Researchers in China have unveiled a new AI framework that could accelerate the discovery of new medicines. DrugCLIP can scan millions of potential drug compounds against thousands of protein targets in just a few hours—ten million times faster than current virtual screening methods.

Typically, when scientists develop new medicines, they use complex computer simulations to fit a 3D drug molecule into a protein pocket. This indicates that it is likely to interact with the protein’s binding site and function. However, the process is incredibly time-consuming and expensive.

Editorial: Proposals to apply clinician-style licensure to AI tools may allow adaptive oversight as AI models grow more complex. Implementation challenges include defining responsible parties and ensuring adequate regulatory expertise.

In this issue of JAMA Internal Medicine, Bressman et al¹ propose a clever thought experiment: what if medical tools incorporating artificial intelligence (AI) were licensed as advanced practitioners, rather than solely regulated by the US Food and Drug Administration (FDA)? This strategy seeks to provide an alternative or complement to FDA clearance in regulation of medical software incorporating AI. The authors suggest this may allow the necessary flexibility to keep up with the pace of change in AI, the breadth of applications for a given model, and the need to ensure that such tools demonstrate clinical utility.²

Many instances of more specific, single-purpose AI applications can be adequately regulated within existing frameworks. However, generative AI may be deployed in a wide range of contexts, and models may continue to develop over time. Because these models are probabilistic rather than deterministic, they may make errors that are analogous to human errors, for example, mistakes due to inadequate knowledge or lapses in judgment. Bressman et al¹ argue that an appropriately flexible framework for certification already exists in the form of licensing oversight of advanced practitioners. With this approach, the extent of supervision depends on the particular activity, with some tasks requiring more oversight than others.

The proposal leaves a number of critical details to be resolved. Any AI licensing system will need to be able to evaluate and address a model’s specific potentials for harm before deployment; thus, some central regulation likely will continue to be required. In addition, determining who will take on the responsibility and oversight for decisions and treatment pathways generated by AI, as well as assume the liability for errors or adverse events, remains a thorny question. These considerations are again analogous to those of clinician licensing, but although medical boards are well positioned for licensing, the extent to which a similar approach could be developed with the necessary expertise for AI in medicine remains to be seen.