The US Copyright Office's AI Report, Part 2: Copyrightability

The United States Copyright Office published Part 2 of its report on copyright and artificial intelligence, subtitled “Copyrightability”, on 29 January 2025. It is the Office’s considered position on what can be registered when a generative system was involved, and its conclusions are narrower and more usable than the headlines suggested.

What was published, and when

The report is the second in a series. Part 1, on digital replicas, was published on 31 July 2024. Part 2, on copyrightability, followed on 29 January 2025. Part 3, on the use of copyrighted works in training generative models, was released as a pre-publication version on 9 May 2025 and is a different subject with a different status — covered in the Part 3 page. All are published at the Copyright Office’s AI initiative page.

Part 2 followed the Office’s 2023 notice of inquiry and the very large body of public comments it drew, and its top-line conclusion is that existing copyright law is adequate to resolve questions of copyrightability in AI-assisted works and that no legislative change is needed for that purpose. That conclusion matters as much as the substantive ones: it is the Office declining to treat generative AI as a category requiring a new statutory regime.

The human authorship conclusion

The Office reaffirmed that copyright protection in the United States requires human authorship, and that material generated wholly by a machine without human creative contribution is not protectable. This is not a new position and it does not originate with generative AI; the Office’s practice has required human authorship for a long time, and the courts have been consistent with it.

The judicial anchor is the litigation over an image the applicant said was autonomously generated by a computer system, in which the applicant named the machine as the author. The District Court for the District of Columbia upheld the Office’s refusal to register in August 2023, and the Court of Appeals for the District of Columbia Circuit affirmed on 18 March 2025, holding that the Copyright Act requires a human author. Note what that case did not decide: the applicant had conceded there was no human author, so it says nothing about works with a mix of human and machine contribution — which is every practical case.

Why prompts were held not to be enough

The most contested conclusion in Part 2 is that prompts alone do not, given the state of the technology the report describes, give a user sufficient control over the expressive elements of the output to make that user the author of it. The Office was explicit that this holds even for long, detailed and iteratively refined prompts.

The reasoning is about the gap between instruction and expression. A prompt describes what is wanted; the system determines how it is expressed, and the same prompt can produce materially different outputs. The Office analogises to the long-standing principle that a person who gives directions or ideas to a creator is not thereby the author of what the creator produces. The unprotectability of ideas is doing the work here.

Two qualifications are worth carrying. The report ties the conclusion to current technology and expressly leaves room for that to change if systems give users determinative control over expressive output. And it is a conclusion about authorship of the output, not about whether the prompt itself is protectable — a sufficiently creative prompt can be a literary work in its own right, which protects the text of the prompt and not the picture.

What can still be protected

The report is more permissive than its reputation, and identifies routes by which a work involving AI is protectable.

• Human-authored expressive inputs. Where a human’s own copyrightable expression — a drawing, a text, a photograph — is put into the system and remains perceptible in the output, that expression is protected in the output as it was outside it.
• Creative selection, coordination and arrangement. A human who selects among generated outputs and arranges them can hold copyright in the selection and arrangement, on the same compilation reasoning that protects an anthology. The protection covers the arrangement, not the underlying generated material.
• Creative modification. A human who materially modifies generated output has protection in the modifications.
• AI as an assistive tool. Using AI for tasks that do not determine expressive content — upscaling, de-noising, spell checking, some editing — does not affect the protectability of the human-authored work.

In every case the protection extends only as far as the human contribution, and the generated material remains outside it. A comic book whose text and layout are human and whose images are generated is protected as to the text and layout, and anyone may reuse the images so far as copyright is concerned. That is precisely the shape of the Office’s treatment of the graphic novel registration that first made the issue concrete.

What this means at registration

Applicants have to disclose AI-generated content that is more than de minimis and must exclude it from the claim, following the registration guidance the Office issued in March 2023 in the Federal Register. In practice that means describing the human contribution in the application and disclaiming the machine-generated material, exactly as one would disclaim pre-existing material in a derivative work.

Three practical consequences follow. Keep records of the human contribution at the time you make it, because reconstructing which elements were human afterwards is close to impossible and the burden is on the applicant. Expect the registration you obtain to be narrower than the work, which affects what you can enforce. And do not read a registration as a finding about the AI question — a certificate reflects the claim as filed, and a court may take a different view of where the line falls.

Finally, none of this is about the input side. Whether training on copyrighted works is lawful is an entirely separate question the Office took up in Part 3 and which the courts are deciding independently; see the four fair-use factors as applied to training. A work being unprotectable and a work being infringing are unrelated conclusions, and conflating them is the most common error in conversations about this report.