Copyright Office Part 3 is a draft, not law

The US Copyright Office's Part 3 report on generative AI training was issued on 9 May 2025. It rejects the claim that training a model is inherently transformative, which is the load-bearing argument in most fair use defences in this area. It gets quoted in briefs, board decks and LinkedIn posts as though it settled something.

It did not. It is a pre-publication draft. As of August 2026 no final version has been issued. And even a final version would not be law — the Copyright Office administers registration, it does not decide fair use.

Getting the weight right on this document is worth twenty minutes, because it is the single most over-cited item in the AI copyright conversation.

Three documents, three very different weights

The Office's AI work came out in parts, and people flatten them into one thing. They are not one thing.

Document	Date	Status	What it actually governs
Registration guidance	16 Mar 2023	Operational Office practice	What you must disclose about AI material when you file
Part 2 report — copyrightability	29 Jan 2025	Issued report	The Office's stated view on authorship, reflected in how it registers
Part 3 report — training	9 May 2025	Pre-publication draft, no final version as of Aug 2026	Nothing directly. Persuasive at most
Thaler v. Perlmutter	Cert denied 2 Mar 2026	D.C. Circuit ruling left standing; not a Supreme Court merits holding	Human authorship required to register

The bottom two rows are the ones that matter to you, and they matter in opposite directions.

Thaler is the operating rule for US registration. A cert denial is not a Supreme Court decision on the merits, but it left the D.C. Circuit's holding in place: human authorship is required, and purely AI-generated works are unregistrable. That is the rule the Office actually applies.

Part 3 is a pre-publication agency report about a question decided by courts. It is not law, and courts have not adopted it.

What Part 2 actually established, and why it has teeth

Part 2 is the one people should be reading and mostly are not.

Its core holding, restated plainly: prompts alone do not make you an author. Not detailed prompts, not iterated prompts, not a hundred rerolls. The prompt is an instruction, and instructing is not authoring.

What is protectable is the human contribution layered around the generation — selection, arrangement, and modification. AI-assisted works are registrable, with disclosure of the AI-generated material.

That has teeth because it maps onto something operational: the Office's own registration practice. If you file, this is the standard your application meets or fails. Part 3 has no equivalent surface. Nothing is granted or refused on its basis.

How much weight an agency report actually carries

Agency views on statutes they administer can be persuasive to a court. They are not binding. The district decisions on training fair use split, and the two that post-date the report both go the other way.

• Thomson Reuters v. Ross (D. Del., 11 Feb 2025) found no fair use — but on a non-generative research tool that directly substituted for the product it copied from, and the opinion predates Part 3. The Third Circuit heard the appeal on 11 June 2026 and has not ruled.
• Bartz v. Anthropic (N.D. Cal., 23 Jun 2025) held that training on lawfully acquired books was fair use and described it as "spectacularly" transformative. That is close to the opposite of Part 3's framing.
• Kadrey v. Meta (N.D. Cal., 25 Jun 2025) granted summary judgment to Meta on training, expressly on a thin record, while signalling that a properly evidenced market-dilution theory could go the other way.

So the district scoreboard is three cases, not four: Ross sits in the report's general direction on unusual facts and is the one now on appeal; Bartz and Kadrey go the other way. The agency draft binds nobody. Anyone presenting that as a settled position is selling something.

The correct weighting, in one line: cite Part 3 as an argument, never as authority.

What is operative for you today

Strip out the training debate entirely and what remains is short, settled, and much more useful.

Human authorship is required in the US. A raw prompt-to-output generation is effectively unprotectable. That is not a prediction; it is the registration rule the Office applies, and the D.C. Circuit holding that a cert denial left standing.

The commercial consequence is the part worth internalising: if your deliverable is a single-generation image, a competitor can reproduce it and you have no copyright claim. The asset is not the generation. The asset is the human work wrapped around it.

Five habits that put you on the right side of that line:

1. Select rather than accept. Choosing among many candidate generations against a brief is a human act, and it is the cheapest one to evidence.
2. Arrange. Composites, layouts, sequences, and multi-shot edits carry authorship in the arrangement even where individual elements do not.
3. Modify. Retiming, colour, compositing, retouching, and re-cutting are the contributions that survive scrutiny.
4. Record the human steps while you work, not afterwards. An edit decision list is literally a record of human decisions — which takes, in what order, at what length — and because the editor keeps one timeline you can re-render, that record persists rather than being reconstructed from memory later.
5. Disclose the AI material when you register. The registration guidance requires it, and an application that hides it is worse than no application.

The longer version for static work is what makes an AI print design legally protectable, which walks the same logic through a product you can actually sell.

Two adjacent things worth keeping straight while you are in here. First, none of this touches your exposure for outputs that resemble someone else's protected work — that is a separate risk with its own live litigation, and the practical version is in our copyright and safety guide. Second, whether your deliverable is protectable and what rights you grant a client are different questions; get the usage rights written down and see legal and licensing for business content for the contract shape.

FAQ

Is the Part 3 report binding on anyone?

No. It is a pre-publication draft with no final version issued as of August 2026, produced by an agency that administers registration rather than adjudicates fair use. Courts may find it persuasive. Several have reasoned the other way.

Does it mean AI training is not fair use?

It means the Office does not accept that training is inherently transformative. Fair use is decided case by case on a full record by a court, and the district decisions so far have split. The first appellate answer is pending in the Third Circuit.

Does the Copyright Office decide whether I own my AI image?

It decides whether it will register the work, applying the human authorship standard and its disclosure requirements. That is not the same as deciding ownership of the file, which is a contract question between you and your tool provider, or between you and your client. Two separate systems that use overlapping words.

Is it worth registering AI-assisted work at all?

If the work has commercial value and real human contribution behind it, yes — registration is what lets you enforce, and the Office does register AI-assisted works when the AI material is disclosed. If the deliverable is a single unmodified generation, registration is not available and the honest answer is to change the deliverable rather than the paperwork. Before choosing tools for work that needs to be protectable and defensible, it is also worth weighing what each provider publishes about its training data.