AI and Copyright: Segregating AI Use When Clients Reject It

AI and copyright has become arguably the most important legal question surrounding generative AI. Music publishers, film studios, newspaper groups, stock photo agencies, and authors accuse the major AI providers of using their works for training without consent – and they are suing. For agencies and smaller companies, this has a very concrete consequence: more and more clients categorically refuse to have their content and data processed by AI.

When a brand-conscious client or a major media house rejects AI processing, good intentions are not enough – you have to guarantee it organizationally and, if challenged, be able to prove it. This article briefly maps the legal situation and then focuses on what really matters for you: how to cleanly segregate AI use internally so that you can reliably serve AI-averse clients while making full use of AI wherever it is permitted.

Note: This article is a journalistic overview, not legal advice. Many proceedings have not been finally decided. Status of proceedings: July 2026.

Germany and the EU: First Landmark Rulings

The most important German case is GEMA v. OpenAI. On November 11, 2025, the Regional Court of Munich I found that OpenAI infringed existing copyright law through the training and operation of ChatGPT.¹ The case concerned nine well-known German song lyrics – including "Atemlos" and "Über den Wolken". The court considered it proven that the lyrics were stored in the model and reproduced in its outputs, and rejected protection under the text-and-data-mining exception. It is the first landmark ruling of its kind in Europe. OpenAI filed an appeal in December 2025; the case is now before the Higher Regional Court of Munich and remains open.¹

In parallel, GEMA is pursuing a second case against the audio AI provider Suno – the first case concerning music AI in Europe.² A ruling is still pending; the court has scheduled its decision for July 31, 2026.²

That courts do not automatically rule in favor of rights holders is shown by Kneschke v. LAION. Here, the Higher Regional Court of Hamburg confirmed in December 2025 that downloading and analyzing images for an AI training dataset can be covered by the text-and-data-mining exception – provided any objection (opt-out) is machine-readable; a reservation expressed in natural language is not sufficient.³ The photographer has appealed to the Federal Court of Justice (BGH). This case, too, remains open.

The regulatory framework is set by the EU AI Act: providers of general-purpose AI models (GPAI) have been required since August 2, 2025 to publish a summary of their training content.⁴ Transparency about training data has thus become a legal obligation – regardless of how individual lawsuits turn out.

The US: Plenty of Movement, Mixed Signals

In the US, the picture is less uniform. Two rulings classified training itself as "fair use" – permitting it in principle. In Kadrey v. Meta, the court ruled in favor of Meta in June 2025 but explicitly stressed that the decision did not mean AI training was lawful in general; the plaintiff authors had simply made the wrong arguments.⁵

How double-edged "fair use" can be is shown by Bartz v. Anthropic. Here too, training itself was deemed fair use – but sourcing the books through illegal "shadow libraries" was not. Anthropic agreed to a settlement of around 1.5 billion US dollars, the largest copyright settlement in US history.⁶ The signal: even where training may be permitted, the provenance of the data remains a substantial risk.

In the music sector, several cases have been partly settled and partly remain open. Universal Music and Warner Music have settled with AI providers and entered licensing partnerships; Sony Music continues to sue Suno and Udio.⁷ In addition, the major music publishers (including Universal, Concord, and ABKCO) are suing Anthropic over the use of song lyrics – that case is ongoing.⁸

Getty Images v. Stability AI likewise shows that rights holders do not win across the board: before the High Court in London, Getty dropped its central copyright claims during the proceedings; the court found only a limited trademark infringement.⁹ A parallel case is pending in the US. And the New York Times, together with other publishers, continues to litigate against OpenAI and Microsoft; the case is in the discovery phase, with no ruling yet.¹⁰ Hollywood is involved as well: Disney, Universal, and DreamWorks are suing the image generator Midjourney, which in turn invokes fair use.¹¹

The Pattern Behind the Cases

However differently the rulings turn out, three questions run through almost all the proceedings:

Training versus sourcing. Several courts treat the training of a model differently from the question of where the data came from. Legally sourced data is judged more leniently than pirated data – the Anthropic settlement makes that expensively tangible.⁶

Output and memorization. When a model reproduces protected content almost verbatim, that weighs more heavily than the training process alone. This was precisely the basis of the Munich GEMA ruling.¹

Opt-out and transparency. In the EU, rights holders can object to the use of their data – but only in machine-readable form.³ And the AI Act obliges providers to disclose their training sources.⁴

The overall picture: the legal situation is in flux, and outcomes vary by country and case constellation – sometimes rights holders prevail, sometimes the providers. Legal certainty in the sense of a final resolution does not yet exist.

