The end is in sight! After delays at (almost) every stage of the legislative process, the EU AI Act will soon be published in the EU’s official journal and eventually enter into force in August. This marks the end of more than 3 years of negotiations since the text was originally proposed by the European Commission in April of 2021.
Naturally, with so much time spent on the file, the text has seen a lot of changes, many of which were inspired by technological developments happening in parallel to the institutional negotiations. Most notable in this regard is perhaps the meteoric rise of generative AI (GenAI) technologies. Whereas in 2021, “ChatGPT” and “large language model” would have been foreign concepts to most of us, they are now nearly omnipresent. As such, it was only a matter of time before GenAI made its way into the AI Act, raising countless questions for legislators to address along the way.
One of the more contested issues arising with the inclusion of GenAI was its interaction with copyright. Very basically, the main questions legislators needed to address were:
- Can copyrighted materials be used to train GenAI?
- What is needed to strengthen the current copyright framework in light of AI?
As to the first question, the AI Act clearly establishes that the opt-out mechanism in Article 4 of the DSM Directive is fit for purpose, meaning that copyrighted materials can be collected with text and data mining (TDM) tools and used for the training of GenAI technologies, under the general purpose TDM exception, unless the right holder explicitly opts out of this process.
Finding an answer to the second question proved to be more difficult, to the point that it almost derailed negotiations right before the finish line.
In the end, legislators agreed that in order to rely on the opt-out mechanism, it had to be effective and trustworthy. In short, right holders need to be certain that their indication of opting out of TDM is respected. To that end, Article 53 of the AI Act requires providers of GenAI models to introduce opt-out compliance policies and introduces transparency provisions for the training data.
If implemented well, these measures will go a long way towards making the existing copyright framework fit for dealing with AI technologies. But while the entry into force of the AI Act marks an important step in this direction, the work on this issue is far from done: Codes of practice will need to be drawn up by the AI Office, particularly on compliance policies; the AI Office will need to provide a template for the implementation of the training data transparency requirement; and the Commission will need to continue to gather knowledge and hold stakeholder discussions on the technical implementation of opt-outs with a view to standardize a small number of identifiers for rights reservation.
All this to say: While the AI Act has been three years in the making, this will certainly not be the last time we discuss the legislation of AI technologies in Europe.
