Faris Adani
Researchers Just Found Something That Could Shake the AI Industry to Its Core
For years, leading artificial intelligence developers such as Google, Meta, Anthropic, OpenAI, and xAI have maintained a critical distinction regarding their large language models (LLMs): they do not technically *store* copyrighted works but rather *learn* from their vast training data in a manner analogous to a human mind. This carefully articulated defense has been a cornerstone of their strategy against a rapidly escalating wave of legal challenges globally, cutting directly to the fundamental principles of copyright law itself. Under the US Copyright Act of 1976, copyright owners possess exclusive rights, including the ability to reproduce, adapt, distribute, publicly perform, and display their original works. However, the “fair use” doctrine provides a crucial exception, allowing the use of copyrighted material for purposes like criticism, commentary, news reporting, teaching, scholarship, or research. This doctrine has been the primary shield for the AI industry, with figures like OpenAI CEO Sam Altman controversially stating that the industry’s future is “over” if it cannot freely leverage copyrighted data for training its models.
This ongoing debate has pitted the immensely valuable AI industry against a coalition of rights holders – authors, journalists, artists, musicians, and other content creators – who accuse AI companies of exploiting their pirated and copyrighted works without proper remuneration. They argue that AI models are effectively monetizing their creative output, leading to a “years-long legal battle” that has already seen significant developments, including a high-profile settlement by Anthropic. The core of this dispute lies in whether the act of training an AI model on copyrighted data, and its subsequent ability to generate output that resembles or even reproduces that data, constitutes copyright infringement or falls under fair use. AI companies typically argue that their models transform the data, creating something new, much like a human learning from various sources.
However, a groundbreaking new study, recently published by researchers from Stanford and Yale, threatens to dismantle this carefully constructed defense and fundamentally alter the landscape of AI copyright litigation. The study presents compelling evidence that certain prominent AI models are not merely “learning” from data but are, in fact, *copying* and reproducing substantial portions of copyrighted works with alarming accuracy. The researchers specifically tested four leading LLMs: OpenAI’s GPT-4.1, Google’s Gemini 2.5 Pro, xAI’s Grok 3, and Anthropic’s Claude 3.7 Sonnet. Their findings revealed that these models could reproduce lengthy excerpts from popular and protected literary works with a stunning degree of fidelity. For instance, Claude was found to output “entire books near-verbatim” with an accuracy rate of 95.8 percent. Gemini replicated the novel “Harry Potter and the Sorcerer’s Stone” with an accuracy of 76.8 percent, while Claude reproduced George Orwell’s dystopian classic “1984” with over 94 percent accuracy compared to the original, still-copyrighted reference material.
The researchers explicitly noted that “While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models.” This direct contradiction to the industry’s long-held assertion is what makes the study so impactful. It suggests that the models’ internal mechanisms might not be as analogous to human learning as previously claimed, raising serious questions about the nature of their data processing.
It is important to note that some of these verbatim reproductions required the researchers to employ a “jailbreaking” technique known as Best-of-N. This method involves bombarding the AI with numerous iterations of the same prompt to elicit a desired response, pushing the model beyond its intended guardrails. While this might seem like an artificial scenario, AI companies have previously used similar arguments in their defense. For example, in the lawsuit filed by *The New York Times*, OpenAI’s lawyers argued that “normal people do not use OpenAI’s products in this way,” implying that any infringing output was the result of deliberate manipulation rather than inherent memorization. However, the very fact that such detailed, accurate reproductions *can* be extracted, even with specific prompting techniques, suggests a level of internal retention that challenges the “not storing” narrative.
The implications of these latest findings are profound and could reverberate through the numerous copyright lawsuits currently unfolding across the country. As Alex Reisner of *The Atlantic* cogently observes, these results significantly undermine the AI industry’s core argument that LLMs “learn” from texts rather than storing and recalling information. This evidence, if accepted by courts, “may be a massive legal liability for AI companies” and could “potentially cost the industry billions of dollars in copyright-infringement judgments.” The fair use doctrine, which hinges on factors like the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use upon the potential market, becomes much harder for AI companies to invoke when near-verbatim reproduction is demonstrable. If a model can output entire books, it directly challenges the notion of “transformative” use or minimal copying.
Despite the mounting evidence, the question of whether AI companies are liable for copyright infringement remains a subject of intense debate. Stanford law professor Mark Lemley, who has represented AI companies in copyright lawsuits, acknowledges the complexity, telling *The Atlantic* that he is “not sure whether an AI model ‘contains’ a copy of a book or can reproduce it ‘on the fly in response to a request.'” This highlights the technical and legal ambiguity that judges will need to navigate. Unsurprisingly, the industry continues to adhere to its original stance. In 2023, Google informed the US Copyright Office that “there is no copy of the training data — whether text, images, or other formats — present in the model itself.” OpenAI echoed this sentiment in the same year, stating that its “models do not store copies of the information that they learn from.”
To critics like *The Atlantic*’s Reisner, the analogy that AI models learn like humans is a “deceptive, feel-good idea that prevents the public discussion we need to have about how AI companies are using the creative and intellectual works upon which they are utterly dependent.” This perspective emphasizes the ethical and societal dimensions of the debate, beyond the purely legalistic arguments. The outcome of these lawsuits will not only determine the financial future of the AI industry but also set precedents for the protection of intellectual property in the digital age. The stakes are undeniably considerable, particularly as content creators, including authors and journalists, face increasing difficulty in making a living in a rapidly changing media landscape, while the AI industry continues to swell to unfathomable valuations.
The new study, by providing concrete evidence of reproduction rather than mere learning, has introduced a critical piece of the puzzle that could fundamentally shift the legal and public perception of AI’s relationship with copyrighted material. Whether judges ultimately agree with the researchers’ findings and the implications they carry for copyright infringement remains to be seen, but the AI industry now finds itself on a far more precarious footing. The era of claiming AI models are just “learning” like humans may be drawing to a close, forcing a reckoning with how these powerful technologies are built and sustained.
More on AI and copyright: *OpenAI’s Copyright Situation Appears to Be Putting It in Huge Danger*
