The last few days have been quite dramatic for anyone paying attention to copyright law and policy. 

• Thursday: The Librarian of Congress, who ultimately oversees the US Copyright Office, was dismissed from her post by President Trump late on Thursday. 
• Friday: The Copyright Office made a highly unusual “pre-publication” release of its third and last copyright and AI report, this one on the most controversial subject of them all – whether fair use applies to use of training data for generative AI.
• Saturday:  Reports started rolling in that President Trump had also sent a notice attempting to dismiss Register of Copyrights Shira Perlmutter. A statement from Rep. Joseph Morelle released that same evening objected to the move and connected Perlmutter’s dismissal with the Office’s AI report, which he construed as being at odds with the administration. 

The Generative AI Training Report

First, let’s start with the report on copyright and generative AI training. This report is the last of three highly anticipated reports from the US Copyright Office, based on a policy study it began in August 2023, receiving over 10,000 comments from the public in response. These kinds of studies and resulting reports are not unusual – the Office has conducted dozens over the years, offering valuable insights and advice to policymakers about how copyright law should react to changes in technology. While we certainly don’t always agree with what the Office says, these reports are typically very helpful. 

The first two reports came in July 2024 on AI Digital Replicas and in January 2025 on copyright protection of AI outputs. These reports were highly anticipated, with members of Congress eagerly asking about their release several times in writing and in oversight hearings. But by far the most anticipated and most controversial was always bound to be the Office’s take on the key copyright question underlying most of the now 40+ pending copyright AI lawsuits: Is it fair use to use copyrighted works for training an AI model, and particularly, is it fair use to use that model in a generative AI system? 

The report tackles this question (and a handful of other foundational questions) in about 113 pages. Much of the reporting so far has been that the report generally favors rightsholders, and takes a stance that is negative for commercial generative AI applications. While the report certainly does have some negative things to say about the fair use case for generative AI training, it’s actually more of a mixed bag, but we will let you decide for yourself – these are some highlights of the Office’s conclusions on key legal issues: 

Transformative use: Whether such uses are “transformative” is a key question under the first fair use factor, asking about the purpose and character of the use. On this the Office concludes: 

In the Office’s view, training a generative AI foundation model on a large and diverse dataset will often be transformative. . . . 

But transformativeness is a matter of degree, and how transformative or justified a use is will depend on the functionality of the model and how it is deployed. On one end of the spectrum, training a model is most transformative when the purpose is to deploy it for research, or in a closed system that constrains it to a non-substitutive task. . . . 

On the other end of the spectrum is training a model to generate outputs that are substantially similar to copyrighted works in the dataset. . . . For example, a foundation image model might be further trained on images from a popular animated series and deployed to generate images of characters from that series. Unlike cases where copying computer programs to access their functional elements was necessary to create new, interoperable works, using images or sound recordings to train a model that generates similar expressive outputs does not merely remove a technical barrier to productive competition. In such cases, unless the original work itself is being targeted for comment or parody, it is hard to see the use as transformative.

Many uses fall somewhere in between. 

“Bad Faith” and Sourcing of Training Data: One of the most controversial issues has been sourcing of data: numerous news reports have reported that AI companies have sourced data from websites like LibGen that contain illegitimate copies of works. Whether this sourcing ultimately matters is a live issue in the current AI lawsuits, and thinking on whether the courts should consider the souring as a matter of “good” or “bad” faith is split. This is the Office’s take: 

In the Office’s view, the knowing use of a dataset that consists of pirated or illegally accessed works should weigh against fair use without being determinative. Courts have expressed some uncertainty about whether good or bad faith generally is relevant to the fair use analysis. The cases in which they have done so, however, involved defendants who used copyrighted works despite the owners’ denial of permission. Training on pirated or illegally accessed material goes a step further. Copyright owners have a right to control access to their works, even if someone seeks to obtain them in order to make a fair use. Gaining unlawful access therefore bears on the character of the use.

