In the first two parts of this series I examined how large language models (LLMs) work and analyzed whether their training process can be justified under copyright law’s fair use doctrine. (Part 1, Part 2). However, I also noted that LLMs sometimes “memorize” content during the training stage and then “regurgitate” that content verbatim or near-verbatim when the model is accessed by users.
The “memorization/regurgitation” issue is featured prominently in The New York Times Company v. Microsoft and OpenAI, pending in federal court in the Southern District of New York. (A reference to OpenAI includes Microsoft, unless the context suggests otherwise). Because the technical details of every LLM and AI model are different I’m going to focus my discussion of this issue mostly on the OpenAI case. However, the issue has implications for every generative LLM trained on copyrighted content without permission.
The Controversy
Here’s the problem faced by OpenAI.
OpenAI claims that the LLM models it creates are not copies of the training data it uses to develop its models, but uncopyrightable patterns. In other words, a ChatGPT LLM is not a conventional database or search engine that stores and retrieves content. As OpenAI’s co-defendant Microsoft explains, “A program called a ‘transformer’ evaluates massive amounts of text, converts that text into trillions of constituent parts, discerns the relationships among all of them, and yields a natural language machine that can respond to human prompts.” (Link, p. 2)
However, the Times has presented compelling evidence that challenges this narrative. The Times showed that it was able to prompt OpenAI’s ChatGPT-4 and Microsoft’s CoPilot to produce lengthy, near-verbatim excerpts from specific Times articles, which the Times then cited in its complaint as proof of infringement.
The Times’ First Amended Complaint includes an exhibit with over 100 examples of ChatGPT “regurgitating” Times content verbatim or near-verbatim in response to specific prompts (copied text highlighted in red; click to enlarge):
This evidence poses a fundamental question: If OpenAI’s models truly transform copyright-protected content into abstract patterns rather than storing it, how can they reproduce exact or nearly exact copies of that content?
The Times argues that this evidence reveals a crucial truth: actual copyrighted expression—not just abstract patterns—is encoded within the model’s parameters. This allegation strikes at the foundation of OpenAI’s legal position and weakens its fair use defense by suggesting its use of copyrighted material is more extensive and less transformative than claimed.
Just how big a problem this is for OpenAI and the AI industry is difficult to determine. I’ve tried to replicate it in a variety of cases on ChatGPT and several other frontier models without success. In fact I can’t get the models to give me the text of Moby Dick, Tom Sawyer or other literary works whose copyrights have long expired.
Nevertheless, the Times was able to do this one hundred times, and it’s safe to assume that it could have continued well past that number, but thought that 100 examples was enough to make the point in its lawsuit.
OpenAI: “The Times Hacked ChatGPT”
What’s OpenAI’s response to this?
To date, OpenAI and Microsoft have not filed answers to the Complaint. However, they have given an indication of how they view these allegations in partial motions to dismiss filed by both companies.
Microsoft’s motion (p. 2) argues that the NYT’s methods to demonstrate how its content could be regurgitated did not represent real-world usage of the GPT tools at issue. “The Times,”it argues, “crafted unrealistic prompts to try to coax the GPT-based tools to output snippets of text matching The Times’s content.” (Emphasis in original) To get the NYT content regurgitated, a user would need to know the “genesis of that content.” “And in any event, the outputs the Complaint cites are not copies of works at all, but mere snippets” that do not rise to the level of copyright infringement.
OpenAI’s motion (p. 12.) argues that the NYT “appears to have [used] prolonged and extensive efforts to hack OpenAI’s models”:
In the real world, people do not use ChatGPT or any other OpenAI product for that purpose, … Nor could they. In the ordinary course, one cannot use ChatGPT to serve up Times articles at will. . . . The truth, which will come out in the course of this case, is that the Times paid someone to hack OpenAI’s products. It took them tens of thousands of attempts to generate the highly anomalous results that make up Exhibit J to the Complaint. They were able to do so only by targeting and exploiting a bug (which OpenAI has committed to addressing) by using deceptive prompts that blatantly violate OpenAI’s terms of use. And even then, they had to feed the tool portions of the very articles they sought to elicit verbatim passages of, virtually all of which already appear on multiple public websites. Normal people do not use OpenAI’s products in this way.
It appears that OpenAI is referring to the provisions in its Terms of Service that prohibit anyone from “Us[ing] our Services in a way that infringes, misappropriates or violates anyone’s rights” or to “extract data.” OpenAI has labeled these “adversarial attacks.”
Copyright owners don’t buy this tortured explanation. As OpenAI has admitted in a submission to the Patent and Trademark Office, “An author’s expression may be implicated . . . because of a similarity between her works and an output of an AI system.” (link, n. 71, emphasis added).
Rights holders claim that their ability to extract memorized content from these systems puts the lie to, for example, OpenAI’s assertion that “an AI system can eventually generate media that shares some commonalities with works in the corpus (in the same way that English sentences share some commonalities with each other by sharing a common grammar and vocabulary) but cannot be found in it.” (link, p. 10).
Moreover, OpenAI’s “hacking” defense would seem to inadvertently support the Times’ position. After all, you cannot hack something that isn’t there. The very fact that this content can be extracted, regardless of the method, suggests it exists in the form of an unauthorized reproduction within the model.
