The UK's unique provision - and its strain
The UK is unusual in having a statutory provision for computer-generated works. Section 9(3) of the CDPA 1988 says the author of a work generated by computer in circumstances where there is no human author (section 178) is the person by whom the arrangements necessary for the creation of the work are undertaken; such works get 50 years of protection (section 12(7)) and no moral rights.
But section 9(3) was drafted in 1988 for deterministic programs, not large language models or image generators. Applied to generative AI, who made the arrangements necessary admits several answers - the model developer, the host of the service, the party who fine-tuned it, or the user who wrote the prompt. Nova Productions Ltd v Mazooma Games Ltd [2007] EWCA Civ 219 (a video-game case) treated the developer, not the player, as the author of computer-generated frames - suggesting the developer rather than the user - but the analogy strains where a prompt involves real creative input. The provision also sits in tension with the post-Infopaq originality standard (the author's own intellectual creation), and its future is under review.
AI cannot be an author or inventor
The courts are clear that AI itself cannot hold rights. In Thaler v Comptroller-General of Patents [2023] UKSC 49, the Supreme Court held that an inventor under the Patents Act 1977 must be a natural person - but Lord Kitchin's reasoning left the door open for AI-assisted inventions where a human is identified as the inventor using the AI as a tool. The practical imperative is to ensure a human inventor is always identified when AI contributes to an inventive process.
The position is consistent across major jurisdictions: Thaler v Vidal (43 F.4th 1207, Fed Cir 2022) reached the same result on US patents, Thaler v Perlmutter (affirmed by the DC Circuit in March 2025) held humanity is required for copyright authorship, and Naruto v Slater (888 F.3d 418, 9th Cir 2018) held non-human entities lack standing to sue for copyright.
The US and EU positions
US law gives no copyright to purely AI-generated material. The US Copyright Office's January 2025 report reaffirmed that prompts alone generally do not make output copyrightable, because the model's behaviour is not controlled by the user - but works containing AI-generated material can be protected where a human exercises sufficient control over the expressive elements. The Zarya of the Dawn decision illustrates this: registration covered the human-authored text and the selection and arrangement of images as a compilation, but the individual AI-generated images were excluded.
Most EU jurisdictions do not recognise computer-generated works at all, requiring human authorship and creativity (Infopaq, Painer). The EU AI Act (Regulation (EU) 2024/1689) does not address ownership of output; its copyright provisions focus on the input side - Article 53 requires providers of general-purpose AI models to publish a sufficiently detailed summary of their training content so rights holders can enforce their reservations.
Getty v Stability AI: territoriality
The first major UK judgment on generative AI turned on territoriality rather than the metaphysics of training. In Getty Images v Stability AI [2025] EWHC 2863 (Ch), Getty abandoned its primary copyright claim because there was no evidence the model had been trained in the UK, and the court dismissed the remaining secondary-infringement claim (importing the model into the UK) on the basis that the model's weights do not store or reproduce the training images, only derived numerical parameters.
The practical signal is that where training occurs, and how a model is distributed, may matter more than whether training data contained protected works - a significant gap for rights holders, and a reminder that AI-IP questions are as much about jurisdiction as about copyright doctrine.
Drafting for AI output
Traditional IP clauses assume human creation, so standard assignment wording - all IP created by the contractor - may not capture AI output that was not legally created by the contractor. The fix is a dedicated AI provision that allocates ownership, takes a present assignment of any rights that do subsist, restricts the use of AI tools (no feeding the client's confidential information into a model; no third-party rights over outputs; no infringement), and requires disclosure of the tools and the human involvement.
The critical addition is a belt-and-suspenders clause: the parties acknowledge that AI outputs may not attract IP protection in some or all jurisdictions, and the contractual obligations - confidentiality, exclusivity, and restrictive covenants - apply whether or not IP rights subsist. Where the output has no copyright (the likely position in the US and EU, and possibly the UK if section 9(3) is repealed), that contractual allocation is the only protection you have.
Use at the desk
Practical checklist
- Treat AI output as a separate category - standard assignment wording may not capture it.
- Identify a human author or inventor whenever AI contributes (Thaler [2023] UKSC 49).
- Do not assume AI output has copyright - the US and most EU give none without human authorship.
- Note s.9(3) CDPA (computer-generated works) is UK-specific and under review.
- Add a belt-and-suspenders clause so confidentiality, exclusivity, and covenants bind whether or not IP rights subsist.
- Restrict feeding confidential information into AI tools and require disclosure of tools and human involvement.
This guide is informational only and is not legal advice. It does not replace advice from licensed counsel on the facts of a specific transaction.
Product demo
Use the guide for context. Use Veqtor for the Word documents.
Watch Claude compare negotiation drafts and create a separate Word document with proposed tracked changes.
See Veqtor work with Word redlines