For years now, AI corporations, together with Google, Meta, Anthropic, and OpenAI, have insisted that their giant language fashions aren’t technically storing copyrighted works of their reminiscence and as an alternative “study” from their coaching knowledge like a human thoughts.
It’s a fastidiously worded distinction that’s been integral to their makes an attempt to defend themselves towards a quickly growing barrage of legal challenges.
It additionally cuts to the core of copyright legislation itself. Copyright is a type of mental property legislation designed to guard authentic works and their creators. Below the US Copyright Act of 1976, a copyright proprietor has the unique proper to “reproduce, adapt, distribute, publicly carry out, and publicly show the work.”
However, crucially, the “fair use” doctrine holds that others can use copyrighted supplies for functions like criticism, journalism, and analysis. That’s been the AI trade’s protection in court towards accusations of infringement; OpenAI CEO Sam Altman has gone so far as to say that it’s “over” if the trade isn’t allowed to freely leverage copyrighted knowledge to coach its fashions.
Rights holders have lengthy cried foul, accusing AI corporations of coaching their fashions on pirated and copyrighted works, successfully monetizing them with out ever pretty remunerating authors, journalists, and artists. It’s a years-long authorized battle that’s already led to a high-profile settlement.
Now, a damning new study may put AI corporations on the defensive. In it, Stanford and Yale researchers discovered compelling proof that AI fashions are literally copying all that knowledge, not “studying” from it. Particularly, 4 outstanding LLMs — OpenAI’s GPT-4.1, Google’s Gemini 2.5 Professional, xAI’s Grok 3, and Anthropic’s Claude 3.7 Sonnet — fortunately reproduced prolonged excerpts from fashionable — and guarded — works, with a shocking diploma of accuracy.
They discovered that Claude outputted “total books near-verbatim” with an accuracy fee of 95.8 p.c. Gemini reproduced the novel “Harry Potter and the Sorcerer’s Stone” with an accuracy of 76.8 p.c, whereas Claude reproduced George Orwell’s “1984” with a better than 94 p.c accuracy in comparison with the unique — and nonetheless copyrighted — reference materials.
“Whereas many imagine that LLMs don’t memorize a lot of their coaching knowledge, latest work reveals that substantial quantities of copyrighted textual content may be extracted from open-weight fashions,” the researchers wrote.
A few of these reproductions required the researchers to jailbreak the fashions with a way called Best-of-N, which basically bombards the AI with completely different iterations of the identical immediate. (These sorts of workarounds have already been utilized by OpenAI to defend itself in a lawsuit filed by the New York Times, with its lawyers arguing that “regular folks don’t use OpenAI’s merchandise on this approach.”)
The implications of the most recent findings could possibly be substantial as copyright lawsuits play out in courts throughout the nation. As The Atlantic‘s Alex Reisner points out, the outcomes additional undermine the AI trade’s argument that LLMs “study” from these texts as an alternative of storing data and recalling it later. It’s proof that “could also be a large authorized legal responsibility for AI corporations” and “doubtlessly price the trade billions of {dollars} in copyright-infringement judgments.”
Whether or not AI corporations are answerable for copyright infringement stays a topic of heated debate. Stanford legislation professor Mark Lemley, who has represented AI corporations in copyright lawsuits, advised The Atlantic that he isn’t positive whether or not an AI mannequin “accommodates” a replica of a guide or can reproduce it “on the fly in response to a request.”
Unsurprisingly, the trade is constant to argue that they’re technically not replicating protected works. In 2023, Google told the US Copyright Office that “there isn’t a copy of the coaching knowledge — whether or not textual content, pictures, or different codecs — current within the mannequin itself.”
OpenAI additionally advised the workplace in the identical yr that its “fashions don’t retailer copies of the knowledge that they study from.”
To The Atlantic‘s Reisner, the analogy that AI fashions study like people is a “misleading, feel-good thought that forestalls the general public dialogue we have to have about how AI corporations are utilizing the inventive and mental works upon which they’re completely dependent.”
However whether or not the judges overseeing the litany of copyright lawsuits will agree with that sentiment stays to be seen. The stakes are appreciable, significantly because it turns into harder and harder for authors, journalists, and different content material creators to make a dwelling — whereas the AI trade swells to unfathomable value.
Extra on AI and copyright: OpenAI’s Copyright Situation Appears to Be Putting It in Huge Danger
