Can you copyright AI-generated content? If your company generates images, code, copy, music, or video with AI, the honest answer to the ownership question is uncomfortable: material created entirely by a generative AI system is not protected by US copyright, and no amount of prompt engineering changes that. The 2025 US Copyright Office guidance is direct about it. That does not make AI useless for your business, but it does mean the thing you thought you owned may be free for anyone to copy, which is a strategic problem worth understanding before you build a product or a brand on top of it.
This guide covers what the current rules actually say, where the line between protectable and unprotectable sits, why the provider's terms of service do not save you, and the practical ways founders protect value when copyright will not. It pairs with the broader AI startup legal checklist and the deeper question of who owns the code a contractor writes.
The rule: copyright protects human authorship
Human authorship is the foundation of US copyright. In its 2025 report on the copyrightability of AI output, the Copyright Office confirmed that a work generated entirely by AI cannot be registered, and that providing a prompt, however detailed, does not by itself make the user the author. The reasoning is that current systems translate a prompt into an output through processes the user does not control at the level of specific expression. You are describing what you want; the machine is deciding how it looks.
This is not a one-off administrative position. A federal appeals court reached the same place in 2025, affirming that a work with no human author cannot be registered. So the human-authorship requirement is not a temporary quirk of one agency. It is the baseline.
Where the line actually sits
The rule is not "anything touched by AI is unprotectable." It is narrower and more workable than the headlines suggest. Copyright can still attach where a human contributes enough original expression. Three patterns tend to qualify:
- Human-authored elements inside an AI-assisted work. If you write the text and use AI to assist, your text is yours. AI in the workflow does not poison the human parts.
- Creative selection and arrangement. If you take AI-generated components and arrange or curate them in a sufficiently creative way, that arrangement can be protected, even if the individual pieces are not.
- Meaningful modification. If a human materially edits, reworks, or adds to the AI output so the final result reflects human creative choices, those choices can carry protection.
What does not qualify is the raw output itself, and the analysis is done case by case. That means two companies using the same tool can end up with very different protection depending on how much a human shaped the result and whether they can show it.
The output your competitor can copy for free is not an asset. The human judgment layered on top of it is where the defensible value lives.
Why the provider's terms do not solve it
Founders often point to the fine print: the major AI providers' terms of service assign the output to the user. That is real and useful, but it is doing less than people assume. A terms-of-service clause transfers whatever rights the provider has. If the output is not copyrightable, there is no copyright to transfer, so you receive a contractual promise about something that may not exist as property. You can use the output, but you may not be able to stop anyone else from using an identical or near-identical version. Ownership by contract and protectability under copyright are two different questions, and only one of them keeps a competitor out.
There is a second exposure hiding in the same place: whether the output infringes someone else's work, and who bears that risk. Some providers now offer limited indemnities; the scope and conditions vary a lot. That belongs in your review of AI vendor and model-provider terms, because output rights, training use, and indemnity all live in the same contract.
How founders actually protect the value
The move is not to avoid AI. It is to stop relying on copyright for things copyright will not cover, and to route the value through protections that do work.
- Add and document human authorship. Where a work matters, make sure a human is genuinely selecting, arranging, and editing, and keep a record of that contribution. If you register the work, be prepared to disclose the AI-generated material and identify the human-authored parts.
- Use trade secret for the parts you never publish. Your fine-tuned model, training data, prompts, pipelines, and internal tooling can be protected as trade secrets if you actually treat them as secret: access controls, confidentiality terms, and need-to-know handling.
- Let contracts do the work copyright cannot. Customer terms, licensing, and confidentiality agreements can create enforceable rights and restrictions around AI-assisted deliverables even where copyright is thin. This is ordinary contract leverage, and it is often the strongest lever you have.
- Protect the brand with trademark. Your name, logo, and product marks are protectable regardless of how the underlying content is made. For a company whose output is hard to copyright, brand becomes disproportionately important. See how to trademark your startup name.
The mistake most founders make
The common error is treating "we generated it, so we own it" as the end of the analysis, then discovering during a financing or an acquisition that the company's core content is not protectable and the paper trail does not show any human authorship. Diligence teams have started asking how AI is used and what the company actually owns, which makes this exactly the kind of thing that is cheap to get right early and expensive to reconstruct later. It sits alongside the other items a buyer examines in a startup legal due diligence review.
- Content generated entirely by AI is not protected by US copyright, and prompts alone do not create authorship.
- Copyright can still protect the human-authored parts: original text, creative selection and arrangement, and meaningful edits to AI output.
- Provider terms assign output by contract, but cannot create a copyright that does not exist, so competitors may be free to copy raw output.
- Protect value through documented human authorship, trade secret, contracts, and trademark, not copyright alone.
- Expect diligence to ask how you use AI and what you own; document it before you need to.
The takeaway is not that AI content is worthless. It is that the value moves. It moves out of the raw output, which anyone can regenerate, and into the human judgment, the proprietary data and models, the contracts, and the brand. Build your protection there, and the AI stops being a liability on the cap table's IP schedule and starts being leverage.
Related reading: the AI startup legal checklist, AI vendor and model-provider terms, and who owns your startup's code. Or start a conversation about your IP position.