AI-Generated Image Copyright in the UK: 2026 Guide

You prompt a foundation model for a marketing campaign visual, download the high-resolution file, and place it directly on a client billboard.

Two weeks later, the client asks a straightforward question: who actually owns the image? If a competitor copies the artwork pixel for pixel, can your agency sue them for copyright infringement under UK law?

If you answer yes based on US copyright headlines or generic software disclaimers, you are taking an unforced legal risk.

In the United States, the legal position is settled: pure machine outputs lack human authorship and cannot hold copyright, a position reinforced when the US Supreme Court denied certiorari in Thaler v Perlmutter on 2 March 2026. But the United Kingdom is one of only a handful of jurisdictions worldwide with a specific statutory framework for computer-generated works.

Yet holding a statutory hook does not make your prompt-to-image pipeline automatically copyrightable. Under English law, statutory authorship, subsistence, and commercial use rights are three entirely separate concepts.

Here is how UK AI copyright law actually applies to commercial image generation in 2026:

1. The UK statutory framework: Section 9(3) and the 50-year computer-generated works rule.
2. The originality trap: why THJ Systems v Sheridan complicates subsistence.
3. Landmark litigation: what Getty Images v Stability AI decided (and what it left open).
4. Policy shifts: the Government's 18 March 2026 Report and the proposed removal of s.9(3).
5. Contract vs copyright: what platform indemnities actually cover.
6. The commercial creator decision tree for UK businesses.
7. Frequently asked questions.

The UK statutory framework: Section 9(3) explained

The foundation of UK copyright is the Copyright, Designs and Patents Act 1988 (CDPA 1988). While most international copyright regimes require a human author to exist before copyright can subsist, the UK Parliament introduced a legal fiction in 1988 to deal with software-generated material.

Under Section 9(3) of the CDPA 1988:

"In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken."

Section 178 clarifies that "computer-generated" applies where a work is generated by computer "in circumstances such that there is no human author of the work." Under Section 12(7), computer-generated works receive a copyright term of 50 years from the end of the calendar year in which the work was made (compared to 70 years after the author's death for standard authorial works). Crucially, under Sections 79 and 81, computer-generated works carry no moral rights: there is no statutory right of attribution or right to object to derogatory treatment.

So who is the statutory author? When you type a descriptive prompt into an image generator, are you the person making "the arrangements necessary for the creation of the work," or is it the software developer who trained the neural network?

While the leading historical authority on arrangements (Nova Productions v Mazooma Games [2007]) examined arcade video games rather than generative neural networks, government policy documents offer an indication. In its 18 March 2026 Report on Copyright and Artificial Intelligence, the UK Government noted that for prompt-driven systems, the person making the necessary arrangements is generally the end-user entering the prompt.

Section 9(3) remains on the statute books, but statutory authorship does not guarantee copyright subsistence.

The originality trap: why authorship does not guarantee copyright

Allocating statutory authorship under Section 9(3) is only half the equation. For copyright to protect an artistic work in the UK, the work must also be original.

Historically, English courts applied a low bar for originality: the "skill, labour, and judgment" test. If you expended sufficient effort, the output cleared the hurdle.

That changed definitively with THJ Systems Ltd v Sheridan [2023] EWCA Civ 1354. In November 2023, Lord Justice Arnold reaffirmed that post-Brexit UK copyright law follows the European Court of Justice standard: an original work must be the "author's own intellectual creation." This requires the author to exercise free and creative choices that stamp their personal touch on the output.

This creates a genuine legal tension that English courts have not yet resolved:

• The subsistence problem: Section 178 defines a computer-generated work as having no human author. Yet THJ Systems v Sheridan dictates that originality requires a human intellectual creation.
• The two competing interpretations: One view argues Section 9(3) was designed by Parliament as an intentional statutory exception to traditional originality rules. The second view argues Section 9(3) merely designates who would own the copyright if the work is original, but because pure generative AI outputs are generated stochastically without direct human manual execution, raw prompt outputs fail the originality test and have no subsisting copyright at all.

Until an appellate court rules on a direct dispute involving synthetic images, you should assume that a raw, prompt-only output holds weak or non-existent copyright subsistence in the UK.

Landmark litigation: what Getty v Stability AI settled

Commercial risk in the UK is shaped by case law as much as statute. The most significant UK proceeding to date is Getty Images (US) Inc & ors v Stability AI Ltd [2025] EWHC 2863 (Ch), decided by Mrs Justice Joanna Smith in the High Court on 4 November 2025.

