General Purpose AI and the EU AI Act
Generative AI or general purpose AI as it is referred to in the AI Act is everywhere right now thanks to the release of incredible cutting edge models like Meta’s Llama, OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini.
Generative AI is broken out into two distinct elements in the AI Act, general purpose AI models and general-purpose AI systems. The intention behind the distinction is to ensure appropriate and tailored regulation. The handful of large international organisations that have the capital and resources to create, train, and manage general-purpose AI models (like Gemini, Claude, Llama, GPT-4 etc) will be regulated in a very different manner to those building AI applications on those general-purpose AI models (general-purpose AI systems).
While the providers of general-purpose AI models will be subject to detailed and complex technical and transparency measures, those building related AI applications like chatbots will be subject to a lighter transparency and disclosures regime.
Those who fine tune or modify third party general-purpose AI models will likely also have a compliance obligation. In this category will be those application providers who fine tune or modify third party models like Claude, Gemini and Llama, for the purpose of creating and providing their own general purpose AI systems to the EU market. The scope of what level of fine tuning or modification will require compliance with specific obligations has not yet been decided, but we should learn more in Q2 2025. To compound matters here, we do not yet know the scope of obligations of these general-purpose AI systems. They are expected to be a limited version of those that apply to the general-purpose AI model providers.
For the moment the new AI model regulator, the AI Office, is focused on agreeing with model providers, industry and academia, the rules that will govern the obligations of the providers of general-purpose AI models (known as the Code of Practice for general purpose AI models). The thresholds for what constitute model modification and fine tuning and the obligations of the organisations that create general purpose AI systems based on those activities are expected to emerge a little later (Q2 2025). Until then a lot of the rules and procedures regarding general purpose AI models and general-purpose AI systems will continue to remain unclear.
For more information on the EU AI Act or its impacts, contact a member of our dedicated Artificial Intelligence team.
The content of these articles are provided for information purposes only and does not constitute legal or other advice.