The year 2026 will be remembered as the moment when the legal architecture governing intellectual property and artificial intelligence stopped being theoretical and became consequential. Three landmark decisions — one from the United States Court of Appeals for the Third Circuit, one from the Munich Regional Court, and one from the Delhi High Court — have produced a fragmented but increasingly legible map of how different jurisdictions are resolving the foundational questions of AI and copyright. Simultaneously, the $1.5 billion settlement in Bartz et al. v. Anthropic PBC has established a financial benchmark for the cost of getting training data provenance wrong. And the USPTO's revised inventorship guidance has clarified, with unusual directness, where the law stands on the question of AI as inventor.
This research article examines the current state of AI intellectual property law across these three domains — training data copyright, output liability, and inventorship — drawing on the primary legal sources and analysing the structural tensions that remain unresolved. The picture that emerges is one of genuine legal evolution, but also of significant jurisdictional divergence that creates both risk and opportunity for organisations operating across borders.
Training Data Copyright: A Jurisdictional Patchwork
The question of whether using copyrighted material to train an AI model constitutes infringement has been litigated across multiple jurisdictions in 2026, with results that reflect fundamentally different legal philosophies rather than mere technical disagreements.
The United States: Thomson Reuters v. ROSS Intelligence
The most significant U.S. development came in September 2026, when the Court of Appeals for the Third Circuit affirmed that ROSS Intelligence had infringed Thomson Reuters' copyrights by using Westlaw headnotes to train its AI legal research tool. The decision, which upheld the district court's finding, established several principles that will shape U.S. AI copyright jurisprudence for years to come.
The court found that the headnotes possessed the requisite originality for copyright protection — a finding that has implications for the broader question of what kinds of curated, structured data can be protected. More significantly, it concluded that ROSS's use was "minimally transformative, at best," because the AI tool was designed to serve as a direct commercial substitute for Westlaw. The market harm factor — the potential damage to the licensing market for training data — weighed heavily against fair use.
Legal observers have been careful to note what the decision does not resolve. The ROSS case involved a non-generative AI tool that returned existing text passages rather than creating new content. The court explicitly acknowledged that its analysis might not apply to general-purpose generative AI models, where the relationship between training data and output is more attenuated. As the LexBlog analysis published in November 2026 notes, the decision "leaves open the broader question of how fair use applies to general-purpose generative AI models" — which is, of course, the question that matters most for the industry.
The ROSS decision establishes that training data provenance is a legal liability, not merely an ethical preference — and that the market for licensing training data is one courts will protect.
Germany: GEMA v. Suno
The ROSS decision establishes that training data provenance is a legal liability, not merely an ethical preference — and that the market for licensing training data is one courts will protect.
The Munich Regional Court's July 2026 ruling in GEMA v. Suno takes a markedly different approach, and its implications extend well beyond the music industry. The court found that Suno's AI music generation models had "memorised" protected musical works — meaning the models contained reproducible versions of the training data — and that the storage of these models on servers in Germany constituted infringement of the reproduction right under German and EU copyright law.
Two aspects of the ruling are particularly significant. First, the court's extraterritorial reach: despite the training having taken place in the United States, the court granted injunctions prohibiting both the use of works to train models and the storage of those works within models. The legal basis was that the models themselves, stored on German servers, constituted infringing copies. This reasoning, if adopted more broadly, would mean that any AI model trained on copyrighted material and deployed in the EU could be subject to EU copyright law regardless of where training occurred.
Second, the court performed its own analysis of U.S. fair use law for the training conducted in the U.S., concluding that it was inapplicable because the AI service acted as a direct, substitutable competitor to the original musical works. This is a remarkable exercise of jurisdictional reach — a German court applying and rejecting U.S. fair use doctrine as part of its own analysis. The Bristows analysis of the decision notes that this approach "effectively provides extra-territorial effect to the injunction," creating a compliance challenge for any AI company with European users.
