Somewhere inside the parameters of a large language model, a Munich court has ruled, sit the recognizable echoes of “Atemlos” and “Über den Wolken” — German pop songs that ChatGPT can reproduce almost verbatim when asked. That finding, delivered in November 2025, made OpenAI the first AI developer in Europe to be held directly liable for copyright infringement. Across the Atlantic, in a federal courthouse in Delaware, a different judge had reached a similar conclusion nine months earlier: that using copyrighted legal research materials to train an AI system is not fair use, no matter how sophisticated the technology doing the training.
Two courts, two continents, one question. What does an AI company owe the people whose work it learnt from?
The answer to that question — being constructed case by case, jurisdiction by jurisdiction, in courtrooms from Munich to Manhattan — will determine whether the generative AI industry is built on a foundation of licensed human creativity or on three decades of unauthorized reproduction at scale. It will reshape music, publishing, journalism, visual art, and film. It will define the compliance obligations of every AI company on earth. And it will do so not through legislation — which has proved consistently too slow — but through the accumulated weight of judicial decisions that are arriving, finally, with real consequence.
The Case That Started the Precedent Clock
Before November 2025’s European rulings, the closest thing to settled U.S. law on AI training data came from a case that had nothing to do with generative AI at all.
Thomson Reuters v. Ross Intelligence, decided on February 11, 2025, by Judge Stephanos Bibas of the U.S. District Court for Delaware, involved Ross Intelligence — a company that built an AI-powered legal research tool by training it on Westlaw headnotes, the proprietary legal summaries that Thomson Reuters publishes as part of its subscription research platform. Ross had obtained those headnotes indirectly, through a third party, and used them to train a system designed to compete directly with Westlaw.
The court found copyright infringement and rejected Ross’s fair use defense — the first time any U.S. court had ever concluded whether AI training data use qualifies as fair use. The ruling’s logic cut to the heart of the AI industry’s standard defense. Ross argued its use was transformative because it was training a machine rather than reproducing the content publicly. Judge Bibas disagreed on the fourth factor — market harm — finding that Ross’s system could damage Thomson Reuters’s potential market for AI training data licenses, even though no such formal market yet fully existed. The ruling established that courts would protect not just existing markets, but the reasonable future markets that copyright owners might develop.
The case was immediately cited in Bartz v. Anthropic — the class action brought by authors whose books were used to train Claude — as courts began applying its reasoning to generative AI. In June 2025, a California federal court ruled that Anthropic may have illegally downloaded some seven million books. Two months later, Anthropic settled for up to $1.5 billion — the largest copyright settlement in U.S. history. The Thomson Reuters decision had become the jurisprudential foundation for a wave of AI copyright litigation in a matter of months.
The European Turn: Munich Rules Against OpenAI
While American courts were carefully hedging their reasoning around the specific facts of each case, Germany took a more direct route.
GEMA — Germany’s music rights collecting society — sued OpenAI over the use of copyrighted song lyrics in training GPT-4 and GPT-4o. The case was narrow in its facts: nine well-known German songs whose lyrics, GEMA argued, had been memorized in the model’s parameters and could be reproduced through simple prompts. The Munich Regional Court I found in GEMA’s favor on November 11, 2025, in a ruling that has since reverberated across the European legal landscape.
The court’s core finding was striking in its specificity. It ruled that when an AI model memorizes training data to the point where it can reproduce it verbatim or near-verbatim on demand, that constitutes reproduction under EU and German copyright law — and that the text-and-data-mining exceptions in EU law apply only to the initial analytical phase of training, not to the permanent embedding of recognizable content in model weights. OpenAI’s argument that it was merely learning statistical correlations, not storing content, was rejected. The court drew an analogy to lossy MP3 compression: even if the storage method differs from traditional copying, if the content can be faithfully reconstructed from what is stored, the law treats it as reproduction.
OpenAI’s secondary argument — that any infringement was the responsibility of users who prompted the reproduction, not OpenAI itself — was also dismissed. The court held OpenAI directly liable for making the lyrics accessible to an unlimited public. GEMA v. Suno, a related case against an AI music generator, is expected to produce a ruling in June 2026.
The contrast with U.S. doctrine is deliberate and significant. As legal analysts at Norton Rose Fulbright noted, the European approach treats copyright as an economic property right requiring prior authorisation, while U.S. law operates through the fair use balancing test, which allows flexibility but also uncertainty. One week before the Munich ruling, the English High Court had delivered its own verdict in Getty Images v. Stability AI — and reached a more limited conclusion. The UK court rejected Getty’s primary copyright claim, finding that because Stability AI’s training occurred largely outside UK territory, there was no basis for UK copyright infringement on the training itself. Getty won only limited trademark claims related to the reproduction of its watermark. Two European courts, one week apart, had drawn the line in different places.
The Biggest Case Still in Motion: NYT v. OpenAI
The case that most of the AI industry is watching most closely remains unresolved: The New York Times v. Microsoft and OpenAI, filed in December 2023 and working through discovery in the Southern District of New York.
The Times alleges that OpenAI and Microsoft copied millions of its articles to train ChatGPT, and that ChatGPT can reproduce those articles in ways that directly substitute for — and compete with — the Times’s own subscription journalism. The suit seeks “billions” in statutory damages. In April 2025, the court consolidated twelve related cases against OpenAI — from authors, news organizations, and an online video creator — into a Multidistrict Litigation, centralizing them in the Southern District of New York to streamline proceedings. Summary judgment in the main Times case was scheduled for April 2026.
