USCO Standard

What the US Copyright Office Requires for AI-Assisted Work

By Nora Whitfield · August 8, 2026

The US Copyright Office has been unusually direct about AI-assisted work, more direct than most applicants realize until they're staring at a rejected or partially canceled registration. This guide pulls together what the Office has actually said, across its guidance documents, its registration decisions, and the reports it's issued since 2023, into one plain-English walkthrough of what it requires from anyone registering a work that involved AI tools.

If you only remember one thing from this page, make it this: the Office doesn't ban AI-assisted work from registration. It requires you to tell the truth about which parts a human authored and which parts came from a machine, and it only protects the human parts.

Why the Office cares in the first place

Copyright registration isn't just a formality that hands you a certificate. It's a legal claim, backed by the federal government, that you own specific expression and can sue anyone who copies it. The Office has to make sure that claim is accurate, because a bad registration doesn't just hurt the applicant. It clouds the public record about what's actually protected and what's free to use.

AI-generated content complicates that record in a way nothing before it has, because for the first time, an applicant can generate large volumes of finished-looking expression without exercising the kind of authorship copyright has always required. The Office's guidance exists to keep the registration system honest in the face of that.

The core guidance, in plain English

The Office has issued formal guidance (most notably a 2023 policy statement and subsequent reports) that boils down to a few concrete rules:

Human authorship is required for protection. This isn't new law invented for AI. It's the existing bedrock of US copyright, applied to a new kind of tool. A work needs a human author to be copyrightable at all.

Prompts alone don't make you the author of the output. The Office has been specific here: even a highly detailed, carefully engineered prompt is treated more like an instruction to a commissioned artist than like the artist's own creative execution. You don't control the specific pixels, words, or notes the model produces, even if you controlled the general direction. That distinction, general direction versus specific expressive control, is the crux of nearly every Office decision on this topic.

You have a duty to disclose AI involvement. When you file an application for a work that includes AI-generated content, you're expected to disclose that fact and identify what was AI-generated as excluded material. This isn't optional or a courtesy. Applications are signed under penalty of perjury, and omitting a known fact about how the work was created is a real legal problem, not a technicality.

Human-authored elements remain protectable even in an AI-assisted work. Registering a work with AI-generated components doesn't waive your claim to the parts you actually authored. The text you wrote, the arrangement you chose, the edits you made: those stay protected, disclosure just makes clear that the AI-generated parts aren't included in that claim.

How to disclose AI use on an application

This is where a lot of applicants freeze up, mostly because the process feels more intimidating than it is. When you file through the standard online registration system, there's a field for describing "material excluded" from the claim, which is exactly where AI-generated content goes.

The description doesn't need to be exhaustive to the sentence level. It needs to be honest and specific enough that an examiner (or a court, later) understands what you're claiming and what you're not. Something like "AI-generated illustrations" or "text generated by [tool name], subsequently edited" is the right level of specificity for most cases. What sinks applications is either omitting the disclosure entirely or describing it so vaguely that it doesn't actually clarify the boundary of the claim.

If you're not sure how much of your work counts as AI-generated versus human-authored, that uncertainty is itself a signal you need a better record of your own process, not a reason to guess or skip the field.

What the "material excluded" field actually wants

The material excluded field exists for more than just AI disclosure (it's also used for things like previously published material or public domain elements incorporated into a new work), but AI content is now one of its most common uses. The Office wants a description that lets a future reader understand the scope of your claim without having to reconstruct your entire creative process from scratch.

A useful mental model: write the field as if you're explaining to a stranger exactly which parts of the finished work are yours. If you generated a background image with Midjourney and didn't materially alter it, that's excluded material. If you generated a rough paragraph and then rewrote it substantially in your own words, the final paragraph is your authored text, not excluded material, because the expression in the final version is yours even though a machine produced an earlier draft.

The three-month window and why it matters more than people think

Separate from the AI-specific rules, there's a timing mechanism in copyright law that applies to every work, AI-assisted or not, and it has real teeth. If you register your work before infringement begins, or within three months of first publication, you preserve your right to statutory damages and attorney's fees if you ever have to sue an infringer. Statutory damages range from $750 to $30,000 per infringed work (more for willful infringement), and they matter because actual damages in a copyright case are often hard to prove and can be small even when the infringement is real.

Miss that window and you can still register later, but you lose access to statutory damages for infringement that happened before you filed. You're left proving actual financial harm, which is a much harder case to win. For AI-assisted work specifically, where the underlying value of the work is already harder to establish because part of it may not be protectable at all, that timing matters even more. Get your human-authored claim registered promptly, while it's clean and well-documented, rather than treating registration as something to handle eventually.

