Copyright limits apply to AI training

Copyright protections continue to apply when online content is used to train AI systems. Rights holders retain exclusive authority over the reproduction, distribution, public communication and adaptation of their work. Lawsuits against OpenAI and Microsoft have brought these limits into focus, including The New York Times’ challenge to the unauthorised use of its archive. The dispute highlights the risk that AI-generated responses can replace website visits or publication purchases, undermining revenue for content owners.

In the European Union, the text and data mining exception balances legitimate access with the ability to expressly reserve rights, rather than making training content freely available. A Munich Regional Court ruling in GEMA v. OpenAI establishes that the exception cannot justify storing protected works and reproducing them in AI outputs. Metadata, opt-out registers and protocols such as ai.txt do not themselves prevent misuse, but can help establish that a system could reasonably identify a reservation of rights.

Works generated entirely by AI raise a separate ownership problem because AI cannot hold rights, while attribution to a natural person remains unclear. Consent, traceability and recognition of the value of creative work are central principles for addressing these legal questions.