The handling of general-purpose AI in the EU AI Act looks like a mess to me!
In particular, Version 1.1 of the Draft Compromise Amendments from May 16, 2023 introduces the notions of "foundation models" and "general purpose AI". In particular, Article 3 (Definitions) states:
(1c) ‘foundation model’ means an AI model that is trained on broad data at scale, is
designed for generality of output, and can be adapted to a wide range of distinctive
tasks;
(1d) ‘general purpose AI system’ means an AI system that can be used in and adapted to
a wide range of applications for which it was not intentionally and specifically
designed;
And the difference between these two would be... what, exactly?
Section (60e) clarifies slightly by asserting "each foundation model can be reused in countless downstream AI or general purpose AI systems". From this, I gather that the authors would see GPT-4 as a foundation model and ChatGPT (adapted from GPT-4) as a general-purpose AI. But, at a technical level, this feels like a superficial and transitory distinction, analogous to two different user interfaces atop the same database.
Beyond these two definitional mentions of "general purpose AI", I can find only one other passing reference to "general purpose AI" in Article 28 1 (ba), which is a clause noting that a "general purpose AI system" is considered an "AI system". This seems like a redundancy analogous to "This applies to all people, including George."
Also, is a "general purpose AI" derived from a foundation model still itself considered a foundation model? This seems like a critical question, because the act says a lot about foundation models, but (as noted above) almost nothing about general purpose AI. And, if not, rules on foundation models could be evaded by making the superficial transition to a "general purpose AI".
I agree with you. I prefer the use of the term 'general purpose AI system'. We have a new paper out operationalising the definition of general purpose AI system: https://arxiv.org/abs/2306.02889. Hopefully, policymakers will remove this confusion during trilogue negotiations.
I am also perplexed. After reading, I ran a quick global search for “General Purpose AI”, assuming I missed something. What’s more, the amendments update penalties for non-compliant Foundational Models, yet the provider requirements are still tailored for “intended use”.
The handling of general-purpose AI in the EU AI Act looks like a mess to me!
In particular, Version 1.1 of the Draft Compromise Amendments from May 16, 2023 introduces the notions of "foundation models" and "general purpose AI". In particular, Article 3 (Definitions) states:
(1c) ‘foundation model’ means an AI model that is trained on broad data at scale, is
designed for generality of output, and can be adapted to a wide range of distinctive
tasks;
(1d) ‘general purpose AI system’ means an AI system that can be used in and adapted to
a wide range of applications for which it was not intentionally and specifically
designed;
And the difference between these two would be... what, exactly?
Section (60e) clarifies slightly by asserting "each foundation model can be reused in countless downstream AI or general purpose AI systems". From this, I gather that the authors would see GPT-4 as a foundation model and ChatGPT (adapted from GPT-4) as a general-purpose AI. But, at a technical level, this feels like a superficial and transitory distinction, analogous to two different user interfaces atop the same database.
Beyond these two definitional mentions of "general purpose AI", I can find only one other passing reference to "general purpose AI" in Article 28 1 (ba), which is a clause noting that a "general purpose AI system" is considered an "AI system". This seems like a redundancy analogous to "This applies to all people, including George."
Also, is a "general purpose AI" derived from a foundation model still itself considered a foundation model? This seems like a critical question, because the act says a lot about foundation models, but (as noted above) almost nothing about general purpose AI. And, if not, rules on foundation models could be evaded by making the superficial transition to a "general purpose AI".
Totally confusing. At least to me.
I agree with you. I prefer the use of the term 'general purpose AI system'. We have a new paper out operationalising the definition of general purpose AI system: https://arxiv.org/abs/2306.02889. Hopefully, policymakers will remove this confusion during trilogue negotiations.
I am also perplexed. After reading, I ran a quick global search for “General Purpose AI”, assuming I missed something. What’s more, the amendments update penalties for non-compliant Foundational Models, yet the provider requirements are still tailored for “intended use”.