The AI Act's extraterritorial reach - Europe’s gamble on the Brussels effect

22.09.2026
Graphic

When arguing that the growing burden of EU legislation on our domestic ecosystem is slowing down European innovation and productivity growth, we are often faced with the argument that we will force the rest of the world to adopt the same policies, thereby leveling the playing field. The idea behind this is that our market is so lucrative, that all vendors will pay the necessary price (by adapting to our rules) to be able to access it. The former commissioner Thierry Breton was a leading champion of this approach, pioneered in the GDPR.

While attractive at first glance, this theory doesn’t hold up to empirical scrutiny. Europe’s legislative frenzy has mainly hamstrung our own industries, forced many of our innovators to relocate to more favourable jurisdictions and is even causing our own financial resources to be invested elsewhere, fueling growth of other nations. The Draghi report leaves no doubt here. 

And yet once again this line of lazy and destructive thinking has seemingly prevailed in the last days of work on the AI Act and found its way into Recital 106, which concerns the issues of copyright policies mandated by Art. 53 AIA. The theory here goes, that we don’t need to worry about how our copyright legal system affects our ability to create and implement AI systems because we will force all foreign providers to follow our rules if they want to place their products on the European market. The proponents of this theory claim that every general purpose AI model offered in Europe must obey our rules on access to copyrighted data in the training stage.

This is the argument that is carefully examined and falsified in the new study by Dr. Ole Jani and Dr. Maximilian Vonthien of CMS: “Does Recital 106 of the AI Act Violate the Principle of Territoriality in International Copyright Law?”, which we are proud to publish. 

The study, commissioned by the AI Chamber, looks at one question: Can the EU require GPAI providers to follow EU copyright rules when they train their models outside the EU? Recital 106 of the AI Act seems to imply such a position. The study examines in detail whether this assumption is correct, compatible with established principles of international law and practice and explains why the attempt is legally unenforceable.

The study makes four main arguments against accepting the extraterritorial effect of the AI Act, as suggested by recital 106:

  1. A recital cannot create an obligation. Under EU law, recitals help interpret the text. They are not binding provisions. So Recital 106 cannot widen Article 53(1)(c) of the AI Act into a worldwide compliance duty. That point alone is enough.
  2. Copyright is inherently territorial. International copyright law (incl. the Berne Convention and TRIPS) and Rome II Regulation) is built on territoriality. Copyright applies only in the territory where protection is claimed. Recital 106 demands compliance with EU law for training that happens entirely outside the EU. That breaches customary international law, which the EU is bound to respect.
  3. It is not a “market entry requirement.” Defenders compare it to laws like the Deforestation Regulation that ban products from the EU market if they don’t follow certain rules. The comparison is not valid. Product safety rules rest on facts that are true everywhere. Whether an act is “copyright-relevant” is a legal question, and the answer changes from country to country.
  4. Foreign copyright rules are not “lower.” The AI Act assumes other regimes fall short and need to be lifted to EU level. But international copyright sets minimum standards, not a league table. Calling a different rule a weaker rule is a political judgment, presented as a legal one. 

Consequences for the wider EU economy

While the study tackled the legal side of this problem, we encourage you to also consider the implications of the alternative view – what if indeed we decide to try to enforce our rules and the providers of frontier models are forced to withdraw them from the European market? Will that serve the EU economy, which has not been able to keep pace with the rapid development of AI and relies heavily on models trained abroad?

Many in the media industry would welcome this, because it seemingly targets only foreign AI companies. That view is short-sighted and completely misses the bigger economic picture.

Territoriality is what makes many media business models work. Geofencing, staggered release windows, exclusive licences country by country – all of it depends on copyright being geofenced. If the EU succeeds in dismantling that principle to try and outlaw AI training abroad, the damage may not stop there. You cannot expect to upend the rules that underpin your business model and expect there will be no negative consequences.

And this is even before we start examining the impact on the EU economy as a whole – the EU doesn’t yet possess the compute infrastructure to catch up to the frontier in the short term. In one swift stroke we would be relegated to being a technological backwater. 

What happens next

Trying to make EU copyright the global default will fragment the EU’s AI ecosystem rather than level it.

There is a remedy. Because the legal basis is weak, Commission enforcement built on Recital 106 can be challenged. Providers can bring annulment proceedings under Article 263 TFEU, or raise a plea of inapplicability under Article 277 TFEU.

Please check the full study AI Study_ Recital 106 AIA.

·

Graphic
04.10.2023

Main page