AI-enabled biological tools such as ESM3 can jointly reason over protein sequence, structure and function, letting users iteratively design novel proteins far faster than in a traditional lab. This paper asks whether such tools fall under the EU AI Act’s obligations for “general-purpose AI models with systemic risk” (GPAISR), a designation that currently requires a model to first qualify as general-purpose AI (GPAI) at all.
Mapping ESM3’s capabilities to a seven-step biorisk chain, we find that four steps, including design, learning, building, and weaponisation, could plausibly be made easier by access to the model, particularly its freely available open-source variant. Turning to the Act itself, ESM3’s flagship variant exceeds the Commission’s indicative GPAI compute threshold by roughly tenfold, yet the Commission’s non-binding GPAI Guidelines also require a model to generate human language, which ESM3 does not. This leaves ESM3 outside GPAI classification under current guidance, and therefore outside the Article 55 obligations to assess and mitigate dual-use risk, despite biological sequences being explicitly listed as a relevant input/output type in the Act’s own Annex XIII.
We identify two reinforcing sources for this gap: ambiguity in the Act’s statutory text and a narrowing interpretation in the Commission’s Guidelines. We set out four non-exclusive remedies, ranging from a Commission designation under Article 51(1)(b), to revising the GPAI Guidelines, amending the Act itself, and complementary instruments such as export controls or the proposed EU Biotech Act.
Paper Authors: Taro Qureshi, Jacob Griffith, Koen Holtman, Marcel Mir Teijeiro, Ze Shen Chin, and Rokas Gipiškis
This paper was presented at the 3rd Technical AI Safety Conference (TAIS 2026) in Oxford.