European Commission Approved Guidelines on the Scope of Obligations for Providers of General-Purpose AI (GPAI) Models under the AI Act

The Guidelines clarifies thresholds and terminologies regarding GPAI. Such clarifications relate more specifically e.g., 

• to GPAI models
• modifications to GPAI models
• GPAI Providers
• placing on the market
• applicability of the qualification "GPAI with systemic risk”
• estimate the computational resources used for training a GPAI model

In respect of GPAI models, the European Commission states that “An AI model is considered to be a general-purpose AI model if it was trained using an amount of computational resources (‘compute’) that exceeds 10^23 floating point operations and if it can generate language (whether in the form of text or audio), text-to-image or text-to-video.”

In respect of modifications to GPAI models, the Guidelines clarify matteres about the lifecycle of a GPAI model. Such clarification is important because the applicability of obligations under the AI Act have a different timeline depending on whether a GPAI model is already existing prior Aug 2nd, 2025, or will be a “new model” to the market as of Aug 2nd, 2025. Alongside further details and exceptions, the Guidelines mention in para 23 "the Commission understands the notion of ‘model’, and consequently its ‘lifecycle’, in a broad sense. In practice, the Commission considers the lifecycle of a general-purpose AI model to begin at the start of the large pre-training run. Any subsequent development of the model downstream of this large pre-training run performed by the provider or on behalf of the provider, whether before or after the model has been placed on the market, forms part of the same model’s lifecycle rather than giving rise to new models".

In respect of GPAI Providers, the Guidelines provide both, a generic methodology approach and distinct examples when the European Commission deems the methodology resulting being a GPAI Provider. Such examples are, pursuant para 51 of the Guidelines

• If actor A develops a general-purpose AI model and places it on the market, then actor A is the provider.
• If actor A has a general-purpose AI model developed on its behalf by actor B and actor A places that model on the market, then actor A is the provider.
• If actor A develops a general-purpose AI model and uploads it to an online repository hosted by actor C, then actor A is the provider.  
  If a collaborative or consortium has a general-purpose AI model developed for it by different individuals or organisations and places the model on the market, then usually the coordinator of the collaborative or the consortium is the provider. Alternatively, the collaborative or the consortium might be the provider. This must be assessed on a case-by-case basis. 

In respect of placing on the market, the Guidelines provide both, a generic methodology approach, including a reference to the Blue Guide, and distinct examples when the European Commission deems the methodology resulting in a placing on the market. Such examples are, pursuant para 53 of the Guidelines

• a general-purpose AI model is made available for the first time on the Union market via a software library or package;
• a general-purpose AI model is made available for the first time on the Union market via an application programming interface (API);
• a general-purpose AI model is uploaded for the first time to a public catalogue, hub, or repository for direct download on the Union market;
• a general-purpose AI model is made available for the first time on the Union market as a physical copy;
• a general-purpose AI model is made available for the first time on the Union market via a cloud computing service;
• a general-purpose AI model is made available for the first time on the Union market by being copied onto a customer’s own infrastructure;
• a general-purpose AI model is integrated into a chatbot made available for the first time on the Union market via a web interface;
• a general-purpose AI model is integrated into a mobile application made available for the first time on the Union market via app stores;
• a general-purpose AI model is used for internal processes that are essential for providing a product or service to third parties or that affect the rights of 
  natural persons in the Union. 

The Guidelines also include provisions addressing GPAI models integrated into AI systems and downstream modifiers of GPAI models affecting both the qualification as GPAI Providers and the understanding of placing on the market. 

In respect of the applicability of the qualification “GPAI with systemic risk”, the Guidelines clarify that either of two conditions must be met. The GPAI model must either 

• have ‘high-impact capabilities’, namely ‘capabilities that match or exceed those recorded in the most advanced models’ (Article 3(64) AI Act);  or
• ‘based on a decision of the Commission, ex officio or following a qualified alert from the scientific panel, have capabilities or an impact equivalent to those set out in [the preceding point] having regard to the criteria set out in 
  Annex XIII’

The methodology to assess evaluate high-impact capabilities are to be determined by delegated acts; such delegated acts also enables the European Commission to adapt thresholds of the AI Act (Art. 51 AI Act), if necessary. In any case, for the time being, a GPAI model is presumed to have high-impact capabilities when the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25, Art. 51.2 AI Act.

In respect of estimate the computational resources used for training a GPAI model, the Guidelines include an entire Annex detailing the European Commission's understanding of methodology and threshold for different approach, such as hardware-based approach or architecture-based approach.

Additionally, the Guidelines clarify on the European Commission's understand of duly notify about GPAI models and / or the due provision of arguments, why a GPAI model might be exempt from obligations under the AI Act as well as to the designation mechanism applicable to GPAI with systemic risk. 

The European Commission highlights that the obligations for GPAI enter into application by Aug 2nd, 2025.

The European Commission confirms the understanding, that any Code of Practice is deemed a temporary tool, only (read also European Commission Received Final Version of the Code of Practice for General Purpose AI), para 100 of the Guidelines. Likewise the European Commission repeats its understanding, that any Code of Practice assess as adequate by the AI Office and the AI Board will provide beneficiary legal effects under the AI Act. The European Commission does not consider its implementing act resulting in general validity a pre-requisite for such legal benefits, see para 94, 99 of the Guidelines. 

The CoP for GPAI will only provide its positive legal effect to the extent it will be implemented. If GPAI providers / models opted-out from entire chapters (or single provisions), the related benefit vanishes. In other words, the European Commission confirms that it will not be necessary to implement the voluntary tool of the CoP for GPAI in its entirety, while there is a notion that opt-in / opt-out applies to the level of entire chapters. 

GPAI Providers which do not adhere to the CoP for GPAI may demonstrate their compliance by other adequate means. The European Commission subnotionably creates the impression that those providers may face higher administrative burdens than those which do adhere to the CoP for GPAI.