The Case for ESM3 as a General-Purpose AI Model with Systemic Risk Under the EU AI Act

Biological AI models like ESM3 currently fall through regulatory gaps: The paper finds that ESM3, a frontier biological foundation model, is not meaningfully regulated under the EU AI Act despite its potential risks, revealing ambiguity in how the law applies to AI systems designed for biological research.

Dual-use biological risks require specific oversight: The authors argue that ESM3 and similar models pose "dual-use risks"—capabilities that could be misused for harmful biological purposes—and recommend that providers of such models be required to assess and mitigate these risks, similar to obligations for other high-risk AI systems.

Current AI Act criteria don't adequately capture biological foundation models: When comparing ESM3's attributes against the EU AI Act's classification criteria, the paper demonstrates that biological models don't clearly trigger the Act's requirements for general-purpose AI with systemic risk, suggesting the law wasn't designed with this category of AI in mind.

Regulatory remedies are needed to close the gap: The paper proposes specific fixes to extend the EU AI Act's obligations to biological foundation models, indicating that policymakers should clarify and expand the law's scope to address emerging risks from AI systems designed for scientific research.

The biorisk chain connects AI capabilities to potential harms: By mapping ESM3 to the "biorisk chain," the authors establish a concrete pathway showing how biological AI models could contribute to harmful outcomes, making the case that regulation should follow this risk trajectory.