Copyright and AI-Generated 3D Models: What Is Protectable

What the Copyright Office says (US only)

Registration guidance, March 2023 (88 Fed. Reg. 16190, published March 16, 2023). Applicants “have a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author’s contributions”. AI-generated content “that is more than de minimis should be explicitly excluded from the application.” Applicants “should not list an AI technology or the company that provided it as an author or co-author”. On prompts: “When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship.” Selection or arrangement of AI material, or sufficiently creative modification, can support a claim, and “copyright will only protect the human-authored aspects of the work.”

Part 2 report, January 2025. “Copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements.” “Based on the functioning of current generally available technology, prompts do not alone provide sufficient control.” Also: “Whether human contributions to AI-generated outputs are sufficient to constitute authorship must be analyzed on a case-by-case basis”, and “The use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output.” The conclusion reads “Prompts alone, however, at this stage are unlikely to satisfy those requirements.”

The report spends time on the common objection that repeated prompting is creative work. Commenters argued both ways, with the Authors Guild comparing it to “spinning a roulette wheel”. The Office’s analysis comes down against prompts alone: “Prompts essentially function as instructions that convey unprotectible ideas.” So many generations of the same crate, however carefully chosen, are not by themselves what the Office looks for.

The Office describes this as one part of a series and says it will continue monitoring developments. Its AI page lists Part 1 (digital replicas, July 31, 2024) and a pre-publication Part 3 (training, May 9, 2025); we did not summarize them. The January 2025 report is the latest copyrightability text we read, and the Office’s AI page, read 2026-11-08, lists nothing later on that question.

What the courts have said (US only)

Thaler v. Perlmutter, No. 23-5233 (D.C. Cir.), decided March 18, 2025: the court read the Copyright Act as “making humanity a necessary condition for authorship”. The applicant had listed a system, the “Creativity Machine”, as sole author. The Supreme Court denied certiorari on March 2, 2026 (No. 25-449), per its docket page. A denial of certiorari is not a ruling on the merits. The opinion says it does not reach Thaler’s argument that he was the author by making and using the system, because that argument was waived before the agency, so it does not settle how much human input suffices. We also saw a pending case, Allen v. Perlmutter, mentioned only in a secondary result and did not verify its status.

An older 3D case is more useful than either for a mesh artist. In Meshwerks v. Toyota (10th Cir., June 17, 2008), a company built digital wire-frame models of Toyota vehicles, working from measurements and then hand-adjusting them. The court held the models were not protected: they “owe their designs and origins to Toyota”, and the opinion describes them as “not so much independent creations as (very good) copies of Toyota’s vehicles.” It also said it did “not doubt for an instant” that digital modeling can create fully protectable work. The Copyright Office report cites the case for that second point. Meshwerks is about faithful replicas of a real object, not AI output, but the logic transfers uncomfortably well, as the next section shows.

Applying this to a 3D model

The passages we read do not mention AI-generated 3D models, so this section is our reading of them, not something they say. The table runs from weakest to strongest human contribution.

The retopology line deserves emphasis because it runs against instinct. Cleanup is the main human work on many AI assets, and the site’s retopology guide treats it as a technical task. That is exactly why it may not count for much here: edge flow chosen to preserve an existing silhouette is closer to Meshwerks than to sculpting. No source we read decides this for AI meshes. If protection matters for an asset, add visible design decisions and record them.

The report also says protection for selections needs more than a token effort: “The selection, coordination, and arrangement of only two or three elements is not generally sufficient” (footnote 138). And the whole work survives even where parts do not: “a film that includes AI-generated special effects or background artwork is copyrightable, even if the AI effects and artwork separately are not.” The report does not mention games; applying this to them is our reading.

Image-to-3D and your own artwork

The report treats expressive inputs as different from prompts. “These types of expressive inputs, while they may be seen as a form of prompts, are different from those that merely communicate desired outcomes.” It gives a registered example, Rose Enigma, where a hand-drawn illustration was used as input and its expressive elements remained “clearly perceptible in the output”. The Office registered it with an annotation limiting the claim to “unaltered human pictorial authorship”, after the applicant disclaimed non-human expression such as “the realistic, three-dimensional representation” of features.

