Biological data governance in an age of AI.

Scientists are making rapid progress in developing artificial intelligence (AI) models trained on biological data and in combining these biological models with general-purpose AI reasoning models and AI agents. After the 50th anniversary Asilomar Conference, an international group of more than 100 researchers endorsed an entreaty to prevent the use of these increasingly capable AI systems for "deliberately harmful applications such as bioweapons development" (1). [T.V.I., M.J.I., R.B.A., and J.P. joined that entreaty.] Expanding on the entreaty's call for a system of data access control, we propose that governments, life-science funders, and other organizations impose tailored access restrictions on a narrow set of pathogen data that pose particular AI-related misuse risks. Unfettered access to data has generally served science and society well and should be preserved in almost all domains. Governments should thus regularly reassess the narrow restrictions we propose with the aim of refining them. [ABSTRACT FROM AUTHOR]