Drafting a copyright policy for the EU AI Act? See the new EU AI Act Code of Practice

The final version of the EU AI Act Code of Practice for General Purpose AI Models has been released. There are three chapters, namely (1) Transparency, (2) Copyright and (3) Safety and Security.

The Copyright chapter is of particular interest, not least to general-purpose AI model providers who are required by Article 53(1)(c) of the EU AI Act to put in place a policy to comply with EU law on copyright and related rights, including complying with reservations made by copyright owners around the use of their copyright works. The Copyright chapter will be instrumental in drawing up these policies.

The Copyright chapter contains five principal measures, namely:

1. To draw up, keep up-to-date and implement a copyright policy. The policy should incorporate the measures included in the Copyright chapter.

2. To reproduce and extract only lawfully accessible copyright protected content when crawling the World Wide Web. This requires signatories (i) not to circumvent technological measures designed to restrict unauthorised acts in respective copyright works; and (ii) to exclude from their web crawling websites which are recognised as persistently and repeatedly infringing copyright and related rights on a commercial scale by courts/public authorities in the EU (a dynamic list of these websites is to be made publicly available on an EU website).

3. To identify and comply with rights reservations when crawling the worldwide web. This requires signatories to commit to (i) employing web crawlers that abide by the Robot Exclusion Protocol (presently the most widely deployed rights reservation mechanism, although not necessarily the most effective); and (ii) identifying and complying with other appropriate machine-readable protocols that are used to express rights reservations.

This does not affect the application of EU copyright law to content which is scraped/crawled by third parties and used by the signatories for text and data mining activities when training their general purpose AI models.

Signatories are also encouraged to engage in discussions with copyright holders in order to develop appropriate machine-readable standards and protocols around rights reservations. This will be crucial. There are new more effective rights reservation standards evolving, but no frontrunner at present. Signatories are also required to make public information about the web crawlers employed, their robot.txt features and other measures that they have adopted to identify and comply with rights reservations.

4. To mitigate the risk of copyright infringing outputs. Signatories are required to commit to implementing appropriate and proportionate technical safeguards to prevent their models from generating copyright infringing outputs. They are also required to prohibit copyright infringing uses of their model in their acceptable use policy, terms and conditions, and other equivalent documents.

Where the general-purpose AI model has been released under free and open source licenses, they are required to alert users to the prohibition of copyright infringing uses of the model in the documentation accompanying the model (regardless of whether the signatory has vertically integrated the model into its own AI system or provided to another entity under a contract).

5. To designate an electronic point of contact and enable the lodging of complaints by copyright holders, regarding non-compliance by the signatory with the code.

There is plenty to pick over in the Copyright chapter. It seems likely that the aspects of a copyright policy which involve rights reservations (which are expressly referenced in Article 53 of the EU AI Act) will necessarily be drafted at a fairly high level, at least at this stage. The reservation of rights – whereby a copyright owner indicates to the general-purpose AI model provider that they do not consent to their copyright works being used in training and development for the AI model – is a key part of the legislative ‘compromise’ between copyright owners and tech AI developers, and yet the technology underpinning it is in a state of flux and somewhat uncertain.