The Copyright Office Issues A Largely Disappointing Report On AI Training, And Once Again A Major Fair Use Analysis Inexplicably Ignores The First Amendment

As this post explains, we have issues with the third installment of the Copyright Office’s report on AI. The rest of this post discusses them, but we note here that, while the report is due criticism, the criticism does not warrant or justify the firing of Register of Copyrights Shira Perlmutter from her duly-appointed position. But we will save comment on that news for other posts and use this one to discuss our chief concerns with the report itself.

At the end of last week the Copyright Office released a prepublication version of the third, and likely final, report of the study it did on the intersection of copyright law and AI. Earlier installments addressed the questions raised by copyright with respect to AI output and digital replicas, whereas this installment addressed whether and how copyright law is implicated by training an AI model on copyrighted works and, in particular, whether such use of works for this purpose was a fair use.

There are some good things to highlight from the report. For example, it acknowledged a concern raised by commenters, including one we raised as the Copia Institute, that if models could only be trained on licensed works it would inherently produce a distorted model tainted by bias and inaccuracy.

But there are also concerning aspects to the report, with one of the most significant being that there was not a single mention of the First Amendment. And we know commenters raised it as an important consideration, because we did in our comments. In particular we discussed how allowing a copyright to bar AI training would interfere with the First Amendment’s protection of the right to read because if people are free to read directly, then they should be able to use tools (like crawlers and bots) to help them do their reading, and if they can’t be free to use tools to do their reading, then are they really free to read after all, which the First Amendment says they are supposed to be. It’s an important question to resolve, but one which the report seems to have entirely ignored.

It is also odd to omit any discussion of the First Amendment in any significant fair use analysis because fair use is an important way copyright law is able to comport with the First Amendment. As we’ve explained before, the Progress Clause of the Constitution says that Congress has the authority to write copyright law, but the First Amendment tempers that authority, just as it tempers all of Congress’s authority to write laws, to ensure that it “makes no law” that abridges freedom of expression. Without fair use, expressive freedom is often abridged, so it is very odd to produce a major document addressing a potential fair use and not directly consider how the Constitution informs the analysis.

Worse, it seems to be part of a growing trend to skip right over that part, which we saw earlier when the Supreme Court issued its own major fair use decision in the case about whether Andy Warhol’s Prince picture was a fair use of the earlier Lynn Goldsmith photograph. Not only did the entire decision fail to mention the First Amendment even once, but its analytical approach, which has been echoed in this report, tends to overemphasize market concerns over transformative concerns.

True, this report did acknowledge that AI “training a generative AI foundation model on a large and diverse dataset will often be transformative,” and being transformative is also a significant way that First Amendment interests are advanced by fair use because it recognizes how a later use adds something that an original use did not, and fair use is about saying yes to that new thing.

But, like the Supreme Court in Warhol, the Copyright Office invited the subordination of the transformative quality of the new use in favor of concerns about market harm for the original works an AI model trains on. Even though, like in Warhol, and also the Second Circuit’s decision in Hachette v. Internet Archive, these concerns are often predicated on dubious evidence about what harm there actually might be and questionable presumptions about what copyright owners should be entitled to say no to when others want to use their works. (It is another point that we made in our comments that if copyright owners can prevent reading, it would significantly expand the list of exclusive rights a copyright grants from what the statutory language currently includes.)

Unlike these decisions, however, the report only collected and collated public comments speaking to these issues; it doesn’t make law itself. The copyright statute is the statute as Congress has written it and courts will interpret it. But the report is influential in how it guides courts and Congress, and so it is important to note that while on its face it appears seemingly exhaustive, plenty of important analysis is missing from it, and thus its ability to effectively influence is commensurately limited.