๐Ÿ’ก Download the complete guide to AI-generated synthetic data!
Go to the ebook
August 22, 2023
7m 39s

AI Act: A Deep Dive into the Future of AI regulation - General Purpose AI

Trascript

๐Ÿ“บ Welcome to our video series on the Draft EU AI Act!

In this episode, we delve deeper into the world of AI definitions, focusing on the intriguing concept of "General Purpose AI Systems." If you've been following along, you already know that we've covered the broad legal definition of AI systems and even explored the specifics of foundation models. But hold onto your seats, because we're about to unlock the secrets of this new subset.

๐Ÿ” What exactly is a general purpose AI system? The definition might remind you of foundation models, but there are nuances that set them apart. We're here to break down those differences for you.

๐Ÿ”„ What is the difference between foundation models and general purpose AI system?

๐Ÿ” What's the significance of these definitions?

๐Ÿ’ผ Are general purpose AI systems also high-risk AI systems?

The future of AI legislation is still in motion, and we're here to keep you in the loop. Join us next time as we dive into the world of "Generative AI."

Transcript

[00:00:01] -Welcome again to our video series on the Draft EU AI Act.

[00:00:06] So far, we have already dealt with the general legal definition of an AI system and looked at foundation models as specific parts of AI systems with their own legal definition, but why stop there?

[00:00:21] For reasons of clarity and comprehensibility, some might argue, well, no. In the wake of the global rise of ChatGPT, the EU Council has included yet another definition for a subset of AI systems.

[00:00:35] Let me introduce general purpose AI systems, and you already know the procedure.

[00:00:42] Let's start with the legal definition of general purpose AI systems, which we find in Article 3(1d) of the Draft AI Act. I just want to highlight that this is not the initial definition as established by the Council but the one later introduced by the Parliament.

[00:00:59] Again, it is important to highlight that this definition is not yet set in stone, as the trialogue is still ongoing, but we'll have a look at it anyways since it is the latest proposal.

[00:01:10] 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.

[00:01:25] Does that remind you of something? Well, I was somewhat reminded of the definition of foundation models which we touched upon in our last video.

[00:01:34] Foundation model means an AI system 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.

[00:01:46] Okay, if we look close enough, there are of course differences and there should be, because the EU legislator wants to set strict rules for foundation models that only apply to such foundation models, but honestly, we have to dissect this quite carefully to differentiate between the two definitions.

[00:02:04] I mean, first of all, and quite importantly,

[00:02:06] Foundation models are AI system models, whereas general purpose AI systems are not models, but well, systems.

[00:02:16] If we look for a legal definition helping us to differentiate between those two terms, systems and models, we are looking in vain. The EU legislator seems to assume that we know the difference anyways. Well, for our purposes, I'd say that a model is an integral part of an AI system. It's kind of the core of the system. The term system therefore is broader.

[00:02:41] In this context, we can recall Recital (60 e) of the Draft AI Act, which states that an AI system with specific intended purpose or general purpose AI systems can be an implementation of a foundation model, so a foundation model can be the basis for general purpose AI systems. So far, so good.

[00:03:04] What other differences can we find between the two legal definitions of general purpose AI and foundation models? Well, the foundation model definition focuses on training data, while input data is not mentioned at all in the definition of general purpose AI.

[00:03:23] Also, the foundation model definition refers to generality of output, while the generality with general purpose AI lies with, well, the general purpose... Hm.

[00:03:37] Last, with the foundation model, we have the adaptability to a wide range of distinctive tasks, while general purpose AI can be used in and adapted to a wide range of non-intended applications.

[00:03:52] All right, or actually it is somewhat confusing, but okay, let's just proceed on the understanding that there is a difference between the two concepts, general purpose AI system and foundation models. After all, a model is only a part of a system.

[00:04:10] Let's move to the next point. We do have a general legal definition of an AI system which encompasses all kinds of AI systems.

[00:04:22] If we also introduce legal definitions for different parts of such AI systems,

[00:04:29,190] for example, models and for different types of systems such as general purpose AI systems, we also need specific rights or obligations for such specific subsets of AI systems. Otherwise, such subclassifications would be useless from a legal perspective. Why would I define a specific category if there are no legal implications only for this category, right?

[00:04:56,250] Last time we saw that Article 28b includes a list of quite onerous obligations for providers of foundation models, so of course it makes sense to define this part of AI system separately. What about general purpose AI? Well, to find a list of obligations for general purpose AI, we need to go back to the council's draft of the AI Act which was published in late 2022. There we find Article 4b entitled Requirements for General Purpose AI Systems and Obligations for Providers of Such Systems.

[00:05:34,169] At the risk of getting ahead of myself, it is important to note at this stage that general purpose AI systems are not automatically considered high-risk AI systems, and we'll talk about the AI Act's risk-based approach in another video, but in short, under the AI Act, all AI systems are categorized in risk categories. The higher the risk the AI system poses, the higher the obligations. If there is only a limited risk, there are only limited obligations. Now, the Article 4b requirements only apply to general purpose AI systems which may be used as high-risk AI systems or as components of high-risk AI systems.

[00:06:18,630] Essentially it says that if general purpose AI is considered high risk, it has to comply with the obligations that all high-risk AI systems have to comply with. This kind of seems to be the point where we enter into circular reasoning territory, but the council's idea was that the EU Commission should continuously implement further specific requirements.

[00:06:39] for general purpose AI via so-called implementing acts,

[00:06:44] but to be honest, I am not really sure whether the definition

[00:06:47] of general purpose AI really adds anything substantial

[00:06:51] to the Draft AI Act at the current stage,

[00:06:54] so we will stop at this point.

[00:06:56] Just remember that general purpose AI is a topic in the AI Act negotiations.

[00:07:03] I'm really looking forward to find out what the final act will state

[00:07:07] in this regard.

[00:07:09] Next time, we will look at yet another subcategory of AI systems,

[00:07:14] generative AI.

[00:07:16] You guessed correctly, generative AI can be

[00:07:19] based on a foundation model and also be general purpose AI.

[00:07:23] Well, if that's not a teaser, I don't know what is.

[00:07:26] For now, I leave you with a tantalizing picture

[00:07:29] of a tired lawyer on a summer break.

[00:07:32] Thanks for watching.

Ready to try synthetic data?

The best way to learn about synthetic data is to experiment with synthetic data generation. Try it for free or get in touch with our sales team for a demo.
magnifiercross