The AI logo you generated might already belong to someone else

Key Takeaways

• An AI-generated logo or mascot can be a fully enforceable trademark. Trademark law protects what a mark does in the marketplace, not who or what created it, so it never asks the human-authorship question that trips up copyright.
• Copyright is the weakest link for AI branding. U.S. law requires a human author, and a logo built from a bare prompt usually cannot be registered. Real copyright only attaches to the human creative choices layered on top.
• Protection comes from sequence, not luck: a clearance search before you print, a federal trademark application filed early, documented use in commerce, and trade dress for the look of the package itself.

A snack-brand owner spends an afternoon with Midjourney, generates a cartoon fox for her packaging, strips the background out with a free online tool so the mascot drops cleanly onto her label, prints ten thousand retail boxes, and six months later spots a near-identical fox on a competitor's product two states over. She calls a lawyer expecting to hear she has been robbed. Instead she learns that the question she has been asking — can I copyright this? — is the wrong one, and that the protection she actually needs has almost nothing to do with copyright.

That gap is what this guide is about. Tens of millions of small businesses now use generative tools like Midjourney, ChatGPT, Gemini, DALL·E, and Adobe Firefly to produce logos, mascots, and labels that end up printed on physical products. A belief has spread alongside the tools. This belief is that anything AI made is unprotectable, leaving the business defenseless against copycats. That belief is false, and acting on it costs founders money and leverage. What follows is a working map of the law as it stands in mid-2026 — where AI-generated works sit in U.S. copyright, how Canada and the EU diverge, how much human input counts as “enough,” the copyright-versus-trademark distinction most coverage botches, what enforcement looks like when someone copies you, and a checklist to run before the next print job.

The core misconception: “AI made it, so nobody owns it” is one of the most common things intellectual-property attorneys hear, and it is wrong. Trademark, trade dress, contract, and publicity rights all continue to apply to AI-assisted work.

The current legal status of AI-generated works in the United States

Let's begin with the rule that everyone has heard and most people misunderstand. U.S. copyright requires a human author. That principle is old, and in the AI era it was tested by Stephen Thaler, who tried to register an image he credited entirely to an autonomous system called the “Creativity Machine.” The Copyright Office refused. Thaler sued, and in Thaler v. Perlmutter the U.S. Court of Appeals for the D.C. Circuit affirmed on 18 March 2025 that the Copyright Act “requires all work to be authored in the first instance by a human being.” The district court (Judge Beryl Howell, August 2023) had called human authorship “a bedrock requirement.” The Supreme Court declined to hear the case on 2 March 2026, so the rule is settled: a machine cannot be an author.

What Thaler did not decide is just as important. It addressed a work listing the machine as sole author. It said nothing about a human who uses AI as a tool and then claims authorship of their own contribution. Aaron Hall, an attorney for business owners in Minneapolis whose explainer videos reach a wide small-business audience, draws the line narrowly. “The Office's core point is sound: something a machine generates with no human creativity is not yours,” Hall said. His disagreement is with the further claim that prompting can never count, however much creative direction goes in. “That is the Copyright Office's current position, not settled law: no court has adopted it, and it is being challenged now.” Read him precisely: courts have affirmed the human-authorship rule, but the sub-question — whether heavy prompt engineering supplies that authorship — has not been decided.

The Copyright Office set out how it draws the line in its January 2025 report Copyright and Artificial Intelligence, Part 2: Copyrightability, published 29 January 2025. Its conclusion: AI output can be protected only where a human determined enough of the expressive elements. A human-authored work that remains perceptible in the output qualifies. So does creative selection, arrangement, or modification of AI material. The bare provision of prompts, on the Office's view, does not — because prompts influence the result without controlling how the system fills in the expressive detail.

