More and more companies are using generative AI to produce all kinds of content. This may include videos featuring non-existent people acting as satisfied customers, images of products made to look slightly more appealing, or complete campaign texts. This raises a number of questions: Do we own the rights to this content? Are AI companies allowed to use this content as training data? And what do we actually need to disclose when we publish AI-generated content? These are no longer theoretical questions. They come up daily in advisory practice, yet the answers are less straightforward than many people assume.
During our webinar of 13 October, we examined these questions along three axes: the input of AI models, the output, and the transparency obligations under the AI Act. Below are the key takeaways from this webinar
May AI be trained on existing works?
Training an AI model requires reproductions of existing material. This directly engages copyright law: reproducing works without permission is in principle an infringement, including for AI companies. For certain types of acts, however, the Copyright Act provides a closed system of exceptions, including the text and data mining exception (TDM exception). The TDM exception derives from the DSM Directive and has been implemented in Articles 15n and 15o of the Dutch Copyright Act.
The exception under Article 4 also allows commercial parties to apply TDM, provided there is lawful access and the rightholder has not exercised an "opt-out". This opt-out must be machine-readable. What exactly this entails remains subject to debate. The Oberlandesgericht Hamburg held in 2025 that a reservation expressed in natural language in terms of use in 2021 did not qualify as machine-readable, while applying a dynamic standard: what was insufficient at the time may now be adequate given the current state of technology. The Amsterdam District Court held during the same period that the publishers had failed to demonstrate sufficiently that their robots.txt file, a small text file on a website containing instructions for search engines, scrapers and AI crawlers, constituted an appropriate machine-readable reservation. The proceedings revealed that the robots.txt in question excluded only a limited number of specific bots (including GPTBot), but did not contain a general machine-readable reservation.
Every limitation on copyright is moreover subject to the "three-step test". That test provides that an exception is only permissible if it covers a special case (1), does not conflict with the normal exploitation of the work (2), and does not unreasonably prejudice the legitimate interests of the author (3). In the Amsterdam case, the court held that the TDM exception did not fail the three-step test in the circumstances, since HowardsHome's use did not conflict with the primary market of the publishers. That is a relevant nuance: the exception applies, but its boundaries are policed by this test.
Is AI output protected by copyright, and can it infringe?
Copyright protection requires an original intellectual creation ("OIC"): the work must reflect the personality of the author through the expression of free and creative choices. With AI-generated output, this is problematic, because the concrete creative elaboration is largely determined by the system. The human typically provides only functional instructions, and the AI model determines the result.
This line is also broadly followed in case law. The Amtsgericht München held in early 2026 that three AI-generated logos were not protected by copyright, even though the user had entered detailed prompts and gone through multiple iteration rounds. The essential creative choices, such as composition and style, were made by the AI model, not by the user. Copyright does not protect effort or time, but exclusively creative choices. The Oberlandesgericht Düsseldorf took the same approach in a case involving an AI-edited photograph of a dog underwater: there too, the AI image did not qualify as a work, because the user had not demonstrated which creative choices they had made.
AI output may itself also infringe existing works. Here the phenomenon of memorisation is relevant. Memorisation occurs when an AI model reproduces fragments of its training data in near-verbatim form, because certain works are so deeply embedded in the model parameters that they recur literally in response to particular prompts. In some cases this may constitute a (near-)verbatim reproduction of a copyright-protected work. The Munich Regional Court held in November 2025, in the case of GEMA v OpenAI, that song lyrics reproduced by ChatGPT virtually word for word qualify as a reproduction within the meaning of copyright law. The fact that the work is not stored as a text file but as numerical probability distributions in model weights did not alter this finding. A notable detail is that the AI provider bears liability here, not the end user who entered the prompt.
What must you disclose under the AI Act?
Anyone using generative AI to produce content is not only subject to copyright law but also to the transparency obligations under the AI Act. These apply regardless of whether you hold rights in the output; they are directed at anyone who publishes AI-generated content.
Article 50(4) of the AI Act contains two independent obligations for the deployer, the party using an AI system under its own responsibility. The first obligation concerns so-called deepfakes: anyone who generates or manipulates images, audio or video that could falsely be mistaken for real must make this clearly and conspicuously known. The definition is broader than the classic notion of a deepfake in which someone's face is misused. It also covers an AI-generated photograph of, for instance, a touched-up product in an advertisement, as long as it is conceivable that it could be mistaken for authentic content. This type of content must be labelled visibly alongside the content itself. The regime is more lenient for "manifestly non-realistic content", such as a cartoon, an abstract illustration or an obvious parody. For that category, a non-intrusive disclosure in credits or metadata suffices.
The second obligation concerns AI-generated text published to inform the public on matters of general interest. The trigger here is not deception but the publication context: a fully AI-written news article falls squarely within scope, even without any manipulation of real persons. An exception applies where the content has genuinely undergone human editorial review and an identifiable person or entity bears editorial responsibility.
Conclusion
Generative AI raises fundamental copyright questions that the law has not yet fully resolved. These questions extend across the lawfulness of AI training, the protection of AI output, and liability for memorisation. Anyone working with generative AI would be well advised to take a number of practical steps without waiting for definitive answers.
A rightholder wishing to protect their work against its use as training data for generative AI is advised to implement a broad robots.txt that goes beyond excluding a handful of well-known bots. It is also advisable to record the reservation in metadata.
For those who wish to be able to protect AI output: document the creative process. Preserve prompts and iteration steps and record which creative choices you made and which the AI made. Without that documentation, you will be left empty-handed in any dispute.
Anyone publishing AI-generated content should assess on a publication-by-publication basis whether the content constitutes realistically appearing generated material or AI-written text on matters of general interest. If so, a labelling obligation applies under Article 50 of the AI Act. When in doubt: label rather than not.
Did you find this topic interesting? Next year, a study day on IP & AI is planned. Further information will follow. Keep an eye on our IP page and sign up for our newsletter!
/58.png?width=600&height=300&name=58.png)
/63.png?width=600&height=300&name=63.png)
