Authors asking whether they can “opt out” of AI training usually point to one document: Part 3 of the U.S. Copyright Office’s report on copyright and artificial intelligence, titled Generative AI Training. The report is worth reading closely, partly because it is widely described as finished when it is not, and partly because what it says about opt-outs is narrower and stranger than the headlines suggested. It does not give authors an opt-out right. It says something closer to the opposite.

What Status the Report Actually Has

The Office released Part 3 on May 9, 2025, labeled a “Pre-Publication Version.” According to the Office’s own AI report page, it was released in response to congressional inquiries and stakeholder interest, and a final version will be published in the future, with no substantive changes expected to the analysis or conclusions. As of the date of this article, that page still lists the pre-publication version and gives no date for a final one. A final version has not been issued, so any summary that calls the report “finalized” is ahead of the record.

The delay has an institutional backdrop. The pre-publication release came amid a dispute over who leads the Office, and a July 2026 law-firm alert from Lowenstein Sandler reported that the Supreme Court had denied a stay application on June 30, 2026, leaving the Register of Copyrights in place. Whether that clears the way for a final version is speculation; nothing has been announced.

Two caveats matter for authors. First, the report is an agency analysis, not law. It binds no court. Second, because it is still labeled pre-publication, its conclusions could in principle be revised, even though the Office has said it does not expect substantive changes.

What the Report Concludes About Training and Fair Use

The report’s core analysis runs in a few steps.

Training implicates exclusive rights. The Office concludes that several stages of building a generative AI system, including assembling datasets and training, involve uses that implicate copyright owners’ exclusive rights. The live question is whether those uses are excused as fair use.

Fair use is case by case, and some training is likely transformative. The report says various uses of copyrighted works in training are likely to be transformative, but fairness depends on what works were used, from what source, for what purpose, and with what controls on outputs. It distinguishes research or analysis uses, where outputs are unlikely to substitute for the training works, from commercial use of vast troves of works to produce expressive content that competes with them. The Office says the latter, “especially where this is accomplished through illegal access,” goes beyond established fair use boundaries.

Pirated sources weigh against fair use. The Office’s view is that knowingly training on a dataset of pirated or illegally accessed works should weigh against fair use, though not decide the question by itself.

Market dilution counts. On the fourth factor, the report says that even where outputs are not substantially similar to a particular book, they can dilute the market for works similar to those in the training data.

Courts have not uniformly followed this reasoning. In June 2025, one federal judge held that training on lawfully acquired books was fair use while declining to excuse piracy, and another granted summary judgment to a developer because the authors had not proven market harm, while writing that in many circumstances unlicensed training would be unlawful. The Office’s report is persuasive material for future litigants, not a verdict. Our overview of where the AI lawsuits stand tracks those cases.

This is the part most often misreported. The Office does not recommend that authors be given a legal opt-out. Its reasoning runs the other way.

The report says that requiring copyright owners to opt out is inconsistent with the basic principle that consent is required for uses within the scope of their statutory rights. Copyright, in other words, is an opt-in system: using a work in a way the statute reserves to the owner requires permission unless an exception such as fair use applies.

It then adds a conditional. If Congress were ever to create an exception for AI training, the report says the ability to opt out could preserve some ability to block unwanted uses or negotiate terms. Even then, it flags significant concerns about whether opt-outs would be effective and available, and it lists design problems: the mechanism should be simple enough for individuals without technical expertise, company-by-company opt-outs would be burdensome, and honoring a platform setting like robots.txt is not enough because many owners do not control the sites where their works appear, including pirate sites.

The report also surveys the European Union’s approach, where a text and data mining exception is subject to rightsholder opt-outs, and records commenters’ criticism that such regimes are hard to operate.

For an author in the United States, the practical reading is this: no statute gives you a right to opt out of training, and the Office’s report does not propose one. What exists is a contested legal position that unauthorized training may infringe, plus whatever contractual control you negotiated.

What It Says About Licensing

On remedies, the Office recommends that the licensing market continue to develop without government intervention. It describes voluntary licensing as already happening in some sectors and says that where licensing is available or likely to be feasible, that consideration disfavors fair use under the fourth factor. It says government intervention would be premature. If specific market failures emerge for particular works, it suggests targeted approaches such as extended collective licensing, a system in which a collective management organization licenses a whole class of works and individual owners can opt out to negotiate separately.

That recommendation has a consequence for authors: the Office is pointing toward deals. Who signs them for a given book depends on the grant language in your publishing contract, which is why the question of what your publisher can license to AI companies matters more than anything the report says about opt-outs. The report itself cites press coverage of an academic publisher’s AI deals in which authors were told there was no opt-out.

What to Do With This

Treat the report as a map of how a federal agency reads the law, not a source of rights. Read the pre-publication text itself, particularly the fair use and licensing sections, rather than a summary. Check your own contracts for the grant clause and any AI-related language, because that is where consent is decided in practice. And if you are negotiating a new agreement, ask for an express clause requiring your written permission before the work is used for AI training. A final report, if and when it appears, will not change that advice.

This article is editorial and informational, not legal advice. Consult a licensed attorney about your specific situation.