Given the sharp rise in usefulness of and attention to generative AI–powered applications, the US Copyright Office issued a notice of inquiry and request for comments in order to conduct a deep dive on how this technology could impact fundamental aspects of copyright law, including some issues we have recently discussed, such as potential infringement based on model training content and copyright protection of generative AI outputs.
We note two key takeaways from this notice. First, the Copyright Office is exploring the facts about and diverse perspectives on generative AI to better analyze and understand the relevant policy considerations, rather than publishing near-final recommendations for public comment. If to understand is to know what to do (or in this case to advise the US Congress on what to do), then this inquiry presents a significant opportunity for many stakeholders, including nonlawyer technical experts, to influence future policy.
Second, the Copyright Office is keenly focused on the training of these models and eager to know, for example, the following:
- Whether, when, what, and how copyrighted works are used to train the models;
- How and from what sources the training materials are collected (including any permissions from or compensation to such sources);
- How the training materials are selected, retained, and tracked;
- Whether copyright owners should have to opt in or opt out of use for training purposes;
- How and by whom the models are controlled as they learn, adapt, and create;
- How the models infer based on training materials and how could they unlearn such inferences if certain training content were removed;
- Whether particular content could be identified as underlying training material for a model or whether AI-generated output should be labeled in some manner;
- Whether use of copyrighted works to train the models could or should be considered fair use; and
- What impact would a licensing requirement have on the development of AI systems.
In conjunction with specific issues such as ownership and infringement, the Copyright Office has overarching questions and concerns about transparency and accountability. For example, should there be requirements to inform copyright owners if their works have been or may be used for AI training?
With respect to infringement, the Copyright Office is asking fundamental questions such as: Can outputs infringe existing works? Who should be liable for substantially similar (i.e., infringing) outputs—the model designer, trainer, provider, or user—and how will that be determined? Are there circumstances where a user of an AI system should be considered an “author” for copyright purposes, e.g., can some number or type of prompts be enough?
On this last point, note that the Copyright Office has issued guidance instructing applicants to specifically exclude non–de minimis AI-generated content from any copyright application.
For those interested in sharing insights with (and perhaps having an influence on) the Copyright Office, written comments are due on October 18, 2023, and written reply comments are due on November 15, 2023. This is a rare chance to help retrain the model for US copyright law.