You generated an image with Midjourney, a track with an AI music tool, or an article with a language model, and now you want to know: is it yours, can you stop copycats, and can you send takedowns? In the US, AI-generated content copyright turns on one blunt question: what did a human actually create? Purely machine-made material has no copyright owner. No one owns it, including you.

The practical picture has more layers: most real work is a human-AI mix, and the line between those layers is where ownership lives. This guide walks that line, what the law protects, how to register AI-assisted work without wrecking the registration, when you can enforce, and the flip-side risk that AI output can infringe someone else's rights as easily as it can be infringed.

The human-authorship rule: no human, no copyright

US copyright protects original works of authorship under 17 U.S.C. § 102(a), and courts and the Copyright Office have read "author" to mean a human for over a century, the Supreme Court's 1884 Burrow-Giles photography decision tied authorship to a human mind directing the creative process. The monkey selfie episode applied the same logic: an image an animal took has no author, so it has no owner.

AI forced the question again, and the answer held. In Thaler v. Perlmutter, a researcher sought registration for an image he said was autonomously created by his AI system; the D.C. Circuit affirmed in March 2025 that the Copyright Act requires a human author. A Colorado federal court reached the same conclusion that year, upholding the Office's refusal to register an award-winning Midjourney image. The Office's January 2025 copyrightability report extends the principle to everyone: protection goes only to what a human authored, judged case by case.

Know what "no owner" means in practice. Machine-only output is unprotectable, effectively public domain on arrival. Anyone can copy it, sell it, remix it, print it on merchandise, and owe you nothing. There is no claim to make, and no takedown that survives a counter-notice, because there are no rights to enforce.

Where the protectable line runs

The Copyright Office's position sorts contributions into what counts and what doesn't.

  • Prompts, even elaborate ones, generally don't count. A prompt is treated as an instruction, not expression: you describe the result you want, and the model makes the expressive choices that produce it. A hundred iterations still leave the machine doing the painting. One wrinkle: your prompt is copyrightable as your own writing, but that protects the words of the prompt, never the output made from them.
  • Your own input does count. Feed the model your sketch, your photograph, your recorded vocal, your outline, where the output reflects your expression, that contribution is protectable. You own what you made, not what the machine added.
  • Substantial editing counts. Compositing, repainting, restructuring, rewriting output until it bears your creative stamp creates a human-authored layer, protected as far as your authorship reaches.
  • Selection and arrangement count. The landmark is the Zarya of the Dawn graphic novel: in 2023 the Office cancelled protection for the Midjourney images but upheld it for the human-written text and the author's arrangement. The compilation was protectable; the raw AI pictures inside remained free for anyone.

The practical test: ask what, exactly, you would claim someone stole. If the plain answer is "what the AI made," you have no claim. If it is "the choices I made," you do, over those choices only.

Registering AI-assisted work: disclose or lose it

Registration is where AI-assisted works most often go wrong, and the failure stays silent until it detonates. When a work contains more than trivial AI-generated material, the Copyright Office requires you to disclose it and disclaim it: the registration covers your human contributions, and the AI material is excluded from the claim. Answer the application's questions about AI use completely, and describe your contribution, the text you wrote, the edits you made, the arrangement you chose.

Skipping that is not a shortcut; it is a landmine. A registration obtained by knowingly concealing AI material can be cancelled, and a US infringement suit cannot be filed without a registration under 17 U.S.C. § 411, a cancelled certificate pulls the lawsuit out from under you. Timing matters too: registering before publication, or within three months after it, preserves statutory damages and attorney's fees under 17 U.S.C. § 412. Whether registering a given work justifies the fee is covered in whether registering is worth it, for AI-assisted work, registration done frankly is the only version that pays.

Enforcing your rights: what a takedown can and can't claim

Enforcement is where the analysis gets surgical. You can enforce the human layer of an AI-assisted work: your text, your edits, your arrangement, your original inputs. Someone copies your AI-assisted book in full? You have a claim over the human parts, which, in practice, means challenging the whole copied package. Someone lifts only the raw AI illustrations and resells them? You probably have no claim at all.

What you cannot do is send a takedown claiming rights in pure machine output. A DMCA notice under 17 U.S.C. § 512(c) must come from someone authorized to act for the holder of an exclusive right, and for machine-only material no such holder exists. A notice asserting rights you know you don't have sets up misrepresentation liability under § 512(f), damages, costs, and fees included. Be precise about who can file a DMCA takedown and how to structure the notice, and understand how Section 512(f) misrepresentation claims work.

One reassuring note: once you are enforcing work you provably authored, the takedown path doesn't care whether AI helped along the way. It becomes a normal copyright dispute.

Outside the US: where AI output can be owned

The UK runs the opposite direction. Its Copyright, Designs and Patents Act contains a provision for computer-generated works that lack a human author: the author is deemed to be the person who made the arrangements necessary for creation, and protection lasts 50 years from the end of the year the work was made, shorter than the standard life-plus-70. Ireland, New Zealand, India, and Hong Kong have similar provisions, while most of continental Europe leans toward the US position, tying authorship to a human's own intellectual creation. The same image can be owned in London and unprotectable in Los Angeles.

