Who Owns AI-Generated Images? Copyright Questions Every Creator Should Understand

If you generate an image with AI and change nothing else, nobody owns its copyright in the United States, not you, not the tool. The image is effectively public domain the moment it appears. You own copyright only in the parts you genuinely created: your text, your substantial edits, your creative arrangement of the results. Platform terms from OpenAI and Midjourney may say you “own” the output, but a contract cannot create copyright that federal law doesn’t recognize.

This is not legal advice. Where the law is still unsettled, a lawyer earns their fee.

So who actually owns an AI-generated image right now?

Three layers of “ownership” sit on every AI image, and creators mix them up constantly.

Copyright is the right to stop others copying a work; in the US it requires a human author: the layer that fails for purely AI-generated images.

Contract is your deal with the platform: OpenAI assigns its output rights to you; Midjourney says you own your assets. Those promises bind you and the company: they can’t conjure a copyright into existence.

Third-party rights are everyone else’s: the photographer your output resembles, the studio whose character you prompted. That’s where lawsuits live now.

Practical answer: you own what you humanly made, the platform adds nothing, and a purely prompted image belongs to no one.

What does the US Copyright Office actually say about prompts?

Short answer: prompting alone doesn’t make you an author, but your edits and arrangement can be protected: a position the Office has held since 2023, collected on its AI initiative hub.

The core rule comes from its registration guidance, published March 16, 2023: works with AI-generated material are registrable only to the extent of a human’s creative contribution, and applicants must identify and disclaim the AI parts. Hide it, and the Office can cancel the registration.

Then the big report. Part 2 of the Office’s AI study, published January 29, 2025, concluded existing copyright principles are adequate: no new legislation needed. Prompts alone, it found, don’t give a user enough human control over the output to make them the author. The flip side stands too: substantial human modifications, and creative selection or arrangement of AI pieces, can be registered: case by case.

The Office’s own decisions show how this plays out:

• Zarya of the Dawn (February 2023). Kris Kashtanova’s Midjourney-illustrated comic kept registration for her text and the arrangement of the whole, not the individual Midjourney images, as Garrigues’ explainer recounts. Prompting wasn’t authorship.
• Théâtre D’opéra Spatial (September 2023). The Review Board refused registration for Jason Allen’s award-winning Midjourney image, finding his prompting and tweaks weren’t enough; he sued in September 2024, and that case is still going.

A third report, on training models on copyrighted works, landed in pre-publication form in May 2025: it matters more to AI companies than image users.

Did any court settle the question? Thaler v. Perlmutter, explained

Yes, for purely AI-generated works, it’s settled.

Stephen Thaler tried to register an image his “Creativity Machine” generated autonomously, listing the machine as author. The Copyright Office refused, and the D.C. Circuit affirmed on March 18, 2025: human authorship is a “bedrock requirement” of the Copyright Act: a work must be “authored in the first instance by a human being.”

Thaler took it to the Supreme Court, which denied certiorari on March 2, 2026 (docket No. 25-449). The petition had asked whether works output by an AI system without a direct, traditional authorial contribution from a natural person can be copyrighted.

Human authorship is a “bedrock requirement”, and after March 2, 2026, it’s a requirement the Supreme Court declined to revisit.

What’s not settled is the AI-assisted middle ground: how much editing or arranging turns a tool user into an author. Allen’s case is the live test. Until then, protect what you humanly made and disclaim what the machine made.

What do the platform terms actually say you own?

