Does day-to-day AI-powered legal work risk breaking the law?

The scenario is so common it barely registers as noteworthy: a lawyer downloads a legal commentary from a subscription legal database, uploads it to ChatGPT with the prompt "summarize the key arguments," and moves on with their day. Efficient, practical, seemingly harmless. 

It might also be illegal. 

A November 2025 Munich court decision against OpenAI, combined with little-noticed licensing restrictions in legal databases and German bar association guidance, has created a compliance gap for legal practitioners across Europe. The risks stem from three sources: copyright law, professional obligations, and licensing agreements. 

The Munich Precedent on AI and Copyright 

On November 11, 2025, the Munich Regional Court issued the first European ruling finding that training AI models on copyrighted works constitutes copyright infringement. In GEMA v. OpenAI, Germany's music collecting society successfully argued that ChatGPT had unlawfully memorized song lyrics during training and could reproduce them on demand.1  

The court's reasoning is simple and clear: when copyrighted content becomes embedded in a model's parameters, that constitutes reproduction under Section 16 of the German Copyright Act. When the model then outputs that content, it constitutes public communication under Section 19a. The technical details, the court held, were irrelevant. 

"Memorization fulfills the requirements for reproduction," the court stated, rejecting OpenAI's argument that statistical representations of works in model weights shouldn't count as copying. The court analogized the process to MP3 compression: even though the data is transformed and distributed across parameters, if the model can "generate statistically probable token sequences that recognizably reproduce" the original work, copyright is infringed. 

Recent scientific research confirms the court's technical reasoning. A January 2026 study from Stanford and Yale researchers demonstrated that production AI models, including Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3, can reproduce copyrighted books near-verbatim. In some cases, researchers extracted up to 95.8 percent (!) of entire books from models, even with safeguards in place.2 The study thus undermines any claims that AI models merely learn patterns without storing actual content. 

The court also rejected the text and data mining exception under Section 44b of the German Copyright Act. That exception permits temporary reproductions for analytical purposes, but the court found it doesn't apply when works are permanently memorized in model parameters rather than merely analyzed and discarded.3 In other words, temporary copying for analysis is permitted, but permanent embedding of copyrighted works in retrievable model weights is not. 

OpenAI has appealed the decision, which is not yet final. But the ruling's implications could reach beyond music generation into daily legal practice. The court's reasoning about permanent memorization has a broader application: if the text and data mining exception doesn't apply when works are permanently memorized in model parameters, then even uploading a copyrighted legal commentary to ChatGPT or any open LLM could constitute reproduction under copyright law. The upload by the end-user isn't temporary analysis - it's feeding content into a system designed to permanently embed it. 

Under this logic, both OpenAI and the lawyer who uploads a document from a legal database face potential copyright liability. 

Germany's Bar Association Saw It Coming 

This compliance risk wasn't entirely unforeseen. Four months before the Munich court validated these concerns, Germany's largest bar association had already flagged the issue. The German Bar Association's Statement 32/2025, published in July 2025, explicitly addressed copyright requirements for AI use in legal practice - guidance that received little attention at the time.4 

The guidance included a clear requirement: "Bei der Nutzung von KI-Systemen sind urheberrechtliche Vorgaben unerlässlich, insbesondere bei der Bearbeitung oder Zusammenfassung geschützter Fachliteratur. Eine Vervielfältigung, beispielsweise durch das Hochladen von Werken auf KI-Anbieter-Server, muss durch Lizenzen oder gesetzliche Erlaubnisse abgedeckt sein." 

Translation: Copyright requirements are essential when using AI systems, particularly when processing or summarizing protected legal literature. Reproduction, for example by uploading works to AI provider servers, must be covered by licenses or legal permissions. 

The statement identified precisely the issue the Munich court would later validate: uploading copyrighted materials to AI systems requires explicit authorization. But how many practitioners actually checked whether their legal database subscriptions granted such authorization? 

The Fine Print in Database Agreements 

Content licensing platforms across industries have begun addressing not just AI training, but also end-user behavior. To take just one example, Spotify's User Guidelines, updated in September 2025 and applicable to every subscriber, prohibit "using any part of the Services or Content to train a machine learning or AI model or otherwise ingesting Spotify Content into a machine learning or AI model."5 The phrase "or otherwise ingesting" goes beyond restricting AI companies from scraping content as training data: it captures the everyday act of a user feeding content into an AI tool. 