For online enforcement this matters less than you'd think. Major platforms run US-style notice-and-takedown, and after a counter-notice, what happens after a counter-notice, the material stays down only if the claimant sues in a US court within the statutory window: rough terrain for a right that exists only in London. The wider picture of whether DMCA works internationally is covered elsewhere. Where the UK position genuinely matters is commercial: it gives AI-heavy studios something they can license and sell.

The infringement edge runs both ways

Ownership is half the risk map. AI output can infringe as readily as it can be unprotectable. Models trained on scraped works sometimes reproduce near-copies of existing images, lyrics, code, or text, memorization, and if your AI-generated image turns out to be substantially similar to an existing artwork, you are the infringer. "The tool made it" is no defense; the publisher answers for what was published. Before commercial use, logos, book covers, merchandise, anything high-visibility, reverse-image-search the output and sanity-check it against likely stylistic sources. The litigation over training data itself is still being fought; we track it in our coverage of the AI training copyright lawsuits.

The reverse direction: your work may have fed the machine. If your photos, articles, or art were scraped into training datasets, the copying-for-training question is exactly what the courts are deciding, and the outcomes will shape licensing for years. What you can enforce cleanly today is narrower: AI outputs that reproduce your specific work, recognizable copies of your image, your text, your character, are ordinary infringement, and the target is the output, not the model. When someone instead uses AI to impersonate you or pass synthetic media off as your brand's, the better routes are impersonation account removal and deepfake removal.

Keeping what you build: a working checklist

Whatever the tool, the playbook for ending up with defensible assets is consistent.

  1. Put human expression in first. Sketch, write, photograph, record, then let the AI extend or refine it. Your input is your ownership.
  2. Edit substantively, and keep the receipts. Layered files, drafts, and edit logs prove the human layer when ownership is challenged; the evidentiary groundwork is covered in proving content ownership.
  3. Register anything valuable, disclosing and disclaiming the AI portions so the certificate survives scrutiny.
  4. Read your tool's terms, and screenshot them today. Some services grant broad contractual output rights, which cannot create copyright, but do create permission, and some restrict commercial use. Terms change without notice.
  5. Mind the trademark channel. A pure-AI logo cannot be owned as authorship, but trademark rights come from use in commerce, not authorship; the difference between trademark and copyright means your unownable mascot may still be defensible as your mark.
  6. Don't build a brand on pure output. An image anyone can legally reuse tomorrow is a weak foundation, a competitor can print the same art on the same mug.

AI content copyright: the questions everyone asks

Can I copyright something I made with AI?

Sometimes. Material a human authored, your text, your edits, your arrangement, the inputs you supplied, is protectable and registrable. Material the model generated on its own is not, regardless of which tool you paid for. Registration must disclose and exclude the AI parts, and your protection reaches only your contributions.

Can I file a DMCA takedown over AI-generated content?

Only for the parts you authored. A takedown notice must assert exclusive rights, and pure machine output has no rights holder anywhere in the US system, claiming it anyway is misrepresentation under 17 U.S.C. § 512(f), with liability for damages and fees. If your human-made elements were copied, scope the notice to those.

Do I own AI output because I paid for the tool?

No. Payment buys contractual permission to use the output, not copyright. Contract terms can still matter, a tool's terms may forbid certain uses or restrict other customers, but they bind only people who agreed to them, and no agreement can create rights the copyright statute withholds. Ownership still turns on human authorship.

Is a detailed prompt enough to make me the author?

Under current Copyright Office guidance, no. A prompt is treated as an instruction: the model makes the expressive choices, and more iteration does not change that. The prompt itself is copyrightable as your writing, but that protection covers the words you typed, not the image or text the model produced from them.

What happens if I hide AI use when registering?

The Office can cancel the registration, and since an infringement suit for a US work requires a valid registration, cancellation can end the lawsuit it was meant to support, usually after you have already spent on litigation. Disclose the AI material and disclaim it; your human contributions stay fully protected.

What to do next

The edges of this field are still moving, courts, the Copyright Office, and Congress are all active on AI, and questions like prompt-based authorship and training-data licensing will keep shifting. The human-authorship core has been stable for a century. Verify the Office's current guidance at copyright.gov before making a large bet on AI-made assets.

  1. Separate your layers today. Collect drafts, layered files, prompts, and edit logs in one place, timestamping your drafts and files turns them into usable evidence of the human contribution.
  2. Register the work you care about, disclosing and disclaiming the AI material so the certificate survives scrutiny.
  3. Before any commercial release, reverse-image-search the AI components and check them against likely sources.
  4. If the work gets copied, scope the claim to what you authored and act on it, our takedown service handles enforcement end to end, and you can see pricing before deciding how much to hand off.