💡 Download the complete guide to AI-generated synthetic data!
Go to the ebook
August 9, 2023
5m 19s

AI Act: A Deep Dive into the Future of AI regulation - Part 3: Foundation Models

The EU recently introduced the new concept of foundation models into the draft of the upcoming AI Act. Find out what this means and what are the expectations of lawmakers when it comes to foundational model development.

Transcript

[00:00:00] Welcome back to our MOSTLY AI video series on the draft AI Act.

[00:00:05] In our last video, we discussed the term AI system, and today, we will look at a specific subset of an AI system, namely the so-called foundation model. But first, let's take one step back.

[00:00:20] As is so often the case, lawmakers are in a constant state of trying to catch up with technological developments since it becomes seemingly impossible for legislators to stay ahead of the game.

[00:00:31] In case of the AI Act, at the time the EU Commission had published the first draft of the Act in 2021, hardly anyone in the lawmaking world was aware of large language models and their potential.

[00:00:43] Then ChatGPT came along and changed the entire landscape. It is not surprising that the EU lawmakers felt the need to react, which is why the EU Parliament recently introduced foundation models as a new concept into the draft AI Act.

[00:00:58] Now, as the term suggests, foundation models can be used as a foundation for other downstream AI systems building on such foundation, and very helpful for us, we find an applicable legal definition in the AI Act.

[00:01:14] According to Article 3(1)(1c), 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.

[00:01:31] Recital (60 e) of the AI ACT provides more background information, and states that foundation models are a recent development in which AI models are developed from algorithms designed to optimize for generality and versatility of output.

[00:01:47] Those models are often trained on a broad range of data sources, and large amounts of data to accomplish a wide range of downstream tasks. This includes some tasks for which these models were not specifically developed and trained.

[00:02:02] The recital then goes on to clarify that AI systems with specific intended purpose or general-purpose AI systems can both be an implementation of a foundation model,

[00:02:14] which means that each foundation model can be reused in countless downstream AI or general-purpose AI systems.

[00:02:23] So, why are we dealing with a separative definition of and recitals on foundation models if we already have a general definition of AI systems, which we looked at in our last video?

[00:02:37] There must be some specific obligations for foundation models that do not apply to all AI systems, otherwise, this more specific definition of a foundation model would make no sense, and indeed, there are such specific obligations for foundation models.

[00:02:55] The EU Parliament has introduced the already infamous article 28(b) into the draft AI Act.

[00:03:02] This provision contains a number of obligations for providers of foundation models, some of which will be quite hard to comply with. Just to name a few.

[00:03:11] The provider of a foundation model shall demonstrate the identification, reduction and mitigation of reasonable foreseeable risks to health, safety, fundamental rights, the environment and democracy, and the rule of law.

[00:03:25] Process and incorporate only data sets that are subject to appropriate data governance measures for foundation models, in particular measures to examine the suitability of the data sources and possible biases and appropriate mitigation.

[00:03:39] We'll come back to this point in a future video.

[00:03:42] Design and develop the foundation model in order to achieve, throughout its lifecycle, appropriate levels of performance, predictability, interpretability, corrigibility, safety, and cybersecurity.

[00:03:54] Design and develop the foundation model, making use of applicable standards to reduce energy use, resource use, and waste, as well as to increase energy efficiency and the overall efficiency of the system.

[00:04:07] Draw up extensive technical documentation, and keep the documentation available for 10 years, and establish a quality management system, and register the foundation model in an EU database.

[00:04:22] If that's not enough for you, don't worry, there's more. Providers of foundation models that are to be used in generative AI systems have additional obligations.

[00:04:31] For example, to train, design, and develop the foundation model in such a way as to ensure adequate safeguards against the generation of content in breach of Union law, in line with the generally acknowledged state-of-the-art, and without prejudice, to fundamental rights, including the freedom of expression, and make publicly available a sufficiently detailed summary of the use of training data protected under copyright law.

[00:04:58] Now, if you manage to comply with all these obligations, you will hopefully have a very solid foundation.

[00:05:07] With this information overload, I leave you with an impressive picture of a lawyer suffering from information overload.

[00:05:14] Thanks for watching.

Ready to try synthetic data generation?

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