The future of generative AI could well be shaped by its relationship to the data that its models are trained on. The top large language models inhale vast quantities of published material scraped from every nook of the internet, much of it subject to copyright, and the rights-holders aren’t all happy about it. A series of court cases, as well as the U.S. Patent and Trademark Office’s interest in the issue, have led many observers to cast 2024 as a make-or-break year for the industry’s future. What will this mean for AI? Despite the focus on 2024, the reality is that the cases now in court could take years to shake out, and it’s unclear when — or whether — the U.S. government will take any kind of new position. But there’s another way to think about the copyright battle: not as a survival threat to generative AI, as tech companies and venture capitalists have warned, but as a force that could re-map the industry globally. Ethan Mollick, a professor at the University of Pennsylvania’s Wharton business school and AI researcher, predicts that copyright laws — which vary from country to country — could well bring on a geographic reshuffling of AI development that favors countries based on their varying IP regimes. “People tend to view this AI copyright stuff in apocalyptic terms,” he told DFD. “The EU law is likely to be different from U.S. law. You can easily imagine companies headquartered, nominally or in reality, in another country doing training.” So where will AI go? One potential benefactor is Japan, where current law generally permits AI models to train on copyrighted material without obtaining a license or permission. If court rulings in America start to go against AI training, Mollick envisions a U.S. generative AI field consisting chiefly of industry giants with the means to handle legal challenges or obtain clear rights through licensing deals, while startups gravitate toward places like Japan. Ryan Abbott, a partner at law firm Brown Neri Smith & Khan, agrees that scenario is certainly plausible, telling DFD that “there’s definitely going to be some level of geographic resorting. Whether it’s a major one or not remains to be seen, as does the geographic landscape for these rules.” In the U.S., at least, legal disputes around how copyright law principles will apply to AI are spread across various forums. It’s conceivable that district and circuit courts will offer different interpretations, leaving an opening for the Supreme Court to weigh in. Obtaining a conclusive answer under U.S. law might take up to a decade, according to Abbott. “That is an eternity in the technologies we're looking at and considering right now,” he added. “It may be that in the light of uncertainty, some companies decide that they would prefer to go to jurisdictions where these things are clearly permitted.” The possible case outcomes appear endless. Courts may ultimately affirm that training models on protected content is covered under fair use. They might decide that the only legally defensible way to train models is through content licensing deals, which OpenAI is negotiating at the moment with dozens of publishers. (POLITICO parent Axel Springer struck its own deal with OpenAI in December.) The New York Times took the adversarial route, and if it prevails, the developer of ChatGPT will have to destroy any chatbot models and training data that took in its material. And of course, Congress could always step in and establish a collective licensing agreement or some mechanism for rightholders to opt out of their creations getting swept up in training sets. Adding to the uncertainty is the highly uneven global landscape. This adds complications domestically, too: Case law also hasn’t fleshed out the bounds of allowed activities after training an AI model in a different country. It could take more lawsuits to determine if it’s infringement to deploy a Japan-trained model in the U.S. — or reproduce, use, or sell its output here. This might trigger worries about a “race to the bottom,” in which countries trying to woo AI businesses keep their copyright protections loose. But there’s a counterargument that regulatory clarity is a good thing for a cutting-edge industry, even if it’s a bit stricter than the industry would like. Mollick says he has yet to witness any relocations, though a mass departure would likely intensify the call for some kind of government intervention. “We are right now in a regulatory arms race, as jurisdictions are working to become AI-friendly in their industrial policies,” Abbott said. “If you saw an exodus of AI companies to what they perceived as friendlier jurisdictions, that would be a major pressure point for U.S. policymakers.”
| 