If music streaming and stock photography platforms already impose such restrictions on their users, legal database providers, whose content consists entirely of copyrighted legal commentary, case annotations, and scholarly analysis, can be expected to do so as well.6 

Should We Expect a Global Enforcement Wave? 

Regarding copyright law, it is now clear that the initial "move fast and break things" approach to training data adopted by pioneer AI companies is giving way to licensing requirements and litigation. Lawsuits against AI companies acceleratedsharply over the past two years. The first cases emerged in early 2023, when visual artists sued Stability AI and authors sued OpenAI and Meta.7 By mid-2024, roughly 25 different lawsuits were working through various courts.8 That number climbed past 65 active lawsuits by late 2025, with major corporate plaintiffs joining individual creators: The New York Times (December 2023), Getty Images, record labels, and film studios.9 In Europe, GEMA, emboldened by its victory, filed a similar lawsuit against music AI platform Suno in January 2025, with a hearing held in late January 2026.10 

On professional guidance, however, no global trend can be recognized: the German Bar Association remains the only bar association to state explicitly that lawyers themselves may commit copyright infringement by uploading copyrighted materials to AI systems, and to address the ethical implications of that conduct. No other European or US bar association has tackled this issue head-on yet.11 

Why the silence? Perhaps the behavior feels too innocuous. Lawyers are accustomed to excerpting, quoting, and summarizing legal materials for research. And it is technically difficult to grasp that uploading copyrighted content to AI systems differs fundamentally from traditional copying (does everyone understand what “reproduction in a form that embeds the work in retrievable model parameters” mean?). 

Perhaps many professionals believed or preferred to believe AI companies' repeated assertions that models do not really “save” information from training data and cannot reproduce it. The Stanford-Yale study proving models can reproduce up to 95.8 percent of copyrighted books demolished that claim. 

Or perhaps law firms without secure AI solutions prefer not to confront the reality of shadow AI use within their ranks. Better not to read, black on white from their bar association, that this constitutes both copyright infringement and a violation of professional rules. 

Therefore, while courts will likely provide clarity on copyright infringement by AI companies training models over the coming months, it might be a while before we read press articles about end users caught uploading copyrighted materials. And then again… 

Headlines in early February 2026 offered a preview of the reputational damage such incidents can trigger. 12, and hundreds of pieces of media coverage.13 

European government agencies and bar associations would be wise to issue clear guidance before such headlines emerge on their side of the Atlantic. 

References:

¹ Landgericht München I, Urteil vom 11. November 2025, Az. 42 O 14139/24, GEMA gegen OpenAI. Pressemitteilung abrufbar unter: Pressemitteilung 11/2025, Bayerisches Staatsministerium der Justiz. Zusammenfassung auf Englisch: CMS Law, „GEMA vs. OpenAI: Munich Regional Court I issues landmark copyright decision," Dezember 2025. https://cms-lawnow.com/en/ealerts/2025/12/gema-vs.-openai-munich-regional-court-i-issues-landmark-copyright-decision 

² Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo und Percy Liang, „Extracting books from production language models," arXiv:2601.02671, 6. Januar 2026. https://arxiv.org/abs/2601.02671 

³ Norton Rose Fulbright, „Germany delivers landmark copyright ruling against OpenAI: What it means for AI and IP," November 2025. https://www.insidetechlaw.com/blog/2025/11/germany-delivers-landmark-copyright-ruling-against-openai-what-it-means-for-ai-and-ip 

⁴ Deutscher Anwaltverein (DAV), Stellungnahme 32/2025: „Einsatz von KI in der Anwaltschaft," 9. Juli 2025. Berichterstattung: Datenschutzticker, 5. August 2025. 