Two registration decisions show the test in practice, and both reward reading because they mark the boundary precisely:

• Zarya of the Dawn (decision letter dated 21 February 2023). Artist Kristina Kashtanova made a comic book with Midjourney images. The Office registered the parts she authored — the written text and the selection, coordination, and arrangement of the visual elements — and excluded the individual AI-generated panels because she did not author the images themselves. Same book, two different answers, split exactly along the human-input line.
• Théâtre D'opéra Spatial (Review Board decision dated 5 September 2023). Jason Allen's image won a competition, then was refused registration. Allen had run more than six hundred text prompts through Midjourney and edited the result, but he would not disclaim the AI-generated portion, and the Board held that what remained was more than a trivial amount of unprotectable machine output.

The hard question is now in front of a federal court. Allen sued the Copyright Office, and Allen v. Perlmutter (U.S. District Court for the District of Colorado, No. 1:24-cv-02665) is at the cross-summary-judgment stage as of mid-2026, with no merits ruling yet. An amicus brief calls it the first time a federal court will review the Office's authorship test as applied to prompt-based image generation. Until it is decided, the Office's “prompts are not enough” position is administrative policy, not a holding any court has adopted.

Key divergences: Canada, the United Kingdom, and the EU

Cross the border and the picture changes, which matters because the founder prints abroad and sells across markets. The jurisdictions disagree on copyright in ways that look durable. Yet two poles hold almost everywhere: a bare prompt rarely earns protection, and substantial human reworking almost always does.

United Kingdom. Britain has something the U.S. lacks: a statute that names an author for works with no human one. The Copyright, Designs and Patents Act 1988, at section 9(3), assigns authorship of a computer-generated work to “the person by whom the arrangements necessary for the creation of the work are undertaken.” Drafted in 1988, long before diffusion models, it has survived repeated reviews. A government consultation ran from December 2024 to February 2025 and drew more than eleven thousand responses; the resulting Report on Copyright and Artificial Intelligence (18 March 2026) proposes removing the section 9(3) protection but has not yet become law.

European Union. The EU runs a human-centered originality test from Infopaq International v. Danske Dagblades Forening (Court of Justice, 16 July 2009), which protects a work only where it is “the author's own intellectual creation.” Pure machine output strains against that. Layered on top is the EU Artificial Intelligence Act (Regulation 2024/1689), whose Article 50 sets transparency duties. The allocation is friendlier to a small brand than founders fear: Article 50(2) puts the duty to machine-mark synthetic output on the provider of the system (Midjourney, OpenAI), while Article 50(4) targets deployers only for deepfakes and AI text on matters of public interest — not a cereal-box mascot. Those duties apply from 2 August 2026, and the Commission published draft Article 50 guidelines on 8 May 2026 to explain them.

Canada. Canada has been consulting rather than legislating. Innovation, Science and Economic Development Canada, with Canadian Heritage, ran a Consultation on Copyright in the Age of Generative Artificial Intelligence, covering text-and-data mining, authorship of AI output, and infringement liability. It is closed; no statute has followed. The practical takeaway: with no settled Canadian answer, the conservative U.S./EU approach travels well there.

The human-authorship threshold: what “enough” input looks like

Every system above converges on one idea: protection tracks human creative control. The operative question is not whether you used AI, but how much of the expression you decided. Katie Charleston, an intellectual-property attorney at Katie Charleston Law in the Indianapolis area, puts it plainly. The Copyright Office, she said, “is not measuring how hard someone worked”; it asks whether the expressive elements were “ultimately determined by a human rather than generated by a machine.” The more a person selects, edits, arranges, revises, and directs the result, she added, “the stronger the argument for copyright protection.” Picture a spectrum: at the weak end, one prompt accepted unchanged, almost nothing to claim; at the strong end, dozens of variations selected, recombined, redrawn, and mostly rejected, leaving a real human work on top of the AI layer.

The photographer analogy (useful, but not settled law)

A camera renders the image automatically, yet a photographer holds copyright because the human choices — framing, lighting, timing — are the authorship. The Supreme Court grounded photographic copyright in exactly that reasoning in Burrow-Giles Lithographic Co. v. Sarony (1884). Whether a prompt and a diffusion model work the same way is debated, and Allen v. Perlmutter is testing it now. Treat the analogy as an argument in play, not a decided rule.

