What Your Company Creates with AI May Not Belong to You

Jia (Cleo) Song / Mao Peng · Artificial Intelligence · 2026-07-13 · 9 min read

This article is Part Two of our series “The 3D Lifecycle of AI” (the output side), current as of mid-2026.

Introduction

Alongside infringement risk sits another problem that companies routinely overlook: the output your company generates with AI may not belong to you at all. Many teams assume that whatever they make with a tool is theirs. Under U.S. law, that assumption may not hold.

In the United States, the question is not whether AI was used. It is how much creative input a human contributed to the final product. Where human authorship falls short, the output may fall into the public domain, protected by no copyright at all.

Key Takeaways

1. U.S. law requires a human author. Purely AI-generated output is not copyrightable, and prompts alone are generally not enough to secure ownership.
2. What can be protected is usually only the layer a human contributed (selection, arrangement, and modification), not the raw AI output underneath.
3. Ownership ultimately comes down to control and documentation: build genuine human creativity into the workflow, and disclose the AI-generated material honestly when you register.

Note: This article is for general informational purposes only and does not constitute legal advice or create an attorney–client relationship; the rules and cases in this area are still evolving. Please consult qualified counsel about specific matters.

1. The Baseline Rule: A Work Must Have a Human Author

The starting point is Thaler v. Perlmutter (D.C. Circuit, March 2025; the Supreme Court declined review in March 2026).

The facts were deliberately extreme. Computer scientist Stephen Thaler built a generative AI system he called the “Creativity Machine,” which autonomously produced an image that Thaler titled “A Recent Entrance to Paradise.” In his registration application, he listed the machine as the work’s sole author and himself only as the owner, stating that the work had been generated autonomously by the machine with no human involvement. The Copyright Office refused registration under its human-authorship requirement, the federal district court upheld that decision, and the D.C. Circuit affirmed. The rule the court established is this: the Copyright Act requires that a work be authored in the first instance by a human being; a machine cannot be the author of a copyrighted work; and output generated entirely by AI on its own, without human creative contribution, cannot be registered.

But the case’s boundaries matter. Thaler engineered the dispute as a test of the extreme scenario. From start to finish he claimed no human contribution whatsoever, asserting only that the machine was the sole author, because he wanted a clean answer to a pure question: can a machine itself be an author? (His separate argument that he should count as the author because he built and ran the machine was deemed forfeited, since he never raised it in the administrative proceedings, and the court declined to address it.) The court therefore answered only that one question. The genuinely hard problem, how much human contribution in a human-machine collaboration is enough to create authorship, was never before the court. What Thaler drew is only a floor.

2. How the Copyright Office Draws the Line

The operative guidance in practice comes from the U.S. Copyright Office’s rules and registration decisions. Its report Copyright and Artificial Intelligence, Part 2 (January 2025) concluded that purely AI-generated output is not copyrightable; that prompts alone, however detailed, are generally insufficient to make the output the user’s own, because a prompt conveys unprotectable “ideas” rather than “expression” controlled by the user; but that a human’s selection, coordination, and arrangement layered on top of AI output, as well as creative modifications to it, can be protected, judged case by case.

The report also identifies a telling signal: the same set of inputs can yield different outputs, which is precisely what shows the user lacks control over the expressive elements. That reasoning does rest on a technical premise. Diffusion models such as Midjourney and Stable Diffusion break a prompt into tokens, match them against training data, and then reconstruct an image from noise, which leaves the user with little precise control over the final expression. But today’s text-to-image systems do not share a single architecture. GPT generates autoregressively, token by token, across modalities; Gemini is natively multimodal; and both follow instructions far more faithfully. So “prompts do not amount to control” reads more like a description of particular tools (diffusion models, for instance) than a rule for all AI, which is one reason the question of how much human control is enough remains contested.

Registration practice supplies more concrete boundaries. In Zarya of the Dawn, the Office concluded that for a comic whose images were generated with Midjourney, the human-written text and the selection and arrangement of the images could be registered, but the individual AI-generated images themselves could not. In Théâtre D’opéra Spatial (Jason Allen’s award-winning image), registration was refused. Allen had run more than 600 prompts through Midjourney, and the Office still found that insufficient to make him the image’s author. Taken together, the two decisions send the same message: what can be protected is the layer the human contributed, not the underlying AI-generated output.

