A report released this week by the European Parliament endorsed statutory compulsory licensing of copyrighted works for AI training as the most workable “welfare-superior” model for balancing the interests of innovation and fair compensation for creators and rights holders.
“Making the licence compulsory ensures wide access to works, a royalty rate set by a regulatory authority balances interests of rightsholders, AI developers and end users,” the 53-page report said, “enabling royalty payments that create incentives for enough new creation to keep AI systems valuable over time.”
The study, The Economics of Copyright & AI, was commissioned by the EU Parliament’s Policy Department for Justice, Civil Liberties and Institutional Affairs at the request of the Committee on Legal Affairs, and prepared by Christian Peukert of the University of Lausanne’s Business and Economics Faculty.
“Relying on [voluntary] licensing markets alone risks suffering from high clearance costs and too little available training data due to opt-outs,” the study concluded. “Copyright exceptions without remuneration risk undercutting incentives for future creation. Copyright exceptions without remuneration and opt-outs appear as the worst policy option, because opt outs reduce access to the existing stock of data and there is no funding for a continued flow of data.”
The study drew on the history of legal and regulatory responses to the impact of earlier disruptive technology innovations, such as the introduction of Napster in 1999, on the market for copyrighted works, for lessons on how not respond to AI.
“Peer-to-peer file-sharing services such as Napster brought mass-scale demand-side copyright infringement, but enforcement through litigation or stricter laws proved largely ineffective,” the study found. “Evidence shows that announcements of anti-piracy laws in France and Sweden led only to short-lived increases in sales, and the effects typically vanished within six months. While legal efforts succeeded in shutting down or blocking particular unlicensed services, alternatives were quick to emerge. Similarly, lawsuits against individual users failed to curb piracy, and self-regulation by the involved actors failed to reduce the supply of unlicensed copies.”
Read more: Copyright, Antitrust, and the Politics of Generative AI
The report is equally pessimistic about the benefits of implementing a broad exemption to copyright law for AI training.
“Abolishing copyright for training data would maximize short-run access, but it risks depressing incentives for future creation, as early evidence from stock-photo and content platforms already suggests,” the report said. “The challenge is to find an institutional design that restores creation incentives up to the maintenance threshold, while avoiding excess transfers and administrative overhead.”
The optimal, politically-feasible solution, the report proposed, is “a statutory license with a modest royalty.” According to report author’s estimate, statutory licensing “generates about $14 billion more annual welfare than the next best alternative. Opt-out regimes are particularly problematic, as they shrink access without restoring incentives for fresh data. Licensing markets (“opt-in”) can work in principle but suffer from high clearance costs and fragmented coverage of the available stock of data.”
The report’s conclusion is likely to be controversial, particularly among creators and rights holders. Many artists and creators object to their work being used at all in AI training, whether compensated or not. A compulsory license would deny them that choice.
Rights holders in copyright markets already operating under statutory licenses, such as music, complain that compulsory licensing prevents them from realizing the full market value of their holdings.
For AI developers, a compulsory license would likely increase their costs, but also provide a measure of certainty. Too high a fee for a required license could also become a barrier to entry for would-be new entrants and entrench the power of the incumbents.
The report concludes, however, that “the majority of welfare gains from AI accrue to consumers, and policy should be designed to preserve these gains while ensuring that creators continue to supply the fresh, high-quality data on which future progress depends.”
