Today I filed an amicus brief in In re OpenAI, Inc. Copyright Infringement Litigation.
There are a lot of different generative AI copyright lawsuits floating around, but this is arguably one of the most important. It is a multidistrict litigation in the U.S. District Court for the Southern District of New York that consolidates copyright suits brought against OpenAI and Microsoft by The New York Times, the Daily News and other newspapers, The Intercept, the Center for Investigative Reporting, Ziff Davis, and a proposed class of book authors led by the Authors Guild.
The cases are before Judge Sidney H. Stein, who is now considering cross-motions for summary judgment on whether training large language models on copyrighted books and news articles is fair use. The ruling will be the first on that question in the Second Circuit and the first anywhere to address the claims of news publishers.
You can read my brief here, but allow me to suggest what might be the most interesting points.
The TLDR
The brief offers a framework for applying established copyright fair use principles to generative AI training. Although the brief formally supports neither party—for reasons that will become apparent shortly—it argues that training LLMs is highly transformative because the purpose of copying copyrighted works is to learn statistical patterns and relationships rather than reproduce their protected expression for a new audience.
The brief develops six central arguments.
(1) AI training is transformative under existing precedent
We have reached the point where this is so obvious it almost goes without saying. To be sure, the training pipeline begins with several acts of literal copying, but the training process itself is not copying in any conventional sense. During training, a model that is essentially a blank slate is exposed to training data, and that data influences the parameters of the model. But the influence of any single piece of data on the model is typically microscopic. Most importantly, what comes out of the model once it has been trained typically doesn’t look anything remotely like a substantially similar copy of any piece of training data. The entire process writ large is transformative in both a legal and descriptive sense.
(2) Real-world outputs matter more than theoretical memorization
We have learned a lot about memorization over the past few years, but it’s still difficult to pinpoint exactly how much AI models memorize and under what conditions. Memorization is pretty clearly very small in percentage terms, but at the scale at which AI models operate, it may not be small in nominal terms (0.001% of trillions of tokens is still a big number). This is interesting for researchers, but mostly irrelevant for the fair use inquiry because what matters is whether models ordinarily reproduce protected expression, not what secrets the model might be keeping to itself, or whether determined researchers can extract fragments through extensive adversarial prompting.
(3) Market dilution is a losing theory
Plaintiffs are unlikely to be able to establish any form of harm that the fair use doctrine has traditionally cared about. If the model outputs look nothing like the inputs, then there is obviously no harm from expressive substitution, and courts have consistently held that there is no intrinsic right to license transformative uses.
My brief points out that the novel argument that somehow generalized competition from AI-generated content counts as harm under the fourth fair use factor is both unpersuasive in theory and will be impossible to prove in practice. The market dilution argument is speculation about collective harms that sits oddly with a copyright regime focused on individual rights with respect to particular works.
(4) Rights clearance for internet-scale training data is impossible, not just difficult
I also argue that comprehensive rights clearance for internet-scale training data is practically impossible. Not just difficult, time-consuming, and expensive, but literally impossible. The problem is that there is a fundamental mismatch between the enormous collective value of training data and the negligible marginal contribution of any individual work. The cost of identifying rights holders, negotiating licenses, and administering payments would routinely exceed the value of the permissions obtained. People look at institutions like ASCAP and think it shows that licensing at scale is possible, but that’s because they don’t understand the difference between distributing money to a few million people and what it would take to distribute money to hundreds of millions or billions of creators. Agreements with major rights holders can’t solve this problem: they demonstrate viable markets for access to valuable content, not for obtaining permissions across the entire internet. There’s no way for the court to change this reality just by wishing a licensing market into existence.
Despite all of the foregoing, my brief acknowledges a couple of issues where the court will have to make a call on the evidence, and that call could go against the defendants.
(5) Some indirect competitive harms may nevertheless matter
My brief recognizes that using extracted information to create a direct substitute for a particular source could undermine the economic incentives for producing that source’s original content, and that this kind of indirect expressive substitution is something the court could weigh under the fourth fair use factor (which generally considers the effect of the use on the market for or the value of the copyrighted work).
(6) The acquisition of training materials from shadow libraries could count against fair use under the fourth factor
Obtaining copyrighted books from shadow libraries may cause legally relevant harm when it bypasses an available market for legitimate access. The court should reject the approach suggested in Bartz, where the acquisition of those materials was characterized as a separate use—at least when it is part of an integrated training process. As I explain in my brief, this is really an issue under the fourth factor because the question is whether the developer bypassed an existing, economically meaningful market for lawful access to copyrighted works. Fair use does not categorically require lawful acquisition, and it’s significant that at the relevant time, publishers generally did not offer digital book collections for computational training. So the central question is whether purchasing and scanning millions of physical books constituted a reasonable alternative to obtaining them from shadow libraries. I think that is a factual determination the court needs to make.
Summing up
I think the core of the plaintiffs’ case is weak: training is transformative, memorization is a sideshow, and market dilution is not a cognizable harm. But there one or two judgment calls that the court needs to make before the defendants are entitled to a clean sweep. The shadow library question and the risk of indirect substitution turn on facts the court has to evaluate and given the complexity of the record, it’s not helpful for an amicus brief to pretend to be in a position to make that evaluation.
My hope is that the court will decide the easy questions clearly and the hard questions carefully, rather than treating the whole thing as all-or-nothing.