Could an AI’s plot twist be more than just fan fiction-it could be a legal liability? That’s the question now facing OpenAI and Microsoft after a federal judge allowed a high-profile class-action lawsuit from George R.R. Martin and other prominent authors to proceed, centering on claims that ChatGPT-generated outputs are “substantially similar” to their copyrighted works.
The battle is over how large language models-ones like ChatGPT, for example-are created and trained. These systems have ingested several millions of books, articles, and even more texts to learn the statistical relationship between words, syntax, and narrative structures. Along the way, text gets converted into “tokens,” numerical forms that enable algorithms to perceive patterns without retaining the original format. But plaintiffs say that when the source materials are copyrighted and ingested without permission, the process itself may constitute infringement.
Here, for example, was one telling demonstration provided by the authors’ legal team: a prompt instructing ChatGPT to outline a sequel to “A Clash of Kings” that was different from “A Storm of Swords.” The AI returned “A Dance with Shadows,” replete with new characters, systems of magic, and twists of politics, redolent of Martin’s original work. Judge Sidney Stein wrote that “a reasonable jury could find that the allegedly infringing outputs are substantially similar to plaintiffs’ works,” and noted that tone, plot, and thematic elements could cross the threshold from inspiration into unlawful reproduction.
From an engineering perspective, the similarity problem speaks to how LLMs store and reproduce learned patterns. While the developers insist that their models do not commit whole works to memory, evidence from the other lawsuits shows something rather different: under the right conditions, generative AI can indeed output text bearing a striking resemblance to its training data. This is most likely to occur when prompts are specific enough to trigger latent “memorized” sequences. Filtering, watermarking, and content provenance tracking are among techniques explored as a way of reducing this risk, but their effectiveness is still being tested.
The legal backdrop is complex. U.S. copyright law’s fair use doctrine considers four factors: purpose and character of the use, nature of the work, amount used, and market impact. In the recent rulings, judges have split on whether AI training is “transformative” enough to qualify as fair use. In the case of Bartz v. Anthropic, training on legally acquired books was considered “spectacularly transformative,” but retaining pirated copies in a permanent library was not. Whereas in the case of Meta LLaMA, another judge warned that the ability of LLMs to “flood the market” with similar works may weigh heavily against fair use if plaintiffs provide stronger evidence of market harm.
Here, Judge Stein has not ruled on fair use, and the door remains open for OpenAI and Microsoft to defend their practices as transformative. The decision may come down to technical evidence: whether ChatGPT’s outputs are competitive substitutes for the original works, and whether the model’s training involved unauthorized copies from “shadow libraries” or other illicit sources. Plaintiffs in related suits have claimed that Microsoft-backed models were trained on datasets containing nearly 200,000 pirated books, highlighting the stakes for data provenance in AI development.
This case also has broader ramifications for the industry: if the courts rule that training without permission on copyrighted works is indeed infringement, AI companies would then need to license enormous datasets, an extremely expensive and complex task considering the scale of LLM training. On the other hand, the wide fair-use ruling might preserve the current pace of AI innovation but perhaps at the expense of eroding authors’ control over their intellectual property.
Technical safeguards will probably become more central in compliance strategies going forward. Watermarking could embed invisible signals in AI outputs, tracing their origin; dataset documentation-already required under the EU’s AI Act-may mean companies publish summaries of training sources. All such steps would support legal defenses but build trust with rights holders by forcing much greater transparency.
For now, the outline of “A Dance with Shadows” serves as more than a creative exercise: it’s a test case of how courts will draw the line between inspiration and infringement in the age of generative AI. Until the question of fair use is resolved, similar lawsuits across jurisdictions are making their way forward. This battle between authors and AI developers could present a sea change in the legal and technical boundaries of machine-generated storytelling.
