AI, Copyright and Deepfakes

Home › AI law › AI, copyright and deepfakes

Generative AI touches copyright at two points: training on third-party works and the output the system produces. Personality rights come in with deepfakes. We act for authors, publishers and rights holders as well as for companies that develop AI or use its output.

AI training and text and data mining under section 44b UrhG

Training involves reproducing works. Without a licence, this is only permitted under an exception. Section 44b of the German Copyright Act (UrhG) allows text and data mining for any purpose, including commercial purposes, unless the rights holder has reserved its rights. For scientific research, the broader exception in section 60d UrhG applies, which does not allow a reservation.

Whether training generative models is fully covered by section 44b is not settled. In GEMA v OpenAI, the Munich I Regional Court held that the exception covers preparatory analysis but not the permanent storage of works in the model (judgment of 11 November 2025, 42 O 14139/24).

Declaring an effective opt-out

For works available online, a reservation is only effective if it is machine-readable (section 44b(3) UrhG). In Kneschke v LAION, the Hamburg Higher Regional Court held that a reservation in natural language in a website’s terms of use was not sufficient; it had to be capable of being processed by machines (judgment of 10 December 2025, 5 U 104/24). Leave to appeal to the Federal Court of Justice was granted. We therefore recommend a reservation on several levels: robots.txt covering the known AI crawlers, protocols such as the TDM Reservation Protocol, metadata in the files, and a statement in the legal notice and terms.

Under Art. 53 of the AI Act, providers of general-purpose AI models must have a copyright policy, respect machine-readable reservations and publish a summary of their training content.

Memorisation: when the model contains works

If a model can largely reproduce works in response to simple prompts, the Munich I Regional Court considers that a reproduction exists in the model itself. It ruled this way on song lyrics against OpenAI and, on 31 July 2026, against the music generator Suno (42 O 763/25). In both cases it held the provider, not the users, responsible for the output. Neither judgment is final.

Protection of AI output

Copyright requires a personal intellectual creation (section 2(2) UrhG). A prompt alone is generally not enough. What a human selects, arranges and creatively edits can, however, be protected. If you use AI content commercially, document your own contribution and do not rely on copyright alone, but also on trade marks, designs, unfair competition law and contractual exclusivity.

Third-party rights in AI output

Even if the model is not supposed to „copy“, a result may contain third-party works, characters, trade marks or faces. Whoever uses the content is liable for publishing it. Check in particular images in the style of well-known characters, music with recognisable melodies and code that may be subject to open-source licences. Whether chatbot answers infringe press publishers‘ rights is pending before the CJEU (C-250/25, Like Company v Google Ireland); the Advocate General delivered his opinion on 3 November 2026, and the judgment is awaited.

Licences for AI training and use

Rights holders can grant training licences, individually or through collecting societies; GEMA has presented a licensing model for generative AI. For companies that offer or acquire content for training, we draft licence agreements with clear provisions on purpose, scope, remuneration, deletion and liability.

Deepfakes and personality rights

Deceptively real images, voices and videos of real people regularly infringe the right to one’s own image (sections 22 and 23 of the German Art Copyright Act, KUG) and the general right of personality. Since August 2026, deployers must label deepfakes as artificially generated under Art. 50(4) AI Act. From 2 December 2026, AI systems that generate intimate images of identifiable people without consent are prohibited.

Current proceedings

Case	Court, date	Status
GEMA v OpenAI (song lyrics)	Munich I Regional Court, 11.11.2025, 42 O 14139/24	Infringement through memorisation and output found; not final
GEMA v Suno (music)	Munich I Regional Court, 31.07.2026, 42 O 763/25	Injunction, information, damages; not final
Kneschke v LAION (training dataset)	Hamburg Higher Regional Court, 10.12.2025, 5 U 104/24	Claim dismissed (section 60d UrhG); appeal on points of law allowed
Like Company v Google (Gemini)	CJEU, C-250/25	AG opinion 03.11.2026; judgment pending

FAQs

May I use AI-generated images in my advertising?

Yes, if the result does not infringe third-party rights and the service’s terms allow commercial use. You will usually not have exclusive protection.

Is a note in the legal notice enough as an opt-out?

After the Hamburg judgment, not reliably. Add machine-readable reservations, for example in robots.txt.

Can I take action against an AI company that used my works?

Yes, if you can prove the use. Training content summaries under Art. 53 AI Act, information claims and tests whether the model reproduces your works all help.

Related topics

• AI law overview
• AI Act
• Online attacks on your reputation
• AI training (in German)
• LAION judgment (in German)
• Deepfakes (in German)