The European Commission recently launched a targeted consultation for feedback on the preparation of guidelines intended to clarify the obligations placed on providers of general-purpose AI models under the AI Act. The consultation, carried out by the AI Office, seeks input from a broad range of stakeholders, including industry representatives, civil society, academia, independent experts, and public authorities. The primary aim is to offer clear guidance on topics such as the definition of "general-purpose AI model", the process of placing these models on the Union market, how to estimate computational resources for training or modifying a model, and supervision and enforcement of the GPAI rules. These clarifications are intended to support thorough compliance with the AI Act.
Central to the consultation is the definition of "general-purpose AI model" which hinges on the question of whether a model displays "significant generality and is capable of competently performing a wide range of distinct tasks". The preliminary view is that training compute is "an imperfect proxy for generality and capabilities but currently the best metric for legal certainty". More specifically, there is a presumption that a model that can generate text and/or images with a training compute of more than 10^22 FLOP will typically qualify as general-purpose. Nonetheless, this can be rebutted by evidence of narrower functionality.
The Commission intends the new guidelines, along with the General-Purpose AI Code of Practice (which will outline commitments for fulfilling obligations regarding documentation, copyright compliance, and - where relevant - systemic risk), to be published by May or June 2025. The consultation remains open until noon (CET) on Thursday 22 May 2025 and the AI Office has invited all interested parties to make submissions.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.