Documentation is the quiet discipline almost nobody keeps: save the prompt history, the rejected variations, the layered edit files, the dated drafts. If your claim to a work ever turns on the human input you added, the evidence that you added it is the asset — the thing that converts “I worked on this” into a provable creative contribution.

Copyright versus trademark: the distinction most coverage gets wrong

Here is the pivot the headlines miss. For a logo, a mascot, or a label, copyright was never the protection that mattered most. Trademark is — and trademark asks a completely different question.

Copyright asks who authored the work. Trademark asks what the mark does in the marketplace. “Trademark rights come from using a mark in commerce to identify the source of goods or services,” Charleston said — and unlike copyright, “trademark law does not require human authorship.” An AI-generated fox can become a fully enforceable trademark on the same terms as a hand-drawn one, because the law measures the fox's job, not its parentage.

That single move dissolves most of the founder's panic. Her copyright position may be murky; her trademark position need not be. “Trademark and trade dress do the heavy lifting,” Hall said. “They protect a distinctive brand identifier in the market regardless of how it was made.” In practice, protection for branded print comes from a few layers working together:

• Federal trademark registration with the U.S. Patent and Trademark Office, which gives the owner a bundle of enforcement advantages — a legal presumption of ownership, nationwide priority, and access to stronger remedies — that an infringer would rather you did not have.
• Documented use in commerce. Even without registration, consistent use of a mark to identify your goods builds common-law rights. Dated evidence of that use is what proves them. The same logic applies to the broader bundle of brand assets a business builds around a mark — colour palette, typography, packaging system — because the more consistently those elements travel with the mark, the stronger the source-identification argument when a copycat appears.
• Trade dress, the overall look and feel of the packaging — colorway, layout, visual impression — protected under the Lanham Act, Section 43(a) (15 U.S.C. § 1125(a)) against a rival who dresses its product up to resemble yours, whether or not your mark is registered.

The step almost everyone skips

“Before investing in packaging, marketing, or a trademark application, businesses should conduct a comprehensive trademark search,” Charleston said. The risk runs in a direction founders rarely expect: a model trained on the world's existing logos can hand you something close to a mark someone else already owns, walking you straight into an infringement you never intended. The fox you think you invented may be wearing someone else's coat — and a print run makes the problem permanent the moment ink hits card.

Is the trademark route a way of dodging the copyright question? Dr. Andres Guadamuz, Reader in Intellectual Property Law at the University of Sussex and Editor-in-Chief of the Journal of World Intellectual Property, thinks the worry “isn't unreasonable” but misreads what is happening. “I'd resist the framing that copyright is being bypassed,” he said. “What's actually happening is that the market is routing around a gap by using the IP regime best suited to the function.” Trademark protects source identification, which is what a mascot does for a living; copyright protects authorship, which AI output struggles to supply. When the work is genuine — your own elements, your selection and arrangement — you can hold copyright too. The protections stack; they do not compete. For the production side once the legal groundwork is done, Jukebox's custom packaging resources cover getting brand assets print-ready.

Real enforcement scenarios: what happens when someone copies you

Scenario A — the lifted mascot. You printed ten thousand boxes with your AI-assisted fox; a competitor copies it. The instinct is to call it copyright infringement, and you might have a claim if you put genuine creative work in and can document it. But the more reliable claim is brand law. If customers identify that fox with you, and the competitor's version is likely to confuse them about who they are buying from, that is trademark infringement and false designation of origin under the Lanham Act — whether or not a copyright exists. Trade dress can protect the look of the box on top of that, and registration unlocks remedies like statutory damages and attorney's fees. Hall reduces the whole problem to timing: “Register the trademark and document your creative role before you print ten thousand boxes,” he said, “not after the copycat appears.”

