Versatile by design and created without an intended scope, general-purpose AI systems have democratised access to and use of AI technologies, even for SMEs. Downstream developers can apply them in a variety of ways with benefits and risks changing depending on application. For this reason, within the AI Act context, adding general-purpose AI as a new class of AI systems might prove problematic and counterintuitive.
In fact, assigning stringent obligations on providers of these systems goes against the objective of pre-determining and clarifying the balance of responsibilities in the AI value chain. Moreover, it could end up in hampering collaboration between enterprises and innovation across Europe. Another unforeseen circumstance is that it would create a compliance gap that could harm consumers if such systems are used in high-risk manners not envisioned by the provider and the provider itself is unaware of the fact due to privacy and security obligations.
Developers are best placed to understand which controls and risk mitigations are most appropriate to their specific use case, and to implement them appropriately, while providers who choose to enable third parties to deploy them in high-risk applications should provide documentation and information about relevant compliance activities that enable deployers to comply with the Act. Ultimately, the AI Act should clarify this in order to ensure a fair and just balance of responsibilities across the value chain.
We will bring together policy makers, SMEs and other stakeholders to start a broader discussion on the proposal of the General-Purpose AI with regard to the French Presidency and beyond.