Here’s what each promises, and the catch:

Platform	What the terms say about output ownership	The catch
OpenAI (ChatGPT / DALL·E / GPT Image)	You “retain your ownership rights in Input and… own the Output”; OpenAI “assign[s] to you all our right, title, and interest, if any, in and to Output”	Applies “to the extent permitted by applicable law”; others may get similar output, so no exclusivity
Midjourney	“You own all Assets You create with the Services to the fullest extent possible under applicable law”	Companies over $1,000,000 USD annual revenue need Pro or Mega to own assets; Midjourney keeps a broad license back; content is public and remixable by default
Google Gemini (API)	“Google won’t claim ownership over that content”	Google “may generate the same or similar content for others”; consumer-app terms differ from API terms
Adobe Firefly	Marketed as “commercially safe,” trained on licensed Adobe Stock and public-domain content; IP indemnity for eligible enterprise customers	Indemnity is a contract, not a copyright; it applies when you use the product within its terms

Quote sources: OpenAI’s Terms of Use, Midjourney’s Terms of Service (effective May 27, 2026), and the Gemini API Additional Terms of Service.

These clauses do less than people assume. OpenAI assigns all its “right, title, and interest, if any”, and with no copyright in the output, there’s nothing to assign. You genuinely get a promise: OpenAI won’t claim the image later, and Midjourney won’t sue you for using your own assets. Worth having. Just not copyright.

Midjourney’s fine print matters in 2026. The terms effective May 27, 2026 state plainly: “If you are a company or any employee of a company with more than $1,000,000 USD a year in revenue, you must be subscribed to a ‘Pro’ or ‘Mega’ plan to own Your Assets.” A startup riding Basic plans has a contract problem, not just a pricing problem.

Adobe is the outlier. Since June 2023 it has offered IP indemnification for Firefly output to enterprise customers. Adobe steps in on covered IP claims when the customer used the product within its terms. Indemnity is a contractual promise to cover your legal exposure, the closest thing this industry sells to a safety net.

Who owns what, scenario by scenario?

Here’s the decision table, assuming US law as of November 2026.

What you did	Who owns the copyright (US)	Why
Typed a prompt, used the raw output as-is	Nobody: effectively public domain	Prompts alone don’t provide enough human control (Part 2 report, Jan 2025); human authorship is a bedrock requirement (Thaler, D.C. Cir. 2025)
Prompted, then substantially edited the output in Photoshop	You own your modifications, not the untouched AI layer	Human modifications are registrable; the AI material must be disclosed and disclaimed
Generated variations, curated and arranged them into a comic or campaign	You own the selection and arrangement, plus your text	Exactly what survived in Zarya of the Dawn
Used AI output as a reference, created the final work yourself	You own the whole new work	The expressive choices are yours; the AI was a tool in your process
Company over $1M revenue prompting on Midjourney Basic or Standard	No ownership even under the contract unless you upgrade to Pro or Mega	Midjourney’s terms effective May 27, 2026 make ownership conditional on plan tier
Used Adobe Firefly through an enterprise contract, output as-is	Contractual indemnity for covered claims: no copyright in the raw output	Indemnity is a contract promise; copyright still depends on human contribution

What’s happening with the training-data lawsuits?

None decide who owns your output, but they shape the risk around it. Where they stood in November 2026:

• The New York Times v. OpenAI and Microsoft (S.D.N.Y.). Both sides filed cross-motions for summary judgment on fair use on September 4, 2026 before Judge Sidney Stein. The Justice Department had filed a statement of interest the day before arguing AI training is fair use; amicus briefs ran through October 30, 2026. No ruling yet.
• Getty Images v. Stability AI. The UK High Court ruled on November 4, 2025 ([2025] EWHC 2863 (Ch)): Getty’s secondary copyright claim failed (the model itself is not an “infringing copy”) and it won only narrowly on trademark over reproduced watermarks. In the US, Getty dismissed its Delaware case in August 2025, refiled in California, and its trademark claims survived a motion to dismiss in April 2026.
• Disney, Universal, and Warner Bros. v. Midjourney (C.D. Cal., filed June 11, 2025). The studios allege the tool reproduces their characters and that Midjourney marketed that ability. Midjourney answered in late 2025 and moved for judgment on the pleadings in July 2026. Still pre-trial.
• Andersen v. Stability AI (the artists’ class action from January 2023). Still in discovery, with expert reports due beginning December 8, 2026.