India: ANI Media v. OpenAI
The Delhi High Court's July 2026 decision in ANI Media v. OpenAI provides a striking contrast. The court dismissed an interim injunction application, finding that OpenAI's use of copyrighted news articles for training was prima facie "fair dealing" under Indian law. The reasoning drew heavily on the 2004 Canadian Supreme Court decision in CCH Canadian Ltd v. Law Society of Upper Canada, which established that fair dealing is a "user right" that should be interpreted broadly, extending even to commercial research contexts.
The court's finding that there was no evidence of significant market harm to the plaintiff distinguishes the Indian approach from both the U.S. and German decisions. Where the Third Circuit and the Munich court both emphasised the potential for AI training to damage the market for the original works, the Delhi High Court found no such harm demonstrated. This reflects a broader difference in how Indian copyright law balances the rights of creators against the interests of users and innovators.
The practical implication is that the same AI training activity — using copyrighted news articles to train a large language model — may be lawful in India, unlawful in Germany, and subject to a complex fair use analysis in the United States. For organisations operating globally, this jurisdictional divergence creates a compliance challenge that cannot be resolved by reference to any single legal framework.
The $1.5 Billion Benchmark: Bartz v. Anthropic
While the appellate decisions have been shaping the doctrinal landscape, the settlement in Bartz et al. v. Anthropic PBC has established the financial stakes with unusual clarity. The $1.5 billion settlement, approved by U.S. District Judge Araceli Martínez-Olguín on 20 July 2026, is the largest known payout in the history of U.S. copyright litigation — a figure that concentrates the mind of any organisation building AI systems on the basis of web-scraped or otherwise unlicensed content.
The settlement's structure reveals the specific legal theory that proved most dangerous for Anthropic. The 2025 summary judgment decision in the same case found that while the use of books to train Claude's large language models was "exceedingly transformative" and potentially protected under fair use, the act of acquiring and storing those works — often sourced from shadow libraries such as Library Genesis and Pirate Library Mirror — was not protected. The distinction is crucial: it is not the training itself that created the primary liability, but the provenance of the training data.
The $1.5 billion Anthropic settlement establishes that the provenance of training data is a legal liability of the first order — and that "transformative use" is not a shield against liability for how data was acquired.
The $1.5 billion Anthropic settlement establishes that the provenance of training data is a legal liability of the first order — and that 'transformative use' is not a shield against liability for how data was acquired.
The settlement terms reflect this analysis. The scope of release is deliberately narrow, covering only claims related to Anthropic's past acquisition and copying of works through August 2025, and explicitly excluding claims concerning AI outputs and future conduct. As part of the non-monetary relief, Anthropic is required to destroy all original files obtained from the specified pirate websites. The average recovery of approximately $3,000 per work — four times the statutory minimum for ordinary infringement — signals that courts are prepared to treat AI training data violations as serious rather than technical infringements.
The Legal 500 analysis published in November 2026 draws the appropriate conclusion: the settlement "serves as a warning to developers that the provenance of training data is subject to increasing scrutiny and that companies may face significant financial exposure regardless of their model's eventual output capabilities." The quality of the model is not a defence against liability for how its training data was obtained.
Inventorship: The Human Requirement Holds
The third major domain of AI intellectual property law — the question of whether AI systems can be named as inventors on patents — has been resolved with unusual clarity in the United States, even as it remains contested in other jurisdictions.
The legal position in the U.S. was definitively established by the Federal Circuit's 2022 decision in Thaler v. Vidal, which held that the Patent Act's reference to an "individual" refers to a human being. The Supreme Court's subsequent refusal to hear the case left this holding intact. In November 2025, the USPTO issued revised inventorship guidance that rescinded its February 2024 guidance and emphasised that no special or modified standards exist for AI-assisted inventions — the same traditional legal standards for inventorship apply regardless of whether AI was used in the process.
The practical implications of this position are becoming clearer as AI-assisted invention becomes more common. Nixon Peabody's November 2026 analysis identifies what it calls the "poison pill" risk: if a patent contains claims lacking human inventorship — because the AI, rather than a human, actually conceived the invention — the entire patent may be invalid, not merely the affected claims. This remains an area of active legal debate following the 2026 Federal Circuit decision in Fortress Iron, LP v. Digger Specialties, Inc., but the risk is real enough to require careful documentation practices.
