General purpose AI models and measures in support of innovation

General purpose AI models

(Chapter V  V, Art. 51-56)

The AI Act is founded on a risk based approach. This regulation, intended to be durable, initially wasn’t associated to the characteristics of any particular model or system, but to the risk associated with its intended use. This was the approach when the proposal of the AI Act was drafted and adopted by the European Commission on 22 April, 2021, when the proposal was discussed at the  Council of the European Union on 6 December, 2022.

However, after the great global and historical success of generative AI tools in the months following the Commission’s proposal, the idea of regulating AI focusing only on its intended use seemed then insufficient. Then, in the 14 June 2023 draft, the concept of “foundation models” (much broader than generative AI) was introduced with associated regulation. During the negotiations in December 2023, some additional proposals were introduced regarding “very capable foundation models” and “general purpose AI systems built on foundation models and used at scale”.

In the final version of the AI Act, there is no reference to “foundation models”, and instead the concept of “general purpose AI models and systems” was adopted. General Purpose AI models (Arts. 51 to 56 ) are distinguished from general purpose AI systems (Arts. 25 and 75). The General Purpose AI systems are based on General Purpose AI models: “when a general purpose AI model is integrated into or forms part of an AI system, this system should be considered a general purpose AI system” if it has the capability to serve a variety of purposes (Recital 100). And, of course, General Purpose AI models are the result of the operation of AI systems that created them.

“General purpose AI model” is defined in Article 3 (63) as “an AI model (…) that displays significant generality and is capable to competently perform a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications”. The definition lacks quality (a model is “general purpose” if it “displays generality” 1 Recital 98 contributes to clarify the concept saying that “generality” means the use of at least a billion of parameters, when the training of the model uses “a large amount of data using self-supervision at scale”. ) and has a remarkable capacity for expansion. Large generative AI models are an example of General Purpose AI models (Recital 99).

The obligations imposed to providers of General Purpose AI models are limited, provided that they don’t have systemic risk. Such obligations include (Art. 53 (I) (a)  (i) to draw up and keep up-to-date technical documentation (as described in Annex XI) available to the national competent authorities, as well as to providers of AI systems who intend to integrate the General Purpose AI system in their AI systems, and (ii) to take some measures in order to respect EU copyright legislation, namely to put in place a policy to identify reservations of rights and to make publicly available a sufficiently detailed summary about the content used. Furthermore, they should have an authorised representative in the EU (Art. 54).

The most important obligations are imposed in Article 55 to providers of General Purpose AI models with systemic risk. The definition of AI models with systemic risk is established in Article 55 in too broad and unsatisfactory terms: “high impact capabilities”. Fortunately, there is a presumption in Article 55.2 that helps: “when the cumulative amount of compute used for its training measured in floating point operations (FLOPs) is greater than 10^25”.

The main additional obligations imposed to General Purpose AI models with systemic risks are (i) to perform model evaluation (including adversarial testing), (ii) to assess and mitigate systemic risks at EU level, (iii), to document and report serious incidents and corrective measures, and (iv) to ensure an adequate level of cybersecurity.

Finally, an “AI system” is “an AI system which is based on a General Purpose AI model, that has the capacity to serve a variety of purposes” (Art. 3 (66)). If General Purpose AI systems can be used directly by deployers for at least one purpose that is classified as high-risk (Art. 75), an evaluation of compliance will need to be done, if there is sufficient reason to consider that the system is not compliant with AI Act.