The U.S. Copyright Office released a report on February 12, 2025 analyzing the impact of different approaches to copyright on the development of artificial intelligence.  Entitled “Identifying the Economic Implications of Artificial Intelligence for Copyright Policy,” the report compiles the views of nine academic authors and the Copyright Office’s Chief Economist.  The report discusses a number of foundational issues and concludes by evaluating several possible access regimes for data used to train generative AI models.

Four basic regimes are addressed: (1) “Unrestricted Access to Copyrighted Materials”; (2) “Full Retrospective and Prospective Restrictions on Access for Training Purposes”; (3) “Grant Retrospective Conditional Amnesty and Enforce Access Restrictions Prospectively”; and (4) “Create a New Statutory Blanket License.”  The discussion begins with the two most extreme approaches and then moves into potential middle-ground solutions.  Below is a concise summary of the report’s analysis of each of these options.

The first option, unrestricted access, would minimize overall transaction costs and prove easiest for AI developers to use.  Indeed, the race to build ever-better AI models virtually forces their developers to act as though this policy is already in place, unless and until the courts tell them otherwise.  (One potential route to unrestricted access would be the development of a judicial consensus that AI training is noninfringing or constitutes fair use.)  However, exempting AI training from copyright law comes at a high cost to human creators who hope to receive licensing fees from that use, even while AI-generated content starts to crowd out the general market for those creators’ works.  Moreover, as the report points out, even those companies which appear to benefit most from such a permissive approach would face a long-term risk from it: if copyright law’s incentive for human authorship becomes too diluted, generative AI may stagnate due to a decline in the production of new training materials. 

The second approach, full prohibition, strictly penalizes developers for any past infringement in AI training and requires them to obtain authorization from the rightsholders for future use.  While recognizing the internal consistency of handling matters this way, the report notes several practical problems with doing so.  It may not be feasible to trace where and when a given work is used in AI training.  Damages for past use, if they apply, “could add up to staggering sums” and “potentially bankrupt some current players” while greatly increasing costs for others.  And the transaction costs of trying to reach agreements with vast numbers of rightsholders are likely to be high.

The third case, backward-looking amnesty coupled with prospective restrictions, gives AI developers the benefit of the doubt for the period when they have been operating under a lack of clear legal precedent, but imposes a strict set of expectations going forward.  This approach eliminates the vast financial consequences of imposing retrospective liability, while giving rightsholders access to a presumably significant revenue stream going forward.  The relative financial burden of this compromise could be adjusted, for example by creating an authors’ compensation fund collected from all of the existing developers.

The final proposal, a statutory blanket license, attempts to optimize the combination of financial fairness and transactional efficiency.  This approach replaces individually negotiated licenses with a collective, standardized fee for the use of copyrighted works in AI training.  There are precedents for such a grand bargain to resolve copyright challenges posed by new technologies, such as SoundExchange for the streaming of recorded music.  The report acknowledges that some efficiency problems would remain, especially with identifying rightsholders and setting appropriate licensing rates.  It considers several variations in the details of the regime, including how to correlate the licensing fee to the value of each individual work and whether it should be an opt-out or an opt-in system.

Going forward, it will be interesting to see how this report interacts with President Trump’s executive order “Removing Barriers to American Leadership in Artificial Intelligence.”  A slow-moving copyright licensing regime seems out of the question.  But it is unclear how forcefully the government would be inclined to intrude on the private market.  For now, the issue is being handled in the myriad of AI/copyright cases pending in the federal courts.  If, however, the upshot of all that litigation amounts to a large roadblock on AI development in the United States, Congress or the Executive Branch could step in and impose some kind of compromise that maximally facilitates AI development while ensuring a portion of the associated revenue flows to the authors of works used for training.