Who Owns AI-Generated Content? Often Nobody Does

If no human made the creative choices, nobody owns the result. The US Copyright Office concluded on 29 January 2025 that copyright does not extend to purely AI-generated material, and that prompts alone do not give the user enough control to count as the author. So when a vendor hands you AI-written work and the contract assigns you all intellectual property in it, the assignment can be transferring nothing at all.

What you can actually own is everything around the output: the repositories, the accounts, the data, the prompts, and the human editing that went on top.

Who owns AI-generated content?

In the United States and the European Union, output generated purely by a model has no copyright owner, because both systems require a human author. The United Kingdom is the outlier. Section 9(3) of the Copyright, Designs and Patents Act 1988 says that for a computer-generated work, “the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”. On 18 March 2026 the UK government proposed removing that protection.

Prompts are instructions, not authorship

In its report Copyright and Artificial Intelligence, Part 2: Copyrightability, published on 29 January 2025, the US Copyright Office found that “prompts do not alone provide sufficient control” over the expressive elements of an output. Rewriting the prompt does not help either. The Office said that repeatedly revising prompts “does not change this analysis”, and compared the practice to re-rolling dice.

The courts landed in the same place. On 18 March 2025 the US Court of Appeals for the D.C. Circuit held in Thaler v. Perlmutter that “the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being”. In written testimony to the Senate Judiciary intellectual property subcommittee on 12 May 2026, the Register of Copyrights confirmed that the Supreme Court denied review, leaving that decision in place.

The part you own is the part you changed

Human work on top of a generated draft does create something protectable. The same 2025 report says human authors hold copyright in their own expression that is perceptible in an output, in the creative selection, coordination or arrangement of AI-generated material, and in creative modifications of it. The protection covers what the human added. It does not reach down into the AI-generated layer underneath.

That route is well travelled. In the same 12 May 2026 testimony, the Copyright Office said it has registered more than 7,000 claims that include AI-generated material where that material was disclaimed. Disclaiming is not optional: registration guidance issued on 16 March 2023 (88 Fed. Reg. 16190) states that applicants “have a duty to disclose the inclusion of AI-generated content”, and that AI-generated content which is more than trivial has to be excluded from the claim.

The UK gives AI output an author, for now

The UK is the only major market where a work with no human author still gets a copyright term. Section 12(7) of the 1988 Act sets that term at 50 years from the end of the year the work was made, against life plus 70 for human authors.

The government now wants it gone. Its report on copyright and AI, published on 18 March 2026 after a consultation that drew 11,520 responses, says that “in the absence of evidence of its ongoing value, we propose that this specific type of protection should be removed, while copyright should continue to protect works created with AI assistance”.

The same report drops the earlier plan for a broad text and data mining exception with an opt-out, and confirms there is no UK obligation to label AI-generated content. Nothing has changed in the statute yet, so a UK business relying on section 9(3) today is relying on a rule its own government has proposed deleting.

The EU AI Act regulates the model, not your ownership

Article 53 of the EU AI Act has applied since 2 August 2025. It requires providers of general-purpose models to run a copyright policy that respects rights reservations made under Article 4(3) of Directive (EU) 2019/790, and to publish “a sufficiently detailed summary about the content used for training”. Those are obligations on the model provider.

They give you no ownership of anything the model writes for you. Regulation (EU) 2026/1744, the AI Digital Omnibus in force since 27 July 2026, amends a long list of AI Act articles, and Article 53 is not among them.

Ownership in the EU still runs through the originality test the Court of Justice set in Infopaq on 16 July 2009: protection applies only to subject matter that is “its author’s own intellectual creation”. A model output with no human creative input does not meet it. Separately, from 2 August 2026 the Commission began enforcing the AI Act transparency rules, which require AI-generated or altered content to carry machine-readable marks. Marking your content is a disclosure duty. It is not a claim of title.

