Culture Action Europe opposes 3rd draft of the EU General-Purpose AI Code of Practice

On 21 March, Culture Action Europe’s Policy Adviser Luiza Moroz participated in the meeting of the AI Office’s Copyright Working Group, which discussed the third draft of the EU General-Purpose AI Code of Practice.

The Code of Practice is intended to guide AI providers in complying with the AI Act. However, the third draft completely overlooks the rights of rightsholders and creates leeway to infringe copyright. Culture Action Europe publicly opposed this version.

👉🏻 Here’s why:

🔹 Creators who want to opt out of having their data used for AI training must employ robots.txt protocols on their websites. Alternative opt-out methods—such as metadata tags, website terms and conditions, public declarations, rights reservation repositories, etc.—are neither detailed nor effectively recognised in the Code because they ‘are not standardised’. (The whole point of the Code is to provide best practices that can be used until standards are developed.)

🔹 Moreover, if rightsholders employ other machine-readable opt-outs beyond robots.txt, AI providers are only encouraged to ‘make best efforts’ to comply with them. Alternative opt-outs appear valid only if rightsholders agree on them through additional EU-level discussions. This unnecessary conditionality weakens existing obligations and delays enforcement: the Copyright Directive explicitly requires to respect machine-readable means of rights reservation.

🔹 Another example where the Code waters down clear obligations is when AI providers are requested to make ‘reasonable efforts’ to avoid scraping piracy websites—websites which are already illegal! Similarly weak phrasing appears in measures ‘encouraging’ AI providers to: 1) publish copyright policies; 2) provide rightsholders with information about web crawlers and robots.txt features; 3) respect opt-outs without negatively affecting content findability; and 4) mitigate risks of models memorising copyrighted content.

🔹 Currently, the only remedy for rightsholders whose opt-outs are ignored is lengthy and costly legal action. The Code introduces complaint forms for reporting copyright infringements, but the current draft leaves resolution entirely to AI providers’ discretion. It explicitly allows them to dismiss complaints they consider ‘repetitive’ or ‘excessive’.

The current version creates more legal uncertainty than clarity and should not be adopted as it stands. We call for attention to this matter at the highest political level and expect the Code to clearly detail obligations for tech companies, not simply encourage them to respect copyright.