Scenario B — the training-data worry. A different anxiety gets tangled with the first: were these tools built on other people's work? That fight is real, but it is being waged at the model-maker end of the chain, not the user end. In Andersen v. Stability AI (Northern District of California, No. 3:23-cv-00201), artists allege their work was scraped to train image models; the court let key claims against Stability AI proceed past dismissal in August 2024, with trial set for September 2026. In the U.K., Getty Images v. Stability AI produced a full High Court judgment on 4 November 2025 ([2025] EWHC 2863 (Ch)) in which Getty abandoned its primary copyright and database claims and won only a narrow watermark point. Whether Stability owes Getty has little to do with whether you can stop a rival from lifting your fox.

Scenario C — the contractor gap. You hired a freelancer who used AI to make your logo and never signed anything. Ownership is now genuinely unclear — not because AI was involved, but because no contract assigned the rights. A written assignment fixes this in advance; silence guarantees a dispute once the brand is worth fighting over.

A practical risk-mitigation checklist

None of the following is legal advice for your situation. It is the sequence intellectual-property attorneys say they walk clients through, ordered the way the work should happen — before the print run, not after the copycat.

• Search before you print. Run a real trademark clearance search on any name or mark headed for packaging. The generative model may hand you someone else's mark in a new coat; the search is how you catch it while changes are still cheap.
• Add real human work, and keep the records. Select, edit, arrange, redraw. Save prompts, rejected variations, and layered edit files. Stronger, better-documented input means a stronger copyright claim sitting behind the brand.
• File the trademark early. Get the application in before the big print run, not after a copy turns up. Timing determines priority, and priority determines who wins.
• Use it, and prove you used it. Trademark rights grow from use in commerce. Date your launches, keep your packaging, and hold the evidence of continuous use.
• Read the AI tool's terms of service. Midjourney, OpenAI, Adobe, and the rest each set their own terms on commercial use, ownership, and indemnity. Read them before you build a brand on top of an output. The same discipline applies to the assets that sit alongside the AI work: even open libraries like Google Fonts ship under licences (most commonly the SIL Open Font License) that still govern commercial use, and a clean licensing audit covers the typography on the packaging as well as the mascot on it.
• Get every contract in writing. If a designer, agency, employee, or contractor touched the work, the agreement should assign all rights to you and confirm they had the right to hand them over.
• File where your customers are, not just where you are. Selling across borders means filing around your markets. The protections rhyme across the U.S., U.K., and EU, but rights are territorial — you have to claim them in each place you sell.

One closing point of perspective. The loudest version of this story pits giant technology firms against giant rightsholders in landmark suits. That story matters, but it is not the one most affected by where the lines fall. “The people most affected by where these lines land are smaller creators, smaller businesses, and individual users,” Guadamuz said, “and they are mostly absent from the conversation, the lobbying, and increasingly from the policy outcomes.” The founder is one of them — and, fox in hand, considerably less defenseless than the headlines led her to believe.

The bottom line

If you are putting AI-generated art on a physical product, stop asking whether you can copyright it and start protecting the brand. Run a clearance search before you commit a design, file for federal trademark registration before your big print run, document both your use in commerce and your own creative input, and read your AI tool's terms. You almost certainly own more than the headlines suggest — but only if you claim it before a copycat shows up on a shelf two states over.

Dr. Andres Guadamuz is Reader in Intellectual Property Law at the University of Sussex and Editor-in-Chief of the Journal of World Intellectual Property. His research focuses on AI and copyright, and his work on the authorship of machine-generated output was cited by the UK House of Lords Communications and Digital Committee in its inquiry into large language models and generative AI.

Aaron Hall is a business attorney based in Minneapolis whose practice covers intellectual property, contracts, and business law. He advises company owners on trademarks, copyrights, and trade secrets, and his educational content on legal issues for small businesses reaches a wide audience.

Katie Charleston is the founding intellectual property attorney at Katie Charleston Law, PC, in the Indianapolis area. Her practice centers on trademark and copyright protection, brand strategy, and enforcement for small and growing businesses.