Output Copyrightability: the Zarya of the Dawn and Thaler Decisions

Two decisions are cited for the proposition that AI output cannot be copyrighted in the United States. Only one of them is a court decision, they concern different facts, and neither says that a work made with AI assistance is unprotectable. The distinction is the whole practical content of the area.

Zarya of the Dawn, February 2023

Zarya of the Dawn is a graphic novel by Kris Kashtanova whose images were generated with Midjourney. It was registered in September 2022; the Copyright Office subsequently learned of the AI involvement and, in a letter dated 21 February 2023, cancelled the original registration and issued a narrower one covering the work’s text and the selection, coordination and arrangement of its written and visual elements — but not the individual images. The letter is published by the Office at copyright.gov.

Two things about its status are routinely misreported. It is an administrative decision of the Register of Copyrights on one application, not a judicial precedent, and it binds nobody as law — though it is a clear statement of the Office’s registration practice, which is what most people actually need to predict. And it is a decision that granted protection as well as refusing it: the human authorship in the arrangement and the text was registrable. The headline “AI comic loses copyright” describes half of the outcome.

Thaler v Perlmutter, 2023 and 2025

Stephen Thaler applied to register an image titled “A Recent Entrance to Paradise”, naming his system, the Creativity Machine, as the author and himself as owner by virtue of ownership of the machine. He asserted throughout that the work was created autonomously with no human creative contribution — a stipulation he made deliberately in order to test the question, and one that shaped everything that followed.

The Copyright Office refused registration. On 18 August 2023 Judge Beryl Howell of the United States District Court for the District of Columbia granted summary judgment for the Office, holding that human authorship is a bedrock requirement of copyright. On 18 March 2025 the United States Court of Appeals for the District of Columbia Circuit affirmed, holding that the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being. That appellate affirmance is the part that matters: it is binding authority in the D.C. Circuit and is the highest court to have ruled on the question.

What Thaler does not decide is at least as important. Because Thaler stipulated to full machine autonomy, the case presented no question about works with mixed human and machine contribution, and the D.C. Circuit said so. It also did not decide whether the human authorship requirement is constitutional in origin or purely statutory, which matters because a purely statutory rule can be changed by Congress.

The rule the two decisions establish

Read together, and alongside the Copyright Office’s guidance and the copyrightability part of its AI report published in January 2025, the operative US rule is narrow and reasonably clear:

• Material generated by a machine without human creative control is not copyrightable, and cannot be registered.
• Human-authored elements of a work that also contains generated material remain protectable — text a person wrote, images a person made, and the selection, coordination and arrangement of the whole.
• Human modification of generated output can supply authorship in the modified expression, to the extent of the modification.
• An applicant must disclose AI-generated content that is more than de minimis and disclaim it in the application. This is a duty of candour to the Office, and a registration obtained without it is vulnerable.

The Office’s report parts are collected at the Copyright Office AI initiative page. Note that the report is an agency’s reasoned position; a court may disagree with any part of it that has not been through litigation.

Why prompts have not counted as authorship

The recurring argument is that a detailed prompt, iterated over many attempts, is creative work and should confer authorship. The Office’s stated reasoning for rejecting it is mechanical rather than aesthetic, and understanding it explains why longer prompts have not helped anyone.

The claim is that a prompt does not determine the expressive output. A generative model maps a prompt to a distribution over outputs and a sampler draws from it; the same prompt yields different images, and the specific expression in any one of them — the arrangement of pixels, the exact line of text — is chosen by the system, not by the person. On that view the prompt is closer to an instruction to a commissioned artist, which has never conferred authorship on the person commissioning, than to the execution of a work. Whether that analysis survives contact with tools that give the user fine, deterministic, iterative control over specific regions of an output is an open question, and one where the Office has signalled that control over expressive elements is the thing it is looking for.

It is also worth being explicit that this is a question about US law. The UK has a distinct statutory provision for computer-generated works with no human author, section 9(3) of the Copyright, Designs and Patents Act 1988, which deems the author to be the person who made the arrangements necessary — a rule with no US analogue, and one whose application to modern generative systems is itself contested. Do not carry a US conclusion across a border.

What this means when you register

If you are filing an application for a work with generated components, the practical shape is: identify the human-authored material, claim it specifically, disclaim the generated material in the application, and keep contemporaneous records of who did what. The records matter because the question is factual and will be asked, if at all, years later.

If you are on the other side — deciding whether you can use somebody else’s AI-generated image freely because it is “public domain” — be careful. The absence of copyright in generated pixels does not dispose of trade mark, publicity rights, contractual terms in the generating service’s licence, or copyright in a human-authored element you did not notice. Several of those bind you regardless of what the Copyright Office thinks.

The adjacent questions are covered separately: for the general position on ownership of model output see AI output copyright, and for the input side — whether the training itself infringes — see the four-factor test applied to training.