Author: James D. Berkley
Amidst a level of intrigue rare to the Library of Congress and U.S. Copyright Office, on Friday, May 9, the Copyright Office released a detailed 108-page report containing its most extended discussion of how copyright law applies to the so-called “training” of generative artificial intelligence (better known as “generative AI”). Generative AI is the rapidly developing field exemplified by products such as ChatGPT, where large statistical models—using computer processes opaque even to their developers—can essentially “learn” by imbibing massive quantities of digitized information and then, through continued practice and fine-tuning, can become adept at generating seemingly intelligent, even seemingly creative, responses to human queries, even to the point of creating new works.
At present, more than 40 copyright lawsuits are pending across the United States involving generative AI, typically pitting owners and creators of copyrighted content against the tech company purveyors of generative AI. These cases pose complex questions over whether infringement occurs if copyrighted works are used to “train” generative AI models, and whether copyright owners are entitled to demand licenses and compensation for such uses.
Titled “Copyright and Artificial Intelligence, Part 3: Generative AI Training,” the Copyright Office’s report was released as an unusual “pre-publication version” on a day between two newsworthy firings. First, on May 8, the Librarian of Congress, of which the U.S. Copyright Office is part, was dismissed after serving in her role for nearly eight years. Then, the day after the report’s release, the Register of Copyrights—under whose imprimatur the report was prepared—was dismissed as well. These events prompted speculation as to whether the positions taken in the report were viewed as insufficiently pro-AI.
That said, the conclusions of the report do not blatantly favor either the pro-copyright or anti-AI camps. As the Introduction states, “The public interest requires striking an effective balance, allowing technological innovation to flourish while maintaining a thriving creative community.”
On his website, Copyright Lately, MSK intellectual property partner Aaron Moss offered his “top five takeaways” from the report. These takeaways are summarized below:
1. Generative AI Can Implicate Different Kinds of “Copying”
As the report details, generative AI models are trained using large “datasets” that can involve making multiple copies of copyrighted works. Works must be digitized, formatted, transferred, and combined, and a completed dataset may be reproduced many times over. To no one’s surprise, the Copyright Office report concludes that “[t]he steps required to produce a training dataset containing copyrighted works clearly implicate the right of reproduction.” Unless these steps are defensible as “fair use,” such copying may constitute infringement.
But as the report notes, downstream steps raise the issue of copying as well. If a model is trained on a set of works, and then creates output that strongly resembles one or more of those works, the output could be found to violate copyright too. A more subtle question is whether the trained model itself may be found infringing. The report considers whether the model’s internal “weights”—the parameters that store learned information—can embody copyrighted expression. According to the Office, if a model can produce outputs that are substantially similar to the training inputs, it has memorized protectable content, and copying or distributing these weights could therefore amount to infringement. This is significant, since it would mean that even apart from the model’s inputs and outputs, downloading or transmitting the trained model could be actionable copyright infringement in its own right.
2. Copying for AI Training May Be “Transformative,” But Context and Purpose Matters
Copying works for purposes of training AI might still be fair use. Section 107 of the U.S. Copyright Act provides a four-factor test to consider if use of a copyrighted work is “fair.” The first factor looks at “the purpose and character of the use, including whether [the] use is of a commercial nature or is for nonprofit educational purposes.”
A line of Supreme Court cases instructs that even if a use is commercial, a key issue under this factor is whether the use is “transformative,” meaning that it “has a further purpose or different character” from the original. And in recent cases—such as its important 2023 decision involving the magazine-cover use of an Andy Warhol portrait based on the work of a professional photographer, Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023)—the Court has instructed that courts look very carefully at the specific facts surrounding the use in question and its purpose.
The Copyright Office report follows suit, noting that for generative AI training, “[f]air use must . . . be evaluated in the context of the overall use.” At one end of a spectrum, copyrighted works could be used in training for purposes having nothing to do with creating new works as output, let alone ones that reproduce the training data. At the other end, AI training might be used to generate works that, in copyright parlance, are “substantially similar” to the original works without any noticeably different purpose, which is not transformative and does not support fair use. The report adds that another factor is whether AI systems can deploy “guardrails” that prevent output from copying protected expression. If they can, using copyrighted works for training may be more transformative, and more likely to qualify as fair use. Finally, the report offers that the source of copyrighted works can also play a role under the first factor, stating that “the knowing use of a dataset that consists of pirated or illegally accessed works should weigh against fair use without being determinative.”