The discovery phase has already produced one landmark order. In November 2025, Magistrate Judge Ona Wang ordered OpenAI to produce 20 million ChatGPT conversation logs to the Times — the first time a court had compelled an AI company to expose its product’s output records at this scale. The order was upheld in January 2026 when the court rejected OpenAI’s objections. The logs are designed to find instances where ChatGPT users attempted to extract verbatim Times journalism through targeted prompting. If such examples exist at scale, they could demonstrate the kind of market substitution that the Thomson Reuters precedent identified as the most damaging form of copyright harm.
The Times separately sued Perplexity AI in December 2025, accusing the AI-powered research assistant of reproducing its journalism to power its product. The suit followed eighteen months of failed licensing negotiations. It signals a shift in strategy: rather than licensing their way into coexistence with AI, major news organizations are now choosing litigation as their primary lever.
Japan’s Different Bet and What It Reveals
Not every major economy is litigating its way toward a framework.
Japan has taken a sharply different approach. Its Copyright Act contains a broad exception — Article 30-4 — that explicitly permits the use of copyrighted works for machine learning and AI training without licensing requirements, provided the use is “for information analysis” and not for “enjoying the thoughts or sentiments expressed in the works.” Japan’s Cabinet Office reinforced this interpretation in government policy guidance in 2024, positioning the country as a deliberately permissive environment for AI development, particularly to attract global AI companies and build domestic AI capability.
This approach has drawn both praise and concern. Proponents argue Japan’s framework allows AI developers to train on the full richness of human creative output, enabling more capable models, and that the country can build a sustainable licensing ecosystem over time without litigation-driven disruption. Critics — including domestic authors, illustrators, and musicians who have organized significant public opposition — argue that Japan has unilaterally granted AI companies access to decades of creative work without compensation to its creators. The debate is ongoing, and Japan’s government has signaled it may revisit the framework as the legal landscape in the U.S. and EU becomes clearer.
The contrast between Tokyo and Munich is the sharpest illustration of the global divergence that now defines AI copyright law. In Germany, storing a song lyric in a model weight is infringement. In Japan, training on that same lyric is explicitly permitted by statute.
What the Rulings Mean in Practice
For AI companies, the practical message of 2025–2026 is that the data you train on is a legal liability, not merely an engineering input.
The Thomson Reuters ruling established that potential future markets for AI training data licenses are protectable even before those markets are fully formed. Anthropic’s $1.5 billion settlement demonstrated what exposure at scale looks like. The Munich ruling extended the logic to generative outputs: if your model can reproduce a song lyric, a news article, or a chapter of a book, the act of embedding that reproducibility in your model’s weights is itself an infringement under European law. The UK ruling added a territorial wrinkle: training conducted outside a jurisdiction’s borders may avoid that jurisdiction’s copyright law, even if the resulting product is deployed there.
The music industry has moved faster than any other creative sector in translating this legal pressure into commercial structure. Warner Music settled with Suno and pivoted into a licensing partnership that allows licensed AI models and artist offtake arrangements. Other labels are watching the GEMA v. Suno ruling in June 2026 to calibrate their next moves. Publishers and news organizations are split between litigation and licensing, with the NYT on one side and a growing list of news organizations that have signed deals with OpenAI — including the Associated Press and several European publishers — on the other.
For companies outside the AI industry that depend on creative output — film studios, advertising agencies, game developers, fashion houses — the court cases matter because they will determine whether AI tools that their teams now use daily were built legally. If the NYT prevails at summary judgment, the secondary liability question — does every company using ChatGPT share in the infringement? — becomes the next front in the litigation.
The Copyright System the Courts Are Being Asked to Build
The fundamental problem these cases are trying to solve was not created by the courts and cannot be fully resolved by them.
Copyright law was designed for a world of discrete copying events: a publisher reproducing a book, a broadcaster airing a recording. It was not designed for a world in which a single training run ingests hundreds of billions of tokens of human-created text and encodes their statistical patterns into a set of numerical weights that no human will ever directly read. Whether that process constitutes copying in the legal sense — and if so, what license it requires and at what price — involves technical, economic, and policy judgments that courts are poorly equipped to make by themselves.
What the cases of 2025 and 2026 have established, collectively, is that the answer to the core question — is AI training copyright infringement? — is neither a universal yes nor a universal no. It depends on what was trained on, how, where, and to what degree the model can reproduce the original. It depends on whether the training displaced a market that the rights-holder had or could reasonably expect to develop. And it depends on which country’s court is deciding the matter.
That fragmentation is itself a consequence of the speed at which AI development has outrun the legal frameworks built to govern creative work. The cases will continue. More rulings will arrive. The New York Times summary judgment decision will move the needle further. The GEMA v. Suno music ruling in June 2026 may settle the most commercially sensitive question in the music industry for a generation.
The courtrooms are doing the work that legislatures have not. And the rules they write — one case at a time, one jurisdiction at a time — will be the rules that every creative industry lives under for the next decade.
Sources: Norton Rose Fulbright, AI in Litigation Series (March 2026); Jenner & Block, Thomson Reuters v. Ross Intelligence Analysis (February 2025); Bird & Bird, GEMA v. OpenAI Ruling Analysis (November 2025); Lexology / Norton Rose Fulbright, Germany Delivers Landmark Copyright Ruling (November 2025); AI Lawsuit Tracker (April 2026 update); Sterne Kessler, AI IP Year in Review (January 2026); GlobalLawLists.org, 10 Most Consequential AI Legal Rulings (March 2026); William Fry, GEMA v. OpenAI — Europe’s Direction (November 2025); Best Law Firms / RSSM, AI’s War in the Courtroom (December 2025); Columbia Law School, GEMA v. OpenAI Doctrinal Analysis (March 2026).
