European Union lawmakers, engaged in over 20 hours of negotiations, have reached a preliminary accord on regulating artificial intelligence (AI). A leaked proposal, reviewed by TechCrunch, indicates progress on handling a crucial aspect: rules for foundational models or general-purpose AIs (GPAIs).
While there has been a strong push, led by the French AI startup Mistral, for a complete regulatory carve-out for foundational models/GPAIs, the EU lawmakers seem to have resisted this approach. The proposal maintains elements of the tiered approach suggested by the parliament earlier in the year, rather than a total exemption.
However, there is a partial carve-out for GPAI systems provided under free and open-source licenses, excluding certain high-risk models. This exception is limited by commercial deployment, meaning the carve-out no longer applies if the open-source model enters the market or is put into service.
The agreement retains the classification of GPAIs with “systemic risk,” with criteria for this designation including “high impact capabilities” and a cumulative compute usage exceeding 10^25 floating-point operations (FLOPs). This level appears to exclude most current models from upfront obligations to assess and mitigate systemic risks, softening the regulatory impact advocated by Mistral.
For GPAIs with systemic risk, providers have additional obligations, such as standardized evaluation protocols, reporting incidents promptly, adversarial testing, ensuring cybersecurity, and disclosing energy consumption.
General obligations for GPAI providers include testing and evaluating models, maintaining technical documentation for regulatory authorities, and providing AI app makers with model overviews. Foundational model makers must also implement a policy respecting EU copyright law and disclose a detailed summary of training data used to build the model.
The proposal references codes of practice that GPAIs may rely on for compliance until a harmonized standard is published, with the AI Office playing a role in developing these codes. The Commission is expected to issue standardization requests starting six months after the regulation enters into force, with regular progress reporting.
Despite ongoing talks to resolve contested elements, today’s trilogue on the AI Act may determine the fate of the regulation. The EU’s internal market commissioner, Thierry Breton, confirmed that talks have broken up for the day, with the trilogue set to resume tomorrow at 9 am Brussels time, aiming to secure the AI rulebook proposed in April 2021. The outcome remains uncertain, given the high stakes and sensitive nature of the file.