On training itself, mid-2025 split: Bartz v. Anthropic found training transformative but wouldn’t excuse storing pirated copies (a $1.5 billion settlement followed), while Kadrey v. Meta was more forgiving. The output-ownership rules won’t move. Your risk moves if you deliberately generate and sell protected characters.

What does the EU AI Act require if you publish AI images?

Labeling: effective August 2, 2026. Article 50 of the EU AI Act puts two duties on image creators. Providers of AI systems generating synthetic images must mark outputs in a machine-readable format “detectable as artificially generated or manipulated.” Deployers: you, when you publish: must disclose deepfakes: images that artificially generate or manipulate real people or events. Artistic, satirical, or fictional works get an exemption, so long as disclosure doesn’t spoil the work.

The European Commission published a supporting Code of Practice in 2026 covering provider marking and deployer labeling.

Notice what this isn’t: copyright. The EU has no special copyright rule for AI-generated works. Selling images there? Treat labeling as a compliance cost like fonts or model releases.

A creator checklist you can actually use

1. Document your human process as you go. Save layered files, versions, and notes showing what you changed after generation: contemporaneous records beat reconstruction.
2. Disclose and disclaim AI material when registering. Follow the March 2023 guidance: identify the AI content and exclude it from your claim. Zarya shows what happens when it’s omitted.
3. Register the human-authored elements, and register promptly. Registration is required before filing a US infringement suit, and timely registration unlocks statutory damages and attorney’s fees.
4. Match your plan tier to your business. Check the current terms for your tool: Midjourney’s $1 million revenue rule is the sharpest edge, and consumer tiers elsewhere may lack commercial rights.
5. Scale indemnity with the stakes. A blog thumbnail and a national ad campaign carry different risk; for the latter, indemnified generation (Adobe Firefly enterprise) or legal review is worth the money.
6. Don’t prompt for protected characters and trade on them. The studios’ suits target the tool, but a user deliberately selling a recognizable character is the obvious next defendant. That is not a court holding.
7. Label AI images for EU audiences. Since August 2, 2026, Article 50 duties apply to deployers publishing synthetic content there.
8. Know when to call a lawyer. Registering a contested work, licensing images to clients, responding to a takedown, building a campaign on AI imagery: all lawyer territory. One consult costs less than a claim.

The short version

• Hobbyists: keep creating. Pure AI output is free to use and free to lose: anyone can copy it. Add real edits if you want rights.
• Freelancers: your deliverable is your human contribution. Document it, register it for work that matters, and put AI-tool disclosure in your contracts.
• Small businesses: check plan tiers against your revenue, and don’t build a brand identity from images you can’t enforce.
• Agencies: indemnity is the product you’re buying. Adobe’s enterprise Firefly terms (since June 2023) are the benchmark: read what’s covered.

FAQ

Who owns the copyright to an AI-generated image?

In the US, nobody, if the image came straight from the tool with no meaningful human creative input. The Copyright Office and the D.C. Circuit (Thaler v. Perlmutter, left standing when the Supreme Court denied review on March 2, 2026) treat human authorship as mandatory. Platform terms may say you own the output, but contract rights can’t substitute for a copyright that was never created.

Can you copyright an image you made with AI?

You can register the human-authored parts: text you wrote, substantial edits you made to the AI output, and creative selection and arrangement. The March 2023 registration guidance requires you to disclose and disclaim AI-generated material. Purely prompted output stays unprotectable, as Jason Allen’s ongoing appeal over Théâtre D’opéra Spatial continues to test.

Are AI-generated images public domain?

Effectively yes in the US. With no author, there’s no copyright to enforce, so anyone can copy, remix, or resell a purely AI-generated image. That cuts both ways: you can use them freely, but you can’t stop anyone else from doing the same. Other countries differ: a Beijing court protected an AI image where the user made enough creative choices.