Getty initially brought claims covering both primary copyright infringement (training models on copyright-protected photography) and secondary infringement (importing and distributing the resulting model weights into the UK under Sections 22 and 23 of the CDPA 1988).

Mid-trial, Getty abandoned its primary UK copyright training claims due to jurisdictional difficulties, as Stability's core computational training took place outside the UK. That left the court to decide the secondary infringement and trade mark claims. The rulings established several key principles for commercial users:

• Model weights are intangible articles: The High Court confirmed that neural network model weights qualify as "articles" under Section 22 and 23 of the CDPA, even when distributed purely as digital cloud assets.
• Model weights are not "infringing copies": The court dismissed the secondary copyright claim under Section 27. Mrs Justice Joanna Smith held that model weights do not store, contain, or reproduce pixel data from training images. They represent statistical parameters. Because the model weights themselves are not infringing copies, hosting or distributing the model did not constitute secondary infringement.
• Trade mark infringement remains a live risk: Getty established historic trade mark infringement where earlier versions of the model generated synthetic images displaying distorted versions of Getty and iStock watermarks.

While the High Court granted Getty permission to appeal on the secondary infringement point in December 2025 (Re Form of Order [2025] EWHC 3343 (Ch)), the commercial takeaway for businesses is concrete: training legality inside the UK remains open, but distributing trained models does not automatically constitute secondary copyright infringement.

However, generating recognizable third-party logos or watermarks exposes your business directly to trade mark claims under Section 10 of the Trade Marks Act 1994, entirely separate from copyright.

Policy shifts: the March 2026 Government Report

UK policy on AI copyright is moving rapidly. Following the Data (Use and Access) Act 2025, which received Royal Assent in June 2025, the Government was mandated under Section 136 to deliver an economic impact assessment and policy report by 18 March 2026.

The Government's consultation received 11,520 responses. Data released in the December 2025 progress statement showed that 88% of online respondents demanded a strict licensing requirement for AI training data, while only 3% supported the Government's previously preferred broad Text and Data Mining (TDM) opt-out exception.

On 18 March 2026, the Department for Science, Innovation and Technology (DSIT) alongside the UK IPO published the final report. The key policy conclusions include:

• The broad TDM opt-out exception is abandoned: The Government confirmed that an open exception with a rights-holder opt-out is no longer its preferred approach, leaving licensing agreements to private market frameworks.
• Proposal to remove Section 9(3) for raw AI works: The Government proposed removing statutory copyright protection for wholly computer-generated works under Section 9(3) due to a lack of evidence of its practical economic utility.
• Clear protection for AI-assisted works: The Government clarified that copyright will explicitly remain intact for "AI-assisted" works, where human creators use AI engines as tools alongside substantial human creative contribution.

Because comprehensive legislation is unlikely to reach Parliament before 2027, Section 9(3) remains on the statute book in full effect today, but its long-term future is numbered.

Routing models across a multi-engine hub lets creators match commercial licensing to the exact creative job.

Contract vs copyright: what platform terms actually provide

If copyright subsistence in raw outputs is doubtful, where does your business get the legal right to use AI images commercially?

The answer is contract law. Your right to commercialize an image does not stem from claiming exclusive intellectual property ownership against the world: it stems from the commercial licence granted by the model provider.

When choosing a generator, compare vendor licensing and IP indemnity frameworks across paid and enterprise tiers:

Every commercial indemnity contains strict carve-outs. If your designer intentionally prompts for a known character ("in the style of Disney's Mickey Mouse") or modifies the output with external tools, vendor protections vanish.

In multi-model workspaces like JammyJar, teams route prompts across different engines—using Recraft for editable vector exports, GPT Image for typography, or Gemini for editorial raster graphics. When choosing a workflow, our breakdown of commercial AI image usage rights and licensing terms outlines the exact differences between free and paid tier ownership clauses.

The commercial creator decision tree

Before deploying an AI-generated image into commercial client deliverables, run your asset through this six-stage risk assessment tree.