Market Harm: Perhaps most controversial has been the question of how AI training may affect the market for original works. The fair use analysis often turns on questions about harm to the market for the original, and traditionally that has focused on lost sales or licensing revenue due to direct substitution. In the pending AI cases, litigants have articulated a much broader view, arguing that AI systems that produce outputs that dilute the market for the original by producing works of the same kind or style should count against a fair use finding: 

While we acknowledge this is uncharted territory, in the Office’s view, the fourth factor should not be read so narrowly. The statute on its face encompasses any “effect” upon the potential market. The speed and scale at which AI systems generate content pose a serious risk of diluting markets for works of the same kind as in their training data. That means more competition for sales of an author’s works and more difficulty for audiences in finding them. If thousands of AI-generated romance novels are put on the market, fewer of the human-authored romance novels that the AI was trained on are likely to be sold.  . . . Market harm can also stem from AI models’ generation of material stylistically similar to works in their training data.” 

And so the Office concludes: 

The copying involved in AI training threatens significant potential harm to the market for or value of copyrighted works. Where a model can produce substantially similar outputs that directly substitute for works in the training data, it can lead to lost sales. Even where a model’s outputs are not substantially similar to any specific copyrighted work, they can dilute the market for works similar to those found in its training data, including by generating material stylistically similar to those works.

On balance, we’re supportive of some of the Office’s conclusions and disappointed with others. The report indicates in several places that the Office believes that research and academic uses should be favored under the fair use analysis, which is something we strongly agree with. On the other hand, the Office’s conclusions about market harm are difficult to square with past case law on market harm and could cause unintended consequences if applied outside the AI context. As we noted in our comments to the Copyright Office, “We can think of no fair use case that has ever assessed market harm by adopting such a broad approach to market harm, but numerous instances in which courts have rejected such an approach.” 

How the report was released and why that matters

The release of this report was highly unusual. We’re not aware of the Office releasing any of its other policy studies in the past under a pre-publication status, and we’re not really sure what that means. The report itself states that

“The Office is releasing this pre-publication version of Part 3 in response to congressional inquiries and expressions of interest from stakeholders. A final version will be published int he near future, without any substantive changes expected in the analysis or conclusions.”

With the dismissal of Perlmutter, whether further changes will be made or whether the report will ever be finalized remains to be seen. 

The timing of the release has generated speculation that the release of the report was politically timed – suggesting that a quick release to get its contents out into the world before it could be restricted by a Trump administration that has (it has been suggested) a very different view that is more aligned with technology companies. Whether that is true we may never know. 

Whatever the motivation for the timing of the release, we think it’s unfortunate for this reason: there are two cases currently pending in the courts, Kadrey v. Meta and Bartz v Anthropic, where the issues being litigated are directly related to the contents of this report. In both of those cases, the briefing on fair use is closed. In one, the court has already held oral argument and in the other oral argument will be held next week. Amicus briefs on behalf of other interested parties have been filed.  

While it’s certainly fair for the Copyright Office to make its views known, even on controversial issues, the timing of this report is problematic because it could put a thumb on the scale for how the courts will resolve these cases. Given that these cases are likely to be among the first to be decided, the downstream impact on all other AI litigation could be considerable. Neither the parties nor the amici who have participated in the case will have a chance to respond or point out flaws or gaps in the Office’s analysis. 

So what will the courts do with this report? It’s possible that the parties could bring the report to the Court’s attention as supplemental authority, and it’s possible that the courts will digest it themselves. The Office’s reports are only entitled to what’s known as “Skidmore” deference (meaning, really, no deference at all except to the extent the court is persuaded by the Office’s arguments). That’s some solace, but it’s still the case that the courts could well read this report as a sort of amicus brief on behalf of the Copyright Office, though contributed out of time, well beyond the normal word limits, and not published or vetted in the normal way. Very unusual. 

President Trump’s attempted dismissal of Shira Perlmutter

Following all this, on Saturday, President Trump attempted to dismiss Register of Copyrights Shira Perlmutter. We haven’t heard from Perlmutter herself yet, so we don’t know if she has accepted this attempted dismissal or will fight it (we note that although removal of Hayden may have been legally permissible, that the President can remove Perlmutter is far less certain and she has substantial legal room to contest the move if she chooses to do so). 