Do I own my Midjourney images?

Midjourney’s terms (effective May 27, 2026) say you own the assets you create “to the fullest extent possible under applicable law,” and ownership survives cancellation. Companies with more than $1 million in annual revenue must be on a Pro or Mega plan to own assets. In exchange, Midjourney takes a broad license, and your content is public and remixable by default.

Can I sell AI-generated images commercially?

Usually yes. No US law bans selling AI-generated images, and paid plans on major platforms grant commercial rights. The practical limits are your platform’s license terms, the fact that unprotected images can be copied freely, and the risk that an output too closely resembles an existing character or photograph. For commercial campaigns, indemnified tools like Adobe Firefly’s enterprise offering exist for a reason.

Does OpenAI claim ownership of images made with DALL-E or GPT Image?

No. OpenAI’s Terms of Use say you retain your input and own the output, with OpenAI assigning whatever rights it might have to you. But the grant applies “to the extent permitted by applicable law,” and OpenAI notes other users may receive similar output, so there’s no exclusivity, and US copyright still requires human authorship.

What happens if an AI image I used infringes someone’s copyright?

Liability is unsettled, and that’s the honest answer. Studios are suing Midjourney over character outputs, and courts haven’t sorted out where tool, user, or nobody bears responsibility. A user knowingly selling a Mickey Mouse lookalike faces real risk; a generic landscape is a different universe. For anything high-stakes, get legal advice. Adobe enterprise contracts can shift some risk contractually.

What are the main alternatives for commercially safer AI images?

Adobe Firefly, which trains on licensed Adobe Stock and public-domain content and offers IP indemnification for eligible enterprise customers. Getty’s generative AI service similarly builds on licensed imagery. “Safer” here means training data and contractual backup, not guaranteed copyright: human authorship rules apply to every tool.

How do I register copyright in an AI-assisted work?

Keep records of your human contributions, apply through the Copyright Office, and identify AI-generated material in the application, disclaiming it in the limitation-of-claim section. Register what’s human: your text, your edits, your arrangement. Registration is required before filing a US infringement suit, and prompt registration is what unlocks statutory damages and attorney’s fees.

What should creators do about AI image ownership right now?

Three habits cover most people: document your creative process so your human contributions are provable; match your platform plan to your actual business (the Midjourney $1 million rule catches companies out); and treat indemnity and legal review as project costs for big commercial uses. If a deal, brand, or campaign rides on an image’s legal status, talk to a copyright lawyer before you publish.

Sources

• U.S. Copyright Office: Copyright and Artificial Intelligence initiative hub
• Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 29, 2025, PDF)
• Federal Register: Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (Mar. 16, 2023)
• SCOTUSblog: Thaler v. Perlmutter case file (certiorari denied Mar. 2, 2026)
• Thaler v. Perlmutter, D.C. Circuit opinion (Mar. 18, 2025, PDF)
• OpenAI: Terms of Use
• Midjourney. Terms of Service (effective May 27, 2026)
• Google: Gemini API Additional Terms of Service
• Adobe. Firefly product page (“commercially safe” positioning)
• Computerworld: Adobe offers copyright indemnification for Firefly users (June 2023)
• Garrigues: Copyright and AI-generated works: Zarya of the Dawn
• Wikipedia. Théâtre D’opéra Spatial (USCO refusal and Allen’s appeal)
• Latham & Watkins: Getty Images v Stability AI: English High Court rejects secondary copyright claim (Nov. 2025)
• Norton Rose Fulbright: AI in litigation series: an update on AI copyright cases in 2026
• Wikipedia: The New York Times v. Microsoft and OpenAI (2026 summary judgment briefing)
• Future of Life Institute: EU AI Act Article 50: Transparency obligations
• European Commission: Code of Practice on Transparency of AI-generated Content
• Wolters Kluwer: Beijing Internet Court grants copyright to AI-generated image