 AI-based medical devices in 2025, but mainly through fast-track procedures that did not require new clinical trials, raising questions about the rigour of the validation process. And whilst publications on the ethics of medical AI have doubled, they remain focused on governance, neglecting biosafety and equitable access on a global scale. AI-generated summaries also appear in 84 to 92 per cent of health-related Google searches, making them the primary filter for interpreting medical information.
Read the full article on the original source
Go to the source →