Is Training an AI Model Infringing? Tulane Law Professor Brian Frye Breaks Down AI's Legal Battles

 

There are currently more than 100 pending copyright infringement lawsuits against AI companies, with more sure to come. Most of those lawsuits claim that training an AI model on copyrighted works is infringing, raising fundamental questions about the fair use of copyrighted works, competition, and the public interest. Tulane Law Professor Brian L. Frye, an intellectual property and copyright law scholar who recently filed an amicus brief in the consolidated OpenAI copyright infringement litigation (New York Times v. OpenAI), explains the legal issues at stake and how they could shape the future of artificial intelligence.

  1. What is at the heart of the copyright lawsuits against AI companies?

AI companies train their generative AI models on vast datasets that may include billions or even trillions of works, many of which are protected by copyright. The copyright owners claim that AI training on copyrighted works is infringing, because AI companies have to make copies in order to train their AI models, which generate competing works. The AI companies respond that AI training is a fair use of copyrighted works, because it uses those works for the transformative purpose of creating AI models, which generate new works.

This dispute is important, because both copyright owners and AI companies see it as existential. Copyright owners are worried about competing with AI-generated content, and AI companies are worried about catastrophic infringement liability. That’s why copyright owners have filed so many lawsuits, and AI companies are defending them so vigorously. It’s also why so many third parties are filing amicus briefs.

  1. Is using copyrighted work to train an AI model fundamentally different from copying or reproducing that work?

Superficially, the copyright infringement claims are simple. The Copyright Act gives copyright owners the exclusive right to reproduce or make copies of their works, and AI companies have to copy works into a dataset in order to train an AI model. The complication is that AI companies are using those works for a new purpose that copyright owners never anticipated and are competing with copyright owners by generating new works. Training an AI model doesn’t involve using the copyrightable elements of a work, but analyzing the uncopyrightable patterns that exist across works, which is why the AI companies are asserting the fair use defense.

The purpose of copyright isn’t to benefit authors and copyright owners, but to benefit the public by encouraging the creation and distribution of new works. The fair use defense limits the scope of copyright in order to promote the public interest. Among other things, it’s supposed to ensure that copyright doesn’t prevent the development of new technologies that benefit the public. For example, copyright owners objected to the VCR, because consumers could use it to infringe, but the Supreme Court held that VCRs were protected by fair use because they had non-infringing uses, like recording something to watch later.

Critically, VCRs were a valuable new technology, and the Supreme Court didn’t want copyright to prohibit them. AI poses a similar problem. It’s obviously an immensely valuable new technology, but it technically requires a limited form of copyright infringement. The question is how the courts should reconcile copyright with the public interest in AI technology.

  1. You recently filed an amicus brief in the OpenAI copyright litigation arguing that AI training may actually benefit journalism. What is your argument?

I wrote an amicus brief with Ron Suskind, a Pulitzer-prize winning journalist and filmmaker, and the CEO of a small AI company that created an AI model intended for journalists. Our brief addresses the fair use analysis, which usually hinges on two factors: Is the accused use transformative and does it harm the market for or value of the original. Several courts have already held that training a generative AI model is transformative, so we focused on whether it harms the market for or value of copyrighted works.

We argue that training an AI model on copyrighted newspaper articles doesn’t harm the market for or value of those articles in any relevant way, because their primary value lies in attribution, not reproduction. In other words, what newspapers are really selling is their brand, not individual articles. As a consequence, AI companies can mitigate any harm to newspapers by simply attributing their output to its sources, which would drive new readers and revenue to newspapers. This is much like what happened in relation to VCRs, when the “losing” copyright owners eventually realized that the technology they initially opposed had created an extremely lucrative new market for home video rental.

Neither party can admit that compromise is possible, for fear of compromising their litigation position. The essence of our argument is that the court should focus on the public interest in generative AI models by requiring a compromise that would benefit not only the parties, but also the public. In other words, when you see market failure with an obvious solution, that’s the right answer.

  1. With cases such as Bartz v. Anthropic and the litigation against OpenAI moving through the courts, what legal precedents are beginning to emerge?

At the moment, the law is all in a muddle. Only a few of the courts hearing AI cases have issued rulings on the merits, the rulings disagree, and the ultimate outcome is unclear. Two district courts have held that training an AI model is a transformative use of its training data, but both held that it might harm the market for or value of the copyrighted works in its training data, in different ways and for different reasons. The Third Circuit has held that training a non-generative AI model on legal headnotes in order to create a competing legal research tool isn’t a fair use at all, but its reasoning is widely criticized. I’ve already filed amicus briefs in that action, and will be filing more.

  1. Looking ahead, how could these lawsuits reshape the future of artificial intelligence and copyright law?

In my opinion, it’s inevitable that the courts will eventually hold that training an AI model is non-infringing for some reason, probably fair use. But it probably won’t definitely happen until one of these lawsuits gets to the Supreme Court. The lower courts try to hew to doctrine as closely as possible, especially in copyright cases. It often takes the Supreme Court to adapt copyright doctrine to social and technological change.

I can’t imagine that the Supreme Court will hold that training AI models is infringing, because it would cripple the United States AI industry. If AI companies had to license all of their training data, it would impose massive transaction costs that their foreign competitors could ignore. I don’t believe for a second that the Supreme Court will interpret copyright law as saying the United States has to lose the AI technology race to its competitors. The Department of Justice already sent a letter to the district court in the OpenAI Copyright Litigation Infringement case, essentially arguing that it would be in the national interest to find that training an AI model is a fair use of copyrighted works. I’m pretty confident the Supreme Court will eventually agree.