It was the the first and the most difficult step towards its definitive adoption. Here's what you need to know about it.

The AI-act has been approved by the COREPER

What is it exactly?

The AI Act is the world’s first comprehensive AI law. Its aim is to define different AI systems with different levels of risk and establish obligations for users and providers of those systems according to those levels. It’s not aimed at changing already existing EU laws (such as those regulating copyright and privacy rights), but it clarifies how those laws apply to AI systems and their applications.

What’s in there?

The Act is very long – over 800 pages – but when it comes to generative AI the more relevant parts are Recital 60 and Article 52.

In European laws, the recitals are those parts of an act that explain the reasons for it. Recital 60 defines the notion of “General purpose models”, which includes “generative models capable of generating text, images, and other content”. It states that “any use of copyright protected content requires the authorization of the rightholder concerned unless relevant copyright exceptions and limitations apply” (more on those later).

It also addresses the fact that these models might pose systemic risks which include “major accidents, disruptions of critical sectors and serious consequences to public health and safety; negative effects on democratic processes, public and economic security: the dissemination of illegal, false or discriminatory content”.

Article 52 defines the transparency obligations for providers and users of General-purpose models.

These includes labelling obligations, requiring the disclosure of any use of AI to generate content (even when the content is “artistic, creative, satirical and fictional”, a kind of content that was exempted from labelling in previous drafts of the act).

Furthermore, the article requires gen AI providers to “put in place a policy to respect Union copyright law” and “draw up and make publicly available a sufficiently detailed summary about the content used for training of the model”. These requirements apply to open license models as well, which were exempted in earlier drafts of the text.

The text is not perfect, let’s see what are the parts that we like and the parts that we don’t.

The good parts

As stated before, at the beginning of 2023, when we started working, the AI Act featured no reference to copyright, no mention of generative AI and no obligation for the providers of those systems. Back in January we were told that the best thing we could have gotten out of it was a passing reference to the need of AI companies for data to train their model, which would have been a way to formally addressing the existence of a “chain of production” of AI models that could have been regulated by a new text of law in a vague future. Things have changed with the rise in popularity of Chat-GPT and the tireless work of activists denouncing the violation of their rights and the damages to their work and lives.

The importance of the lines in the Recital stating the need for an authorization to use copyrighted content for training shouldn’t be underestimated.

Their legal basis is the 790/2019 Directive (TDSM Directive), which describes how copyright applies to text and data mining practices. Both members of the Parliament and the EU Commission have been consistent throughout the year in indicating that copyright applies to AI training and that some of the rules for it were to be found in this Directive (sometimes going as far as saying that the way AI companies have been acting has been not compliant with these rules). That said, having it written clearly in a publicly available text so important for AI regulation worldwide is an important first step. The transparency requirements are relevant as well: too often AI companies hide their datasets, and this kind of disclosure might help future lawsuits and interventions from national authorities. The fact that the summary required isn’t limited to the copyrighted content used anymore is also relevant, as it might give citizens a way to find out if other rights not related to authorship have been infringed (think of privacy rights).

Our work isn’t over yet and there’s much to be done:

We have to keep watching over the drafting of the official AI Act text, as we know that some people are not happy with it and will try to erase everything we have accomplished. After that we will have to work closely with those Institutions that will have to enforce what’s written in the Act. We will also need to fight opt-out as a default system and prove that it shouldn’t apply to Gen AI, while also studying ways for the European creative community to protect itself (expect a comprehensive downloadable guide in the near future!).

All these things will come in the next weeks and months.

As for now, we take a moment to take a deep breath, rest and celebrate what we have achieved.