As AI advances rapidly into every sector of the economy, the tech policy world has no shortage of major questions to tackle. But there is a question currently playing out in the courts that could completely reshape the trajectory of the AI revolution in the coming years: how is copyright law going to be applied to AI training? In particular, will AI training be considered “fair use?”
The potential answers to this question won’t just impact the large companies behind the biggest foundation models, but small businesses and startups as well. After all, small businesses are on all sides of the AI market: they are users of AI products and services; they create and hold their own intellectual property on which AI models could be trained; and, as DeepSeek showed earlier this year, they may be training their own models as well, even without the resources of the biggest players.
Copyright law protects creative works of expression, including literary, dramatic, musical, and even the expressive components of software. It also supports collaborative ecosystems such as open-source software licensing.
The key aspect here is fair use. This judge-made legal doctrine allows certain uses of copyrighted works unauthorized by the copyright holder—such as criticism, commentary, news reporting, teaching, scholarship, and research—when these uses serve the public interest. Courts evaluate fair use by considering: (1) the purpose and character of the use, (2) the nature of the original work, (3) the amount and substantiality of the portion used, and (4) the effect on the potential market.
At the IAPP Governance Conference in Boston in mid-September 2025, I moderated a timely and insightful panel titled, “Addressing the Intersection of AI Training and Copyright.” I was joined by panelists Simone Procas, VP and assistant general counsel for the New York Times, Van Lindberg, partner at Taylor English Duma LLP specializing in IP, and Catie Zaller Rowland, VP and general counsel for the Copyright Clearance Center and former official at the U.S. Copyright Office, who brought perspectives from all sides of the AI market to discuss the state of play regarding whether AI training is or should be considered fair use.
To a packed house of interested privacy and AI governance professionals, the panelists described the latest cases on the topic, Bartz v. Anthropic and Kadrey v. Meta, as well as the New York Times’ recent lawsuit against OpenAI and Microsoft. The judge in Bartz did not hold that using copyrighted material for model training violated the law, but was much more leery of Anthropic’s maintenance of a library of pirated books. The judge in Kadrey dismissed the suit against Meta, but in so doing may have provided a roadmap for future suits to succeed. Rowland also took the audience through the details of the U.S. Copyright Office’s recently released third report on the intersection of copyright and AI, including the feasibility of licensing regimes for training models.
The panelists also debated whether liability should ultimately fall on companies using copyrighted material to train their models, or on users who use AI products to generate work that is found to violate copyright. Lindberg argued that model developers should not be held responsible for users who intend to produce violative material, likening the situation to the landmark 1984 Supreme Court case holding that manufacturers of VCRs were not liable for individuals copying television programs to tapes. Procas disagreed, arguing that training models on copyrighted material is not as innocent as manufacturing a VCR, and that users may not appreciate a system where large companies shift liability to them. All agreed that some sort of licensing regime will end up emerging from these ongoing disagreements.
As these and other lines of argument work their way through the courts, small businesses who are positioned to both take the greatest advantage of new AI capabilities and face the greatest risk from an adverse liability regime are watching closely.