Biological Tools and the EU AI Act

The European Union AI Act (EU AI Act) is a regulatory framework that aims to govern the development, deployment, and use of AI systems within the EU. It focuses on large-scale AI systems classified as general-purpose AI (GPAI). It is currently undetermined whether AI-enabled biological tools are subject to regulation under the EU AI Act. In this report, we apply the relevant GPAI and systemic risk definitions in the EU AI Act to 50 AI biological models (both narrow AI-enabled biological tools and biological foundation models) and examine the implications.

Summary Findings

1. 1. We assess that ESM C 6B and ESM3—probably currently the most advanced protein language and generative bioFMs available, respectively—almost certainly¹ meet the generality and downstream integration GPAI criteria and likely have been made available commercially in the Union market, negating the R&D exemption. Therefore, we assess overall it is likely that ESM C 6B and ESM3 are GPAI models. 
   2. We assess that six additional bioFMs (Evo, EvoDiff, Chai-1, xTrimoPGLM-100B, ProtT5-XL-UniRef50 and Nucleotide Transformer) likely meet the generality and downstream integration criteria in the GPAI definition, but only four (Evo, EvoDiff, Chai-1, ProtT5-XL-UniRef50) would be classed as GPAI since the remaining two are likely to fall under the R&D exemption.
   3. Of the non-bioFM models that we considered, we found three examples of experimental design, planning, and simulation tools (AutoOED, Synthace, Benchling for Lab Automation) and two examples of autonomous experimental platforms (Eve, BioAutomata)—where there was a realistic possibility that these models meet the generality and integration criteria for GPAI. These AI systems are commercially available, however, we determined it was unlikely that they met the systemic risk criteria.
   4. On examination of the above potential biological GPAIs, we determined it is a realistic possibility that ESM3 and ESM C 6B meet the high-impact capabilities threshold for protein language and protein generation models, respectively, and therefore pose systemic risk as defined by the EU AI Act.
   5. We think that it is highly likely that some future biological GPAIs will be classified as posing systemic risk given that ESM3 was trained with 10^24 FLOPs, close to the 10^25 FLOPs threshold specified in Chapter 5, Article 51. Future models exceeding this will be automatically considered to possess high-impact capabilities and pose systemic risk.

¹ We use the terminology of the Professional Head of Intelligence Assessment’s Probability Yardstick to specify our level of certainty for our conclusions. This yardstick splits the probability scale into seven ranges: remote chance; highly unlikely; unlikely; realistic possibility; likely; highly likely; almost certain.

For the full report, please click the ‘Download’ button below. If you have any queries about this report, please get in touch with biosecurity@longtermresilience.org