AI Copyright Law: Settlements, Rulings, and Developer Exposure

A $1.5 billion settlement, a Supreme Court refusal to hear a landmark authorship case, and a UK ruling that split liability between training data and AI outputs: by late 2026 the legal framework around AI copyright has moved fast enough that companies building or deploying AI tools face concrete exposure on multiple fronts. What follows draws on the most significant court decisions and U.S. Copyright Office guidance to date.

Training Data Liability

The central legal battleground remains whether using copyrighted material to train AI models constitutes infringement. Courts have begun drawing distinctions that matter for anyone involved in AI development.

Getty Images and the Authors Guild

In early 2023 Getty Images sued Stability AI in the UK, alleging copyright and trademark infringement over the use of millions of its images to train the Stable Diffusion model without authorisation. A parallel case was filed in the U.S. District Court for the District of Delaware. The UK High Court’s November 4, 2025 ruling largely rejected Getty’s secondary copyright infringement claim, finding that Stable Diffusion models do not contain or store reproductions of trained works and therefore are not “infringing copies” for secondary infringement purposes. The court did find limited trademark liability: where AI-generated outputs included Getty’s trademarks, responsibility lay with Stability AI, not the end user. That ruling confirms that model providers cannot redirect trademark liability to the people using their tools.

The Authors Guild, alongside writers including George R.R. Martin and Jonathan Franzen, filed a class-action against OpenAI and Microsoft in September 2023, later consolidated into In re OpenAI, Inc., Copyright Infringement Litigation. The authors allege that OpenAI’s large language models were trained on their copyrighted books without consent or compensation. OpenAI has asserted a fair use defence, arguing that model training is a transformative use. As of September 2026, the case is moving toward summary judgment. That same month, The Seattle Times and Newsday joined a growing list of publishers suing OpenAI and Microsoft, alleging their journalism was scraped to train AI models while bypassing paywalls and ignoring terms of service.

Fair Use and Its Limits

AI developers routinely invoke fair use, arguing that training on copyrighted works is transformative because it produces new technology rather than reproducing existing content. U.S. courts weigh four statutory factors: the purpose and character of the use, the nature of the copyrighted work, the amount used and the effect on the market for the original.

In cases such as Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc., arguments have been made that AI training can qualify as fair use, though courts have also indicated that output-side infringement is a separate exposure. A “market dilution” theory has been proposed in some proceedings, suggesting that AI-generated content could infringe simply by competing with copyrighted works, even without direct copying. How widely courts adopt that theory remains to be seen.

Authorship and AI Output

Separate from how AI systems are trained, courts and regulators have addressed a second question: who, if anyone, owns what an AI produces?

The Thaler Decision

On March 2, 2026, the Supreme Court denied certiorari in Thaler v. Perlmutter leaving intact lower court rulings that copyright requires human authorship. The case concerned an artwork titled “A Recent Entrance to Paradise,” which Stephen Thaler claimed was autonomously created by his “Creativity Machine” AI system. The Copyright Office had rejected his application in 2019. Both the district court and the D.C. Circuit affirmed that human authorship is a “bedrock requirement of copyright,” and the Supreme Court’s refusal to hear the appeal settled the question for now.

The practical consequence for businesses is clear. Works produced entirely by AI, with no meaningful human involvement, receive no copyright protection and can be freely copied or distributed by anyone. The Copyright Office has confirmed that applicants using AI may claim protection only for their own creative contributions and must disclose AI-generated material in registration applications. Where a human selects, arranges or substantially modifies AI-generated content, protection may apply to those contributions on a case-by-case basis. The broader evolution of AI copyright doctrine has been building toward exactly this kind of case-by-case assessment.

Output Infringement Risk

The Andersen v. Stability AI class action, filed in January 2023 by artists Sarah Andersen, Kelly McKernan and Karla Ortiz, alleges that Stable Diffusion produces “infringing derivative works” that copy the style of existing artists. Many initial claims were dismissed, but the direct copyright infringement claim against Stability AI was allowed to proceed. To survive, the plaintiffs must show both that compressed training images stored within the model constitute direct infringement and that outputs are substantially similar to the originals, a threshold that remains contested.

For organisations deploying generative AI, reviewing outputs before use is not optional. Content that closely imitates a specific artist’s style, or that reproduces recognisable passages from copyrighted works, carries real infringement risk regardless of whether the training data was licensed. Originality checking against copyright databases, clear internal guidelines on human review and documented editorial decisions in the generation process all strengthen any authorship claim if content is later challenged.

Licensing and Settlement

With litigation costs mounting on both sides, a licensing market has developed in parallel, and one landmark settlement has reset expectations about the financial stakes.

The Licensing Market Takes Shape

Since 2023, major AI companies including OpenAI, Google, Amazon, Meta and Microsoft have signed licensing, partnership and revenue-sharing agreements with publishers and data platforms. News Corp signed a deal with OpenAI reportedly worth up to $250 million over five years. Reddit disclosed $203 million in aggregate data-licensing contract value in its IPO filing. These arrangements typically cover two distinct rights: training licences, which permit use of archives to build models, and inference licences, which allow real-time content retrieval for user queries.

The U.S. Copyright Office confirmed in 2025 that AI training on copyrighted content is not automatically fair use, giving content owners more leverage in negotiations. For rights holders, licensing converts archives into recurring revenue; for AI developers, signed agreements reduce litigation exposure and provide cleaner chain of title for training data. How the per-work benchmarks emerging from litigation interact with these licensing rates will shape what the market eventually settles on.

Anthropic’s $1.5 Billion Settlement

In July 2026, a federal judge in San Francisco approved a $1.5 billion settlement between Anthropic and a group of authors, the largest known copyright settlement in U.S. history and the first major AI copyright case to end this way. The litigation began in 2024, with writers alleging Anthropic used pirated copies of their books without permission to train its Claude model. A federal judge had previously found that training on copyrighted books qualified as fair use, but also found that Anthropic had stored over 7 million pirated books in a centralised library, a separate infringement that exposed the company to potential damages in the hundreds of billions. That finding drove the settlement. More than 91% of eligible authors and publishers claimed their share, though some objected that the amount was too low.

The Anthropic settlement draws a distinction that will matter in every subsequent case: fair use for model training and liability for how training data is stored and sourced are not the same question. A company can prevail on the first and still face catastrophic exposure on the second. For AI developers still building or expanding training datasets, that gap is the sharpest lesson the litigation record has produced so far.