Perlmutter has been Register since October 2020 and she has been a good one. She was a public servant in the Copyright group at USPTO for years before becoming Register, and came to the Copyright Office with great expertise. She has been willing to listen to different points of view, and has expanded her Office’s own expertise (for example, by hiring the Office’s first chief economist). She has supported continued development of copyright resources such as the Office’s “fair use index,” approved numerous DMCA exemptions that benefit the public (including our own text data mining exemption), and pushed back on expanding copyright in areas where it would be of questionable value such as with a new press publishers’ right. 

Under Perlmutter’s leadership, Copyright Office staff also served as expert advisors to the American Law Institute’s Restatement of the Law of Copyright. Molly Van Houweling, who serves as one of the Associate Reporters for the Restatement (and also an Authors Alliance board member), recounts that: “The Copyright Office participation in the Restatement work during Register Perlmutter’s tenure has been active and extremely constructive. The Office has been particularly helpful in explaining intricate points of registration practice, while also offering bigger picture insights about copyright law and policy.”

So we think the efforts to push Perlmutter out are a shame. What exactly motivated them is still an open question. As I mentioned already, the immediate reaction to her removal was that it must have been related to this AI report and the policy positions the Office has taken on AI under her leadership, particularly as some important AI cases are currently hearing, or have just heard, oral arguments directly addressing points raised by the US Copyright Office’s report. 

That doesn’t seem like a crazy deduction, but there may be some other factors at play. We note, for example, that there were calls over a week before from a conservative political foundation, calling for the removal of “deep state” librarians, naming Carla Hayden and Shira Perlmutter (who is not actually a librarian, to be clear). That call was based in part on their actions related to DEI and gender, but also referenced their positions on intellectual property.  

Here is a timeline of some relevant recent events to help think through what has happened: 

• April 25th:  Bartz v Anthropic, in its Opposition to Anthropic’s Motion for Summary Judgment, the plaintiffs focused heavily on Anthropic’s use of pirated works as contrary to fair use (“So they cheated. Instead of relying on publicly-available data or paying for a license – which would have taken longer and cost more money – they went to known pirated websites and downloaded books for free.”)

• April 30: Kadrey v. Meta,  Judge Chhabria posed 12 questions to the parties. These included a question related to market substitution and its impact on fourth factor fair use analysis and a question related to Meta’s use of “pirated” works.

• April 30: American Accountability Foundation urges President Trump to fire “deep state” librarians, targeting Carla Hayden and Shira Perlmutter. AAF identifies Hayden’s support of Biden policies, particularly regarding transgender policies, and specifically targets their approach to intellectual property. (“The President and his team have done an admirable and long-needed job cleaning out deep state liberals from the federal government. It is time they show Carla Hayden and Shira Perlmutter the door and return an America First agenda to the nation’s intellectual property regulation,” Tom Jones, the president of the American Accountability Foundation, told the Daily Mail.)

• May 1:  Kadrey v. Meta, Judge Chhabria  heard oral arguments on cross motions for summary judgment, and appears to have been particularly interested in the substitutive effect of AI generated works on those works used to train the AI (“Companies can create a product capable of producing infinite number of competing works … and dramatically change and obliterate the market for the [original] works.”)

• May 8: Rubio remarks. AAF reposts Secretary of State Marco Rubio’s remarks about his various roles in government and joking about taking on the Library of Congress.

• May 9: Hayden fired. President Trump fires Carla Hayden. White House Press Secretary Karoline Leavitt defends the firing  (stating that Hayden “did not fit the needs of the American people” and “there were quite concerning things that she had done at the Library of Congress in the pursuit of DEI and putting inappropriate books in the library for children.”)

• May 10: Copyright Office Generative AI Training Data Report released, under unusual circumstances as a “prepublication,” offering some negative conclusions about the fair use case for using copyrighted materials for training AI models for deployment in AI systems.

• May 11: Register of Copyrights Shira Perlmutter receives notice of dismissal from President Trump, according to news reports.

• May 12: President Trump announces appointments of an Acting Librarian of Congress, Deputy Librarian, and Acting Register of Copyrights, as reported by CBS News.