The USPTO's current guidance treats AI systems as tools that assist human inventors, analogous to laboratory equipment or software. For an invention to be patentable, a human being must have performed "conception" — the mental formation of a definite and permanent idea of the complete and operative invention. If a human has not significantly contributed to the conception of the invention, the subject matter may be found ineligible for patent protection.
The Documentation Imperative
The practical response to this legal framework is a documentation imperative that many organisations have not yet fully internalised. Patent applicants are encouraged to document human contributions to substantiate that a natural person — not the AI — conceived the invention. This includes evidence that a human provided specific instructions, refined AI-generated outputs, or performed post-AI inventive work that meets the conception standard.
The challenge is that the boundary between "using AI as a tool" and "having AI conceive the invention" is not always clear in practice. When a researcher prompts an AI system with a specific technical problem and the AI generates a novel solution that the researcher then validates and refines, who conceived the invention? The USPTO's guidance suggests that the answer depends on the specificity and creativity of the human's contribution — but this is a fact-intensive inquiry that will generate litigation for years to come.
The inventorship question is not merely a legal technicality — it is a fundamental question about the nature of human creativity and its relationship to machine intelligence that patent law is only beginning to address.
The inventorship question is not merely a legal technicality — it is a fundamental question about the nature of human creativity and its relationship to machine intelligence that patent law is only beginning to address.
The Structural Tensions
Across all three domains — training data copyright, output liability, and inventorship — the same structural tensions recur. The first is the tension between innovation and protection: IP law exists to incentivise creation by granting temporary monopolies, but AI systems trained on existing creative works challenge the assumption that the creator of the training data and the creator of the AI output are different parties with different interests.
The second tension is jurisdictional. The internet does not respect national borders, but copyright law does. An AI model trained in the United States, deployed globally, and generating outputs consumed in Germany is subject to at least three different legal frameworks simultaneously. The Munich court's willingness to apply German copyright law to training that occurred in the U.S. suggests that this jurisdictional complexity will only increase as courts assert extraterritorial reach.
The third tension is temporal. IP law was designed for a world in which the creation of new works was a relatively slow, human-intensive process. AI systems can generate thousands of outputs per second, each of which may raise copyright questions. The legal system's capacity to adjudicate these questions at scale is genuinely uncertain.
Implications for Organisations
The research findings point to several practical implications for organisations building or deploying AI systems.
First, training data provenance is now a legal liability of the first order. The Anthropic settlement makes clear that the source of training data — not just its use — is subject to legal scrutiny. Organisations that cannot demonstrate the lawful provenance of their training data face exposure that is not mitigated by the quality or transformativeness of their models.
Second, jurisdictional strategy matters. The divergence between U.S., EU, and Indian approaches to AI copyright creates both risk and opportunity. Organisations that understand the legal landscape in each jurisdiction can structure their training, deployment, and licensing activities to minimise exposure. Those that assume a single global standard will be caught out.
Third, inventorship documentation is not optional. As AI-assisted invention becomes standard practice, the documentation of human creative contribution becomes a prerequisite for patent validity. Organisations that cannot demonstrate human conception of their AI-assisted inventions risk having their patents invalidated — a risk that is particularly acute for companies whose competitive position depends on patent protection.
Fourth, the licensing market for training data is real and growing. The Third Circuit's finding that the market for licensing training data is one courts will protect — combined with the Anthropic settlement's financial scale — signals that the era of treating web-scraped content as a free resource is ending. Organisations that invest in licensed, high-quality training data now will be better positioned as the legal framework continues to tighten.
The intellectual property landscape for AI is not yet settled. But the direction of travel is increasingly clear: towards greater protection for creators, greater scrutiny of training data provenance, and greater complexity for organisations operating across jurisdictions. The organisations that navigate this landscape successfully will be those that treat IP governance not as a legal compliance function but as a strategic capability.




