USCO Standard

The US Copyright Office's 3-Part AI Report, Explained

By Nora Whitfield · August 29, 2026

The US Copyright Office's 3-Part AI Report, Explained

If you've seen references to "Copyright Office AI Part 2" or "Part 3" and weren't sure what document people were even talking about, here's the answer: the US Copyright Office is publishing a formal three-part report on copyright and artificial intelligence, and each part tackles a genuinely different question. This isn't one document. It's a series, released piece by piece, and each installment stands on its own. Here's what's actually in each one, and why it matters if you make anything with AI involved.

You can read the source material directly at copyright.gov/ai, which hosts all three parts as they're released.

Part 1: Digital Replicas

Published July 31, 2024, Part 1 is the narrowest of the three and the one that gets the least attention outside legal circles, mostly because it's not the part most creators run into day to day. It addresses digital replicas: AI-generated or AI-manipulated depictions of a real person's voice or likeness, the deepfake and voice-cloning problem. The report examined whether existing law adequately protects people against unauthorized AI replicas of their image or voice and recommended a federal law to fill the gaps that current state-by-state right of publicity statutes leave open.

If you're a musician, actor, or public figure worried about someone using AI to clone your voice or face without permission, Part 1 is the relevant one. If you're asking whether you can copyright something you made using AI, skip ahead. That's Part 2.

Part 2: Copyrightability

This is the one nearly everyone means when they ask about "the AI copyright report." Published January 29, 2025, Part 2 tackles the question at the center of this entire product: when is AI-generated or AI-assisted output eligible for copyright protection at all?

The core finding is worth quoting almost exactly, because it's the standard the Office has been applying in actual registration decisions ever since: AI outputs are copyrightable only where a human determined sufficient expressive elements. The report is explicit that mere prompting is not enough on its own, no matter how detailed or iterative the prompting process is. But it's just as explicit about the other side of that coin. Human-authored elements that are perceptible in the output, or a human's creative arrangement or modification of AI-generated material, can absolutely be protected.

That second half matters as much as the first, because a lot of people read "prompting isn't enough" and stop there, assuming AI-assisted work is a dead end. It isn't. The report draws a real distinction between passively accepting whatever the AI produces and actively shaping it. If you write your own text and only use AI for editing suggestions, that's yours. If you generate images and then meaningfully arrange, select, and combine them into something new, the arrangement and selection can be protected even where the raw images can't be. This is exactly the line the Office drew in its actual registration decisions on cases like Zarya of the Dawn and A Single Piece of American Cheese, both of which we cover in detail elsewhere on the site.

Read the full Part 2 report directly: Copyright and Artificial Intelligence, Part 2: Copyrightability.

Part 3: Generative AI Training

The pre-publication version of Part 3 came out May 9, 2025, and it's the most contentious of the three because it addresses the question with the most money riding on it: the legality of training generative AI models on copyrighted works, licensing considerations for that training data, and how liability should be allocated when a model is trained on material its developer didn't have the rights to use.

Part 3 leans toward a more skeptical read of blanket fair use claims for AI training than many AI companies had hoped for, and it lands squarely in the middle of active litigation covering exactly this question, including the cases against OpenAI, Anthropic, and Stability AI that we track in our complete guide to every AI copyright lawsuit. The Office noted that a final version without substantive changes was expected to follow the pre-publication release.

There's a strange footnote to Part 3's release that's worth knowing if you want the full picture: the pre-publication version went out less than 24 hours before the sitting Register of Copyrights, Shira Perlmutter, was dismissed. Whether the two events were connected has never been officially clarified, but the timing added a layer of political uncertainty to a document that's supposed to function as stable, authoritative guidance.

Why the three-part structure matters

It would have been simpler for the Office to publish one long report covering "AI and copyright" as a single topic. They didn't, and the reason is that these are genuinely different legal questions that don't share an analytical framework. Digital replicas is about the right of publicity and personal likeness, closer to privacy law than authorship law. Copyrightability is about who counts as an author and how much human input is required. Training data is about fair use and the rights of the people whose work went into building the AI system in the first place.

Treating these as one undifferentiated blob, "the AI copyright issue," is how a lot of confused takes on this subject get generated. If you're only going to remember one thing from this report series, remember that Part 2 is the one that governs whether you personally can copyright something you made with AI, and its answer turns entirely on how much creative control you exercised over the final expression, not on how the underlying model was built or trained. That second question, covered in Part 3, is a real and important one, but it's a separate fight between rights holders and AI companies, not something that changes whether your own work is protectable.

For the practical version of what the Copyright Office actually wants to see on an application, including how to handle the AI disclosure fields, see our guide to what the Copyright Office requires for AI-assisted work.