OpenAI: “We Are Protected by the Betamax Case”
How will OpenAI and Microsoft respond to these allegations under copyright law?
To date, OpenAI and Microsoft have yet to file formal answers to the Times’ complaint. However, they have given us a hint of their defense strategy in their motions to dismiss, and it is based in part on the Supreme Court’s 1984 decision in Sony v. Universal City Studios, a case often referred to as “the Betamax case.”
In the Betamax case a group of entertainment companies sued Sony for copyright infringement, arguing that consumers used Sony VCRs to infringe by recording programs broadcast on television. The Supreme Court held that Sony could not be held contributorily liable for infringements committed by VCR owners. “[T]he sale of copying equipment . . . does not constitute contributory infringement if the product is . . . capable of substantial noninfringing uses.”
The take-away from this case is that under copyright law if a product can be put to either a legal or illegal purpose by end-users (a “dual-use”), it is not infringing so long the opportunity for noninfringing use is substantial.
OpenAI and Microsoft assert that the Betamax case applies because, like the VCR, ChatGPT is a “dual-use technology.” While end users may be able to use “adversarial prompts” to “coax” a model to produce a verbatim copy of training data, the system itself is a neutral, general-purpose tool. In most instances it will be put to a non-infringing use. Citing the Betamax case Microsoft argues that “copyright law is no more an obstacle to the LLM than it was to the VCR (or the player piano, copy machine, personal computer, internet, or search engine)”—all dual-use technologies.
No doubt, in support of this argument OpenAI will place strong emphasis on ChatGPT’s many important non-infringing uses. The model can create original content, analyze public domain texts, process user-provided content, educate, generate software code, and more.
However, OpenAI’s reliance on the Betamax dual-use doctrine faces a challenge central to the doctrine itself. The Betamax case was based on secondary liability—whether Sony could be held responsible for consumers using VCRs to record television programs. The alleged infringements occurred through consumer action, not through any action taken by the device’s manufacturer.
But with generative LLMs such as ChatGPT the initial copying happens during training when the model memorizes copyrighted works. This is direct infringement by the AI company itself, not secondary infringement based on user prompts. When an AI company creates a model that memorizes and can reproduce copyrighted works, the company itself is doing the copying—making this fundamentally different from Betamax.
Before leaving this topic it’s important to note that the full scope of memorization within AI models of GPT-4’s scale may be technically unverifiable. While the models’ creators can detect some instances of memorization through testing, due to the scale and complexity of the models they cannot comprehensively examine their internal representations to determine the full extent of memorized copyrighted content. While the Times’ complaint demonstrated one hundred instances of verbatim copying, this could represent just the tip of the iceberg, or conversely, the outer limit of the problem. This uncertainty itself poses a significant challenge for courts attempting to apply traditional copyright principles.
Technical Solutions
While these legal issues work their way through the courts, AI companies aren’t standing still. They recognize that their long-term success may depend on their ability to prevent or minimize memorization, regardless of how courts ultimately rule on the legal issues.
Their approaches to this challenge vary. OpenAI has told the public that it is taking measures to prevent the types of copying illustrated in the Times’ lawsuit: “we are continually making our systems more resistant to adversarial attacks to regurgitate training data, and have already made much progress in our recent models.” (link) This includes filtering or modifying user prompts to reject certain requests before they are submitted as prompts to the model and aligning the models to refuse to produce certain types of data. Try asking ChatGPT to give you the lyrics to Arlo Guthrie’s “Alice’s Restaurant Massacree” or Taylor Swift’s “Cruel Summer.” It will tell you that copyright law prohibits it from doing so.
And, it’s important to note that different AI companies are taking different approaches to this problem. For example, Google (which owns Gemini) uses supervised fine tuning (explained here). Anthropic (which owns Claude) focuses on what it calls “constitutional AI” – a training methodology that builds in constraints against certain behaviors, including the reproduction of copyrighted content. (link here). Meta (LLaMA models) has implemented what it calls “deduplication” during the training process – actively removing duplicate or near-duplicate content from training data to reduce the likelihood of memorization. Additionally, Meta has developed techniques to detect and filter out potential memorized content during the model’s response generation phase. (link here).
Conclusion
The AI industry faces a fundamental challenge that sits at the intersection of technology and law. Current research suggests that some degree of memorization may be inherent to large language models – raising a crucial question for courts: If memorization cannot be eliminated without sacrificing model performance, how should copyright law respond?
The answer could reshape both AI development and copyright doctrine. AI companies may need to accept reduced performance in exchange for legal compliance, while content creators must decide whether to license their works for AI training despite the risk of memorization. The industry’s ability to develop systems that truly learn patterns without memorizing specific expressions – or courts’ willingness to adapt copyright law to this technological reality – may determine its future.
The outcome of the Times lawsuit may establish crucial precedents for how copyright law treats AI systems that can memorize and reproduce protected content. At stake is not just the legality of current AI models, but the broader question of how to balance technological innovation with the rights of content creators in an era where the line between learning and copying has become increasingly blurred.