AI Act: The Risk of Being Classified as a Provider When Customising AI Models
The AI Act applies to a wide range of actors involved in the AI value and supply chain. While providers of AI systems and general‑purpose AI models are the primary targets, organisations that merely use such tools must not underestimate the risks associated with their own legal classification under the Regulation.
As a rule, an organisation that limits itself to the professional use of an AI system is classified as a deployer and is subject to more limited obligations than those imposed on providers. However, this classification is neither straightforward nor fixed. It depends closely on the concrete use of the AI tool and may evolve over time.
Determining your organisation’s exact role is therefore crucial, as it directly determines the applicable obligations and, consequently, your level of compliance with the AI Regulation.
Customisation of General‑Purpose AI Models: A Turning Point
Particular attention must be paid to situations in which an organisation modifies or customises (even to a limited extent) a general‑purpose AI model.
In practice, many organisations adapt existing models, such as GPT, to deploy an AI system (for example, a chatbot) internally or to make it available to clients or public users. Beyond the question of the organisation’s likely qualification as a provider of the resulting AI system, a key issue arises: can such model adaptation lead to the organisation being reclassified as a provider of a general‑purpose AI model?
The stakes are high. The obligations imposed on providers of general‑purpose AI models are particularly extensive, including the preparation of technical documentation, the implementation of a policy to comply with EU copyright and related rights legislation, and the publication of a sufficiently detailed summary of the content used to train the model.
European Commission Guidelines: When a Modification Leads to Reclassification
The AI Act does not explicitly resolve the issue of reclassifying an organisation that modifies a general‑purpose AI model. However, the European Commission has provided important clarifications in its guidelines on the scope of obligations applicable to providers of general‑purpose AI models.
According to these guidelines, an organisation becomes the provider of the modified model where the modification results in a significant change to the generality, capabilities or systemic risk of the original model.
By way of indication, the Commission notably relies on a criterion related to the computing power used for the training associated with the modification. This threshold is met when the computing power used exceeds one third of that used for the initial training of the model. Where such information is not disclosed by the original provider and cannot be estimated, the Commission provides for alternative thresholds (which may evolve over time):
- one third of 10²³ floating‑point operations (FLOPs) for general‑purpose AI models not presenting systemic risk;
- one third of 10²⁵ floating‑point operations (FLOPs) for general‑purpose AI models presenting systemic risk.
The Need for Prior Legal Analysis and Ongoing Oversight
The Commission’s guidelines make clear that an organisation modifying a general‑purpose AI model is not automatically classified as the provider of the modified model. Nevertheless, a rigorous legal assessment is essential and must be properly documented, particularly in the event of an inspection by a competent authority or discussions with contractual partners.
Such vigilance must also be maintained over time. Your organisation’s classification may evolve depending on operations performed on the model (for instance, improvements or further training). To avoid falling within the particularly stringent regime applicable to providers of general‑purpose AI models, it is crucial to establish clear and robust internal rules governing the development and use of AI tools.
Our advice:
As part of your compliance with the AI Act, it is essential to give priority attention to determining your organisation’s role and the legal qualification of the AI tools concerned. This involves:
- a rigorous ex‑ante legal analysis;
- clear internal policies governing AI development and use (AI policy, AI governance and training);
- robust contracts with service providers, incorporating AI‑specific clauses.
Our team of experts supports you at every stage of this process, notably through advantageous service packages such as our AI Regulation (AI Act) Compliance Pack and our AI Policy Workshop Pack.