Ownership of purely AI-generated output, side by side

Where	Who owns output with no human creative input	Term	Status in August 2026
United States	Nobody. Human authorship is required	Not applicable	Copyright Office report 29 Jan 2025; D.C. Circuit 18 Mar 2025, review denied
United Kingdom	The person who made the arrangements, s.9(3) CDPA 1988	50 years from creation, s.12(7)	Government proposed removing the protection, 18 Mar 2026
European Union	Nobody. Needs to be the author’s own intellectual creation	Not applicable	CJEU Infopaq, 16 Jul 2009; AI Act Art. 53 duties sit on model providers

Your model vendor assigns you the output, “if any”

Read the assignment clause closely and you will find the vendors already know all of this. Anthropic’s Commercial Terms of Service, effective 17 June 2025, say the customer “owns its Outputs” and that Anthropic “hereby assigns to Customer its right, title and interest (if any) in and to Outputs”. Two words carry the whole risk. Google takes a different route in its Workspace service terms, last modified 16 July 2026: Google “does not assert any ownership rights in any new intellectual property created in the Generated Output”. A disclaimer, not a grant.

Neither clause manufactures a copyright that never existed. What the big vendors do offer is an indemnity, and every one of them is written on the assumption that you are the party being sued.

Microsoft’s Customer Copyright Commitment defends customers against third-party IP claims about output, but only if content filters were not disabled or evaded, you had sufficient rights to the input, you did not distribute output you should have known was infringing, and, for Azure OpenAI offerings, you retained the required red team evaluation report and can hand it over when a claim arrives. Trademark claims are excluded.

Google’s Workspace indemnity covers unmodified output only, and falls away if you ignore the safety tooling or keep publishing after an infringement notice.

One gap worth stating plainly: OpenAI’s policy pages returned 403 to automated retrieval on 22 August 2026, so the terms quoted here are only the ones we could read on the day.

Training claims got settled. Output claims did not

On 20 July 2026 the Northern District of California granted final approval to the 1.5 billion dollar class settlement in Bartz v. Anthropic, over books copied for training. The order is precise about what class members gave up: they release claims “broadly related to past AI inputs”, but “they do not give up claims about past AI outputs, nor claims of any kind about future conduct (on or after August 25, 2025)”.

The output side is live elsewhere too. In Getty Images v Stability AI [2025] EWHC 2863 (Ch), decided on 4 November 2025, the English High Court found that Getty and iStock watermarks reproduced in generated images infringed the trade marks, while holding that the model weights are not themselves an infringing copy. The defendant there was the model provider. The infringing artefact was the generated image.

The clause most AI delivery contracts are missing

The fix is not a longer IP assignment. It is a delivery clause that transfers the things that can actually be transferred. Ours covers five:

1. The artefacts. Repositories, database, prompts and configuration, plus every hosting and API account created in the client’s name from day one.
2. The record of human contribution. Who edited what, and when, so a registration with a proper AI disclaimer is possible later.
3. Disclosure by deliverable. Which outputs were model-generated, and how much human editing sits on top of each.
4. The indemnity chain. Which vendor indemnity applies, what conditions keep it alive, and who stores the filter settings and evaluation evidence that the indemnity demands.
5. Honest exclusivity. No vendor can promise exclusivity in text nobody can own. What they can promise, in writing, is that your inputs, data and prompts are not reused for another client.

Point 5 is where most agency contracts quietly overreach. If your supplier will not put points 1 to 4 in writing, the reason is usually that the answer to who owns the repository at the end is not you.

What this looks like in practice

Every install we ship hands over the code, the data and the accounts, plus the prompts and configuration behind each tomte, our word for one production agent with one defined job.

As of July 2026 our own group ran 41 registered agents, 9 of them live, logging 5,450 runs with zero failures across 13 measured days, and the same handover rule applies to every one of them: see the fleet page for how that is tracked. Ownership of a working system is a list of accounts and repositories, not a paragraph of legal boilerplate about title in the output.

If the regulatory side of this is also open at your end, the map of which AI rules apply to your business covers the compliance dates. And if you would rather have someone own this ongoing, a fractional AI officer runs the roadmap, ships one automation a month and reports weekly in writing for $2,900 a month, cancel monthly.

Start async: describe your setup at /start and get a written answer within one business day. Send us the contract clause you were about to sign, and we will tell you what it actually transfers. No meetings, no discovery call.