The principle stated: “where a human inputs their own copyrightable work and that work is perceptible in the output, they will be the author of at least that portion of the output.” It “would not extend to the AI-generated elements standing alone.” The report contrasts a refusal, Suryast, where a photograph was restyled and the output “did not clearly show” the input work.

For a 3D developer, our reading is:

• Using your own concept art as the input is the stronger starting point, and the design that survives into the mesh is what you may be able to claim.
• The depth, back faces and surface detail the generator invented are likely not yours. Those are the parts it filled in.
• Someone else’s art as input gives you no such claim, and raises its own infringement problem.
• Keep the original drawing, dated. It is the evidence that the input was yours.

If you register

The guidance gives actual form instructions (US only). Describe your contribution in the “Author Created” field. Exclude more-than-de-minimis AI content in the “Limitation of the Claim” section, under “Material Excluded” in the “Other” field, with a brief description such as “[description of content] generated by artificial intelligence.” Additional information can go in the “Note to CO” field. If you are unsure, the guidance says you may “simply provide a general statement that a work contains AI-generated material”, and the Office will follow up.

Mistakes have consequences. For an application already registered, “the applicant should correct the public record by submitting a supplementary registration”. Applicants who do not “risk losing the benefits of the registration”, the Office may take steps to cancel, and “a court may disregard a registration” if inaccurate information was knowingly provided.

Outside the US, briefly

These notes are not US law and are not a survey.

The Office’s own summary, January 2025. Part 2 reports that Korean guidance limits registration to human creative contribution, that Japanese guidelines use a case-by-case test including prompts, number of attempts, selection and later human changes, that the Beijing Internet Court in 2023 protected a Stable Diffusion image based on prompts and adjustments (the ruling is non-precedential), and that most EU member states said AI output is eligible “only if the human input in [the] creative process was significant”. Check current local sources before relying on these.

United Kingdom. The Copyright, Designs and Patents Act 1988 s.9(3) says that for a computer-generated work, “the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”, and s.178 defines computer-generated as there being “no human author of the work”. The UK government published its report on copyright and AI on 18 March 2026. According to IPKat’s commentary, no immediate change was made and the Government will keep s.9(3) under review, “with a preference to remove the provision”. Treat s.9(3) as live but unsettled.

European Union. The AI Act, Article 50(2), requires providers of AI systems generating synthetic audio, image, video or text to mark outputs in a machine-readable format (the text we read does not name 3D models); that is a transparency duty on providers, not a copyright rule. The page we read gave 2 August 2026 as the date Article 50 applies. According to law-firm commentary (Garrigues, July 24, 2026), the AI Digital Omnibus Regulation gives systems placed on the market before that date until 2 December 2026 to meet the output-labeling duty. We did not read the Official Journal text.

What this means in practice

Do not treat an AI output as a protectable asset by default. Contractual rights from the generator let you use the file, but they do not necessarily give you something you can enforce against copiers. For shipped games the practical stakes are limited (selling a game); for a business that sells models as such the stakes are higher. Anything high-stakes needs a lawyer in your jurisdiction.

Four habits keep your options open, whichever way the law moves:

1. Separate human and generated work in your files. Keep the raw generator export next to the edited version.
2. Log what a person changed and why, per asset. A dated change list beats memory.
3. For hero characters and logos, consider commissioning an artist with a written rights assignment, so ownership does not rest on this analysis. Trademark is a separate tool for names and logos and is outside this page.
4. Do not represent AI work as fully original in a contract or marketplace listing. Some sellers’ terms ask you to warrant exactly that.

More technical records and methods are in the cleanup hub.

Not covered

Countries beyond the notes above, the legal status of training data, whether a particular generator’s output infringes anything, trademark, and any development after the dates above.

More in the licensing hub.