UK computer-generated works and generative AI

Every few months someone points at section 9(3) of the Copyright, Designs and Patents Act 1988 and concludes the UK solved AI copyright forty years early. The provision is real, it is still on the statute book, and the March 2026 government report left it there. What it is not is a plan you can build a rights position on.

The honest status: contested, untested against generative models, and load-bearing in exactly the place it is weakest. Here is what the section says, where it breaks, and what to do instead.

This is a map of the argument, not legal advice.

What the section actually says

Section 9(3) provides that where a literary, dramatic, musical or artistic work is computer-generated, the author is taken to be the person by whom the arrangements necessary for the creation of the work are undertaken. Section 178 defines "computer-generated" as generated by computer in circumstances such that there is no human author of the work.

Two features follow that people rarely mention when they cite it.

First, it is a fallback. By its own definition it only engages where there is no human author. If your contribution rises to the level of authorship, you are outside s.9(3) entirely and inside the ordinary rules. The section is for the residue.

Second, the protection it grants is not the ordinary protection. Computer-generated works get a fixed term measured in decades from creation rather than the author's life plus seventy, and no moral rights attach. Whatever s.9(3) is, it is a thinner thing than the copyright people assume they are claiming.

Note also what it does not cover: films and sound recordings are separate categories in the Act with their own authorship rules. A generated video is not obviously "an artistic work" in the s.9(3) sense, which is its own unresolved problem for anyone in this business.

Three places it breaks on generative AI

Who undertook the arrangements? Read literally, the phrase could point at the person who typed the prompt, at whoever built and trained the model, or at the platform operator who assembled the pipeline the generation ran through. Nothing in the section resolves it, and no reported decision has applied it to a generative model. A drafter in 1988 was picturing something much more like a bespoke program run by the person who commissioned it. Modern generation splits that role across at least three parties who have never negotiated with each other about it.

The originality problem. UK originality is now understood through the "author's own intellectual creation" standard. That formulation presupposes an author and their intellectual creation. A work defined by s.178 as having no human author sits awkwardly against a standard framed entirely around human choice. Academics have been pointing at this tension for years. It has never been resolved by a court, which means an s.9(3) claim carries a live argument that the work fails the originality threshold it needs to clear.

Scope, if it works at all. Suppose you win the "arrangements" argument and clear originality. You hold a fixed-term right in a specific generated file. That does not stop anyone generating something similar from the same model, and it does not stop the model producing near-identical output for the next user with the next prompt. A right that only bites on literal copying of one file is worth less in practice than the confidence people place in it.

The March 2026 report did not fix this

The government's copyright and AI report, published on 18 March 2026, abandoned the proposed broad text-and-data-mining exception with a rightsholder opt-out and maintained the status quo. Section 9(3) survived the process untouched and unclarified. That is the important part for anyone who was waiting for a legislative answer: the consultation that could have settled the computer-generated works question closed without settling it.

So the position going into the second half of 2026 is the position that existed before: a 1988 provision, no applying authority, and an active academic argument that it cannot do the job.

Getty did not fix it either

The obvious hope was that litigation would produce the answer instead. Getty Images v. Stability AI went to trial in the England and Wales High Court and delivered judgment on 4 November 2025, and it produced almost nothing on the underlying questions. Getty dropped its training and output claims mid-trial for want of UK territorial evidence. The secondary-infringement claim failed. What survived was an extremely limited trade mark finding on watermarks.

UK training legality was never decided, and neither was anything about ownership of generated output. The case that was supposed to be the UK's landmark turned into an evidential cautionary tale about where model training physically happens.

What to rely on instead

The route that works is the one that works everywhere: put a human author in the chain and record what they did.

• Build the human contribution into the work, not around it. Selection from a larger pool, arrangement, and modification of specific outputs are recognised forms of authorship that do not depend on s.9(3) at all. An EDL-based editor makes those decisions durable and inspectable, because the timeline is literally the list of them.
• Keep the record at generation time. Model and version, prompt, references, seeds, and what a human changed afterwards. Asset naming and version discipline is the boring infrastructure this sits on.
• Handle the commercial question by contract. Ownership of the file, warranties, indemnities and permitted uses are all allocable in writing regardless of whether copyright subsists. Legal and licensing for AI content in business and the usage rights definition cover how that allocation is normally written.
• Check the model's own licence separately. Statutory copyright and the terms you generated under are independent constraints, and open-weight releases in particular carry conditions of their own. Open-weight video licences and what you can ship walks through the ones that actually restrict commercial output.

The pattern to avoid is the one where a UK client asks "do we own this?", someone cites s.9(3) from memory, and the deliverable ships on a provision no court has ever applied to a diffusion model. If ownership genuinely matters for the asset, build authorship into it and say so in the contract. What to tell clients about using AI covers how to have that conversation without overpromising.

FAQ

Is section 9(3) still in force?

Yes. It remains on the statute book, and the 18 March 2026 report maintained the status quo rather than amending or repealing it. Being in force and being applicable to generative AI are different questions, and only the first one has a clear answer.

If s.9(3) applies, who owns the output — me or the model provider?

Unresolved. The section assigns authorship to whoever undertook "the arrangements necessary for the creation of the work", and there is no authority applying that phrase to a generative pipeline where the prompt, the model and the infrastructure come from different parties. In practice the answer you actually operate on comes from your agreement with the provider, not from the statute.

Does the UK position help if my work is used in the US?

No. US law requires human authorship, a position reinforced when the Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026. A UK computer-generated works argument does not travel, and most content businesses need a position that holds in both markets. The human-contribution route is the one that does.

Should I stop telling clients the UK is friendlier on AI copyright?

Yes, unless you can also explain the three open questions above. "The UK has a provision that might apply and has never been tested" is an accurate sentence. "The UK grants you copyright in AI output" is not, and it is the kind of assurance that turns into a warranty claim when someone reads it back to you.