The Core Issue: Segregating AI Use Internally

For agencies and smaller companies, the real challenge is not the outcome of individual lawsuits but dealing with clients who fundamentally reject AI processing of their content. These are often the most demanding and most valuable clients: strong brands and media houses that rigorously protect their intellectual property. Disney, which is taking action against AI image generators,¹¹ or the New York Times, which is suing OpenAI,¹⁰ exemplify a stance that is spreading: "Under no circumstances does our content go into an AI system."

Anyone serving such clients needs more than good intentions. They need a clear internal boundary – so that one client's content never touches an AI tool while AI is used to the full for another. This segregation rests on six building blocks:

• Classify clients by AI permission. Keep a simple register: which client permits AI, which only with restrictions, which not at all? This classification is the foundation for everything else.
• Separate content technically. For "AI-prohibited" clients, assets, texts, and data do not belong in AI-powered tools – not even "just to try it out". Separate workflows, folders, and access rights prevent mixing.
• Label assets. Mark client material so that everyone on the team can see at a glance whether AI use is permitted. Anything not explicitly cleared stays out.
• Document usage. Record which tools were used for which client. Only then can you prove, if challenged, that no AI was used – evidence that demanding clients increasingly require.
• Clarify it contractually. Establish in writing before the project starts which AI use a client permits, and anchor it in the contract. Train the team on these boundaries.
• Where AI is permitted, work cleanly. For clients who allow AI, prefer tools with reliable data provenance (such as models positioned as "commercially safe"), review providers' IP indemnification commitments, and check outputs for excessive similarity to protected works.

This segregation is not a brake. It is the precondition for retaining AI-averse clients while still benefiting from AI – instead of having to choose between the two.

Conclusion

The legal situation around AI and copyright remains uncertain for now – sometimes rights holders win, sometimes the providers. For agencies and companies, however, the decisive question is not who prevails in the end, but whether they can deploy AI in a way that keeps clients on board who reject AI processing. The key is internal segregation: classify clients by AI permission, separate their content technically, document usage. Those who draw this line cleanly use AI to the full where it is permitted – and reliably keep it out where a client demands it.

If you want to draw this line for your company, talk to us – a structured look at your starting position, no sales pitch, no slides.

Sources & References

1. Regional Court of Munich I (case no. 42 O 14139/24), ruling of Nov 11, 2025, GEMA v. OpenAI; appeal pending before the Higher Regional Court of Munich: press release of the court, justiz.bayern.de · gema.de
2. Regional Court of Munich I (case no. 42 O 763/25), GEMA v. Suno, decision scheduled for July 31, 2026: court press release, justiz.bayern.de · gema.de
3. Higher Regional Court of Hamburg (case no. 5 U 104/24), Kneschke v. LAION, appellate ruling Dec 2025 (appeal filed with the Federal Court of Justice): ki-kanzlei.de
4. European Commission: GPAI transparency obligations since Aug 2, 2025, template for disclosing training data: digital-strategy.ec.europa.eu
5. Kadrey v. Meta, N.D. California, ruling of June 25, 2025 (training as fair use, narrowly reasoned): analysis by Goodwin, goodwinlaw.com
6. Bartz v. Anthropic, settlement of ~1.5 billion USD (2025): Authors Guild, authorsguild.org
7. Music labels (UMG, Sony, Warner) v. Suno & Udio – settlements and ongoing proceedings, overview: musicbusinessworldwide.com
8. Concord Music Group et al. v. Anthropic (song lyrics), ongoing proceedings: musicbusinessworldwide.com
9. Getty Images v. Stability AI, High Court of England & Wales, ruling of Nov 4, 2025 (largely in favor of Stability, limited trademark infringement): analysis by Bird & Bird, twobirds.com
10. The New York Times et al. v. OpenAI/Microsoft, S.D.N.Y., ongoing proceedings: techcrunch.com
11. Disney, Universal, and DreamWorks v. Midjourney (lawsuit filed June 2025, C.D. California; Midjourney invokes fair use), reported by NPR: npr.org