3. The First Direct Judicial Test: Allen v. Perlmutter

Jason Allen has sued over that refusal (Allen v. Perlmutter, No. 1:24-cv-02665, U.S. District Court for the District of Colorado), asking the court to set aside the Office’s determination. Allen moved for summary judgment in August 2025, the Copyright Office cross-moved, and briefing closed in February 2026. As of June 2026, the court has not ruled.

The case will be the first direct judicial test of a core question: does intensive prompt engineering amount to the kind of human control that creates authorship? The Copyright Office’s current position is no. Whether the court agrees is worth watching closely.

4. What This Means for Companies

In practical terms, if a work in its final form is essentially “type a prompt, get a result,” the company likely can neither register it nor enforce rights in it. That means AI-generated logos, marketing copy, interface art, or code may be copied outright by a competitor, and the company will have no effective way to stop it.

Just as important, even where an output can be registered, the scope of protection is often narrow. It covers only the human-contributed selection, arrangement, and modification, not the underlying raw AI output. Put differently, a competitor cannot copy your finished product wholesale, but it can take the same raw material and make its own.

A commonly misunderstood risk hides here as well. The Copyright Office’s examination depends on applicants’ honest disclosures; it cannot screen every work for AI, and in practice plenty of works with undisclosed AI involvement have received registrations anyway. But holding a registration certificate is not the same as holding a valid, enforceable copyright. There are two reasons.

First, once the Office later learns the truth, it can cancel or narrow the registration. In Zarya of the Dawn, the Office discovered through social media that the comic’s images had been generated with Midjourney and never excluded from the claim, and it reissued the registration to cover only the human-written and human-arranged material.

Second, and more importantly, a registration certificate is only prima facie, rebuttable evidence of validity. At trial, the defendant does not have to prove the work was actually created by AI. It need only come forward with some evidence pointing to the absence of a human author, at which point the burden shifts back to the plaintiff to prove sufficient human creativity. In other words, the defendant does not have to prove “this was AI”; the plaintiff has to prove “this was human.” A registration obtained by concealing AI involvement is closer to a liability than an asset.

Accordingly, for any output the company intends to own and protect, such as brand assets, distinctive product output, or content slated for licensing, the workflow should build in genuine human creativity (selection, arrangement, editing, and creative modification), and the company should keep the working files and version history that document that process. At registration, disclose the AI-generated material honestly and describe the human author’s contribution clearly. If even the prompts themselves were AI-generated, the human contribution that can be claimed is thinner still; disclose that candidly rather than falling back on “we wrote the prompts.”

5. A Preview of the Divergence: China Takes the Opposite View

Companies operating across borders should take particular note: on this question, China and the United States reach nearly opposite conclusions. In a 2023 case, the Beijing Internet Court held that an image generated with an AI tool could qualify as a protected work, reasoning that the user had invested intellectual effort and made original creative choices during the generation process. The same AI output may therefore be ownable in China but not in the United States. Companies should structure their ownership strategy separately for each jurisdiction in which they intend to assert rights. For a detailed look at the Chinese practice, see our companion piece in this series, “One AI Product, Two Rulebooks: China vs. the U.S.”

6. Conclusion

At bottom, ownership of AI output is a question of control and documentation: can the company prove that the final product embodies genuine human creativity? Only a company that can prove it is in a position to own the work, protect it, and price it in a deal. The next installment turns to the other face of the same output side. When the output replicates a real person’s face or voice, the primary risk is no longer copyright.

Key Sources

Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied, No. 25-449 (U.S. Mar. 2, 2026); Allen v. Perlmutter, No. 1:24-cv-02665 (D. Colo. filed Sept. 26, 2024); Unicolors, Inc. v. H&M Hennes & Mauritz, L.P., 595 U.S. 178 (2022) (§411(b) standard); Ent. Research Grp. v. Genesis Creative Grp., 122 F.3d 1211 (9th Cir. 1997); Spyder Games LLC v. Mementum Lab, No. 3:25-cv-10248 (N.D. Cal. filed 2025); 17 U.S.C. §§ 410(c), 411(b); Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023); U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2 (Jan. 2025), and the registration/review decisions Zarya of the Dawn (2023) and Théâtre D’opéra Spatial (Review Board, 2023).

In This Series

This is Part Two (the output side) of our series “The 3D Lifecycle of AI.” Previously, Part One (the input side): Training AI on Copyrighted Works: What the Key Rulings Actually Established. See the full series index.