The filing basics

Registration through the Office costs roughly $65 for a basic single-work claim as of this writing (fees are set by the Office and do change, so confirm the current number at copyright.gov before filing). The process is designed to be usable without a lawyer for a straightforward single-author claim: you create an account, describe the work, disclose any excluded material, upload a copy, and pay the fee. Processing time varies, often several months, which is a separate reason not to wait until you're worried about infringement to start the process.

What examiners are actually looking for

Office guidance and public statements from examiners point to a few consistent things they check on AI-assisted applications: whether the disclosure is present and specific, whether the claimed authorship is plausible given what's disclosed (a claim to "the entire novel" alongside a disclosure that says "text generated by AI, lightly edited" is going to draw scrutiny), and whether the application is internally consistent. Inconsistency, more than the mere presence of AI, is what triggers a closer look or a rejection.

This is exactly why a real-time record of your process is worth more than a good memory. If an examiner has questions, or if your registration is challenged years later in litigation, the applicant who can produce a contemporaneous account of exactly what they wrote, edited, and generated is in a completely different position than one trying to reconstruct events from memory.

What happens if you skip disclosure

Registrations obtained without proper AI disclosure aren't automatically void, but they're vulnerable. If the omission surfaces later (in litigation discovery, in a licensing negotiation's due diligence, in a routine audit), the registration can be canceled or found unenforceable for the undisclosed portions. That's a far worse outcome than disclosing honestly from the start and registering a smaller, accurate claim. A smaller registration that holds up beats a larger one that collapses under examination.

Registering an AI-assisted book or long-form work, step by step

For longer works like books, the practical version of all of this looks like: keep track of which chapters or sections involved heavy AI drafting versus your own writing as you go, note significant AI-generated elements (a generated cover image, generated chapter illustrations) separately from your prose, and when you file, describe the excluded AI content specifically enough that an examiner understands the scope without needing your entire manuscript history. Waiting until the manuscript is finished to reconstruct which parts were AI-assisted is exactly the failure mode this whole guide is trying to help you avoid.

Common mistakes that trip up AI-assisted applications

Claiming the whole work as human-authored when part of it wasn't. This is the single most common problem, and it usually isn't malicious. People generate an image or a passage, keep it because it's good, and just don't think of it as a separate thing from the rest of the work they made. The Office doesn't care about your intent here. It cares whether the disclosure matches reality.

Describing excluded material too vaguely to mean anything. "Some AI assistance was used" tells an examiner almost nothing. It doesn't establish what you're claiming versus what you're not, which defeats the purpose of the field. Be specific about which elements, even if you can't be specific about exactly how much editing happened to each one.

Waiting to figure out authorship until the application is due. By the time you're filling out the form, you're relying on memory to reconstruct a creative process that may have unfolded over months. This is the exact problem a contemporaneous record solves. If you know while you're working which passages you wrote from scratch, which you heavily edited from AI drafts, and which you kept largely as generated, filling out the disclosure becomes a lookup, not a reconstruction project.

Assuming a registration, once granted, is permanent regardless of what's later discovered. It isn't. Registrations can be challenged and canceled if it turns out the application misrepresented authorship, which is exactly why getting the disclosure right the first time matters more than getting the registration fast.

How this differs from patent law, briefly

Writers and artists sometimes assume copyright and patent AI rules are the same question with different names. They aren't, and the difference matters if you're building something that has both creative and technical dimensions, like software with a novel interface or a game with original mechanics. Copyright protects the specific expression, the actual text, code, images, or sounds you authored. Patent protects a novel functional invention, an entirely different legal test with its own AI-inventorship questions currently working through the courts. If your work has both a creative expression side and a genuine technical invention side, you may need to think about both separately, and the human-authorship record that helps your copyright claim isn't automatically the same evidence a patent claim would need.

Quick answers to what people actually ask

Can I amend a registration if I find out I disclosed wrong? Yes, the Office has a supplementary registration process for correcting or amplifying information in an existing registration, including authorship disclosures. It's not instant and it's not free, but it exists precisely because people do sometimes get the disclosure wrong the first time, especially on long or complex works. Getting it right initially is still cheaper and cleaner than fixing it later.

Does the Office treat different AI tools differently? No. The guidance is tool-agnostic. It doesn't matter whether you used Midjourney, ChatGPT, Claude, Suno, GitHub Copilot, or something released next month. The test is always the same: did a human control the specific expression, regardless of which tool assisted.

What if my collaborator used AI and I didn't know the extent of it? This is a real risk in co-authored or commissioned work, and it's one more reason to have your own clear record of your own contributions, independent of what a collaborator tells you about theirs. If you're jointly registering a work, you're both making representations to the Office, and you want your own side of that claim to be solid regardless of what happens with a collaborator's.

The practical takeaway

None of the Office's requirements are designed to punish you for using AI tools. They're designed to keep the registration system's claims accurate, and accuracy happens to require the same thing good practice always required: knowing what you actually made, and being able to show it. Build that record while you work, not after.