Defining Hazardous Capabilities of Biological AI Models

Artificial intelligence (AI) models are used increasingly in biological research. Many AI models contain some degree of biological knowledge. Developers are also creating AI models that are able to carry out an array of increasingly sophisticated tasks, either under human guidance or autonomously, including by vesting some models with biological capabilities. Identifying which biological AI capabilities pose significant potential risk to global public health and which do not is an essential step toward ensuring responsible AI development. Both model developers and policymakers have been attentive to the possibility of AI models causing public harm, but further specificity regarding how these harms may arise in the realm of biology is needed. 

On June 3, 2024, the Johns Hopkins Center for Health Security and RAND convened a group of scientists, AI-model developers, biosecurity experts, and policymakers to discuss the potentially hazardous capabilities of biological AI models—models that are trained on or capable of meaningfully manipulating substantial quantities of biological data. The interdisciplinary group clarified the definitions of concerning biological capabilities that would raise significant public health concerns, acknowledging that one of the capabilities that was identified as concerning already exists. Some capabilities have not yet been achieved but are under active development, while others may be years away or may never materialize. Accurately forecasting the direction of AI advances was not considered a prerequisite to initiating this discussion on security. Instead, the aim was to determine specific model capabilities that, if achieved, would pose a serious threat to global public health.