AI Training: Another Option

Eric C. Williams

Managing Director @ Detroit Justice Center | Founder @ Eric C. Williams, PLLC

May 7, 2024

Would Compulsory Licensing Work for AI Training?

The use of copyrighted materials to train artificial intelligence systems has become a heated topic, both in and outside the courtroom. To date, guidance from the copyright office is limited and none of the two dozen or so cases in federal courts have been fully adjudicated. In most of these cases, the outcome will turn on the “fair use” defense advanced by tech companies. If the defense succeeds, the ability of copyright holders to monetize their creative works will be severely compromised. On the other hand, rejection of the fair use defense and a holding that AI training constitutes infringement could bring AI development in the U.S. to a sudden halt, with massive cultural and economic consequences. There is, however, a compromise solution: compulsory licensing.

Compulsory licensing, long an accepted part of the music industry, would take advantage of the lessons emerging from other countries. After all, the United States is not alone in trying to reconcile the interests of copyright holders and AI developers. In 2019, Japan went all in on supporting AI development, revising its copyright laws to permit the use of copyrighted works for virtually any type of information analysis, including for the of training AI models. Although Japan is reconsidering this approach, it is currently the friendliest country to AI and, consequently, the most hostile to the rights of creatives.

At the other end of the spectrum is the European Union, which decided to approach the issue through the regulation of AI, not copyright. The EU’s recently adopted Artificial Intelligence Act explicitly provides that AI is not permitted to infringe on the existing rights of copyright holders. Those rights include the ability to optout of data mining, such as AI training. There are numerous questions regarding the reach and impact of the new law, its applicability to systems developed outside the EU, and its impact on AI development. However, it appears the EU has prioritized the established rights of copyright holders over the desires of tech companies.

Compulsory licensing, which is currently being considered by the U.S. Copyright Office, splits the baby. Most people are familiar with compulsory licensing in the music or video context. Under section 115 of the Copyright Act, reproduction and distribution of nondramatic musical compositions are permitted when the copyright holder has distributed the work to the public in a fixed medium.[1] In return, the copyright holder receives a royalty in an amount set by the Copyright Royalty Board. This is the legal basis for cover songs. It is also used to include copyrighted music in cable television broadcasts and satellite radio.

An AI training regime that incorporates compulsory licensing would compensate creatives while granting AI developers access to a vast amounts of copyrighted materials for training purposes. It would also establish clear guidelines for AI developers and streamline the training process, which could actually accelerate the development of AI. For creatives, the royalties/fees would be a new revenue stream.

There are certainly a number of other factors to consider. Would the certainty of access to copyrighted materials compel developers to abandon their pursuit of free/stolen content? Would creatives need additional protections, for example prohibitions on output that mimicked their style or directly competed in the market? Who would set the royalty rates? However, even with all these questions, a compulsory licensing approach would seem an improvement over the current state of AI-copyright relations.

 [1] The Copyright Act refers to phonorecords, but that includes any material object in which sounds are fixed by any method now known or later developed and from which the sounds can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device.