⁵ Spotify, Nutzungsrichtlinien, https://www.spotify.com/us/legal/user-guidelines/ 

⁶ Ein Beispiel aus den Allgemeinen Geschäftsbedingungen von LexisNexis, Abschnitt 1.2(h): „The uploading or submission of Materials into third party applications, software or websites that utilize AI Technologies is prohibited unless approved in writing by LN", gültig ab 31. Oktober 2025. https://www.lexisnexis.com/en-us/terms/general/default.page 

⁷ Die bildenden Künstlerinnen Sarah Andersen, Kelly McKernan und Karla Ortiz reichten am 13. Januar 2023 eine Sammelklage gegen Stability AI, Midjourney und DeviantArt ein (Andersen et al. v. Stability AI LTD et al., N.D. Cal. 3:23-cv-00201). Die Autorinnen und Autoren Sarah Silverman, Richard Kadrey und Christopher Golden klagten gegen OpenAI (28. Juni 2023) und Meta Platforms (7. Juli 2023). Siehe CNN Business, „Sarah Silverman sues OpenAI and Meta alleging copyrightinfringement," 10. Juli 2023, https://www.cnn.com/2023/07/10/tech/sarah-silverman-openai-meta-lawsuit/index.html; Copyright Alliance, „Takeaways from the Andersen v. Stability AI Copyright Case," 12. November 2024, https://copyrightalliance.org/andersen-v-stability-ai-copyright-case/ 

⁸ Copyright Alliance, „AI Lawsuit Developments in 2024 (Mid-Year Review)," 15. Oktober 2024, https://copyrightalliance.org/ai-lawsuit-developments-2024/ 

⁹ eDiscovery Today, „65 AI Copyright Lawsuits, One Company Virtually Unscathed," 8. Dezember 2025, https://ediscoverytoday.com/2025/12/08/65-ai-copyright-lawsuits-one-company-virtually-unscathed-artificial-intelligence-trends/ 

¹⁰ Music Ally, „GEMA scores court victory in its copyright battle with OpenAI," 11. November 2025, https://musically.com/2025/11/11/gema-scores-court-victory-in-its-copyright-battle-with-openai/ 

¹¹ Auswertung der Leitlinien folgender Rechtsanwaltskammern und -verbände: ABA (Formal Opinion 512, Juli 2024), Law Society of England and Wales (Generative AI: the essentials, Mai 2025), CCBE (Leitfaden zum Einsatz generativer KI durch Anwälte, 2025), französischer CNB (Leitfaden zum Einsatz generativer KI, 2024-2025), italienischer CNF (Informationsschreiben zum KI-Einsatz, Oktober 2025), spanischer CGAE (Schulungen und Leitlinien, 2024-2025), belgischer Orde van Vlaamse Balies (KI-Leitlinien, Januar 2025), niederländischer NOvA (KI-Empfehlungen, November 2025), Law Society of Singapore sowie über 30 US-amerikanische Anwaltskammern. Die Auswertung zeigt, dass alle genannten Stellen zwar Fragen der Vertraulichkeit, Sorgfaltspflicht, Überprüfung, Abrechnung und Mandanteneinwilligung beim KI-Einsatz ansprechen, keine jedoch die urheberrechtlichen Konsequenzen des Hochladens geschützter Materialien aus Fachdatenbanken in KI-Systeme durch Anwältinnen und Anwälte selbst thematisiert. Weitere Quellen: CNB, „Guide pratique - Utilisation des systèmes d'intelligence artificielle générative pour les Avocats," 2024-2025; italienischer CNF, Informationsschreiben gemäß Gesetz 132/2025, Oktober 2025; belgischer OVB, „Richtlijnen voor advocaten rond gebruik van artificiële intelligentie," 20. Januar 2025, https://www.ordevanvlaamsebalies.be/nl/kennisbank/digitalisering/richtlijnen-voor-advocaten-rond-gebruik-van-artificiele-intelligentie; niederländischer NOvA, „Aanbevelingen AI in de advocatuur," November 2025, https://www.advocatenorde.nl/dossiers/digitalisering-en-ai 

¹² Politico berichtete als erstes über den Vorfall am 28. Januar 2026. Siehe auch: Cybersecurity Asia, „ChatGPT Risk at Centre of Controversy After Acting US CISA Head Uploads Government Documents," 30. Januar 2026, https://cybersecurityasia.net/chatgpt-rist-at-centre-of-controversy-cisa/; The Daily Caller, „US Cyber Defense Agency Head Posted Sensitive Information Online," 28. Januar 2026, https://dailycaller.com/2026/01/28/madhu-gottumukkala-cisa-posted-sensitive-information/; Digit, „CISA ChatGPT leak: Acting director Madhu Gottumukkala investigation explained," 28. Januar 2026, https://www.digit.in/features/general/cisa-chatgpt-leak-acting-director-madhu-gottumukkala-investigation-explained.html