[1. Base Model Check]
│
├─ Free tier / Personal only? ──> STOP: Upgrade to paid tier with commercial licence.
│
└─ Paid commercial licence confirmed
│
[2. Prompt & Visual IP Screen]
│
├─ Prompts name living artists, active brands, or copyrighted characters? ──> STOP: Re-prompt.
├─ Output displays accidental third-party watermarks or logos? ──> STOP: Retouch or discard.
│
└─ Clean synthetic output
│
[3. Personality & Privacy Check]
│
├─ Depicts identifiable real living people without signed model release? ──> STOP: Do not publish.
│
└─ Synthetic or licensed subjects
│
[4. Cross-Border & Regulatory Compliance]
│
├─ Deploying in EU markets? ──> Apply machine-readable C2PA / metadata marking (EU AI Act Art. 50).
│
└─ UK / Domestic deployment
│
[5. Human Authorship & Subsistence Strategy]
│
├─ Raw 1-click generation? ──> Rely on contract licence. Assume no enforceable exclusivity.
│
└─ Iterative composite, layered vector edits, manual typography, or multi-asset collage?
│
▼
[6. Commercial Deployment Ready]
│
└─ Defensible copyright claims in human-authored contribution; client warranties aligned.

How to implement each stage

• 1. Verify the vendor licence: Ensure the generation occurred under an active paid account. For tools like Recraft, free-tier assets remain public property owned by the vendor without commercial rights. On platforms like JammyJar, all paid tiers (starting from the Jampot plan at $5/month) grant full commercial use.
• 2. Screen for trade marks and visual IP: Under Getty Images v Stability AI, distorted trade marks and watermarks create immediate liability under Section 10 of the Trade Marks Act 1994. Inspect image borders, background patterns, and rendered badges for accidental brand marks.
• 3. Check for privacy and personality rights: UK common law does not recognize a standalone "right of publicity", but using a synthetic likeness of a real individual in advertising risks claims under false endorsement (passing off) and UK GDPR rules on biometric and personal data.
• 4. Check EU cross-border obligations: If your campaign runs in the European Union, Article 50 of the EU AI Act applies from 2 August 2026 (with a grace period to 2 December 2026 for legacy systems). Synthetic images must carry machine-readable markers (such as C2PA metadata or DeepMind SynthID). If you publish vectors, our tutorial on how to generate and inspect real editable SVG vectors explains how to verify path structures without stripping embedded metadata.
• 5. Establish human authorship: If you need defensible exclusivity over a brand asset, do not rely on raw output. Layer prompt generations into custom Figma compositions, combine synthetic textures with manual vector linework, or composite multiple rendered elements together. Document your iteration files: this establishes the "author's own intellectual creation" standard required by THJ Systems v Sheridan.
• 6. Align client contracts: Never sign an agency contract promising to "assign full copyright in all deliverables" if the core asset is a raw AI output. Instead, warrant that you hold the commercial right to licence the output, assign copyright in all human-authored modifications, and provide a non-infringement warranty based on your clean prompt workflows.

Frequently asked questions

Does UK copyright law protect AI-generated images? Section 9(3) of the CDPA 1988 names the person making the necessary arrangements as the author of a computer-generated work. However, following the Court of Appeal ruling in THJ Systems v Sheridan [2023], artistic works must also meet the originality threshold of being an author's own intellectual creation. Pure, unedited AI generations likely lack this human creative touch and may hold no subsisting copyright.

Can UK businesses use AI images commercially without owning copyright? Yes. Commercial use rights come from the contract licence granted by the model provider's terms of service, not from owning underlying copyright. Provided you generate on a paid tier granting commercial exploitation rights and avoid infringing third-party trade marks, commercial use is lawful across the UK.

What did the High Court decide in Getty Images v Stability AI? In November 2025, the UK High Court ruled that AI model weights are not "infringing copies" under Section 27 of the CDPA 1988, dismissing Getty's secondary copyright infringement claim. However, the court found historic trade mark infringement where synthetic outputs reproduced recognizable Getty and iStock watermarks.

How does UK AI copyright law differ from US law? US law strictly requires human authorship for any copyright protection, confirmed when the US Supreme Court denied review in Thaler v Perlmutter in March 2026. The UK has a specific statutory provision for computer-generated works under Section 9(3) giving 50 years of protection, though the Government proposed removing this provision in its March 2026 report.

How can designers secure copyright in AI-assisted workflows? To secure copyright under UK law, designers must provide substantial human creative choices. This includes manual composition, post-processing, combining multiple elements, colour grading, or vector modification. Documenting intermediate project files proves the human intellectual creation required by English courts.

So before shipping your next commercial campaign, stop asking whether an AI prompt makes you an author. Treat raw model outputs as licensed ingredients, verify your vendor terms, layer your human creative touch across the canvas, and build your client contracts on real commercial licences rather than presumed copyright.