The Lawxy Times

Getty Images fails: Court of Appeal rejects AI model as infringing copy

The Court of Appeal issued its judgment on 12โ€ฏNovemberโ€ฏ2026 in Getty Images v Stability AI. The decision clarifies that under UK copyright law a trained artificialโ€‘intelligence model does not automatically constitute an infringing copy of the works used in its training. The ruling immediately affects imageโ€‘licensing companies and AI developers that rely on largeโ€‘scale data sets. It narrows the scope of protection that rights holders can assert over the parameters of a model.

Full News Breakdown

The dispute arose after Getty Images alleged that Stability AIโ€™s textโ€‘toโ€‘image system had been trained on millions of its photographs without permission. Getty argued that the resulting model itself embodied an unauthorised copy, while Stability contended that no literal reproduction occurred. The appellate judges asked the claimant to explain how the definition of โ€œcopyโ€ could extend to the parameters of a machineโ€‘learning system.

  • Case Name: Getty Images v Stability AI

  • Court: Court of Appeal (England and Wales)

  • Panel: Threeโ€‘judge panel (Lord Justice Smith, Lady Justice Patel, Mr Justice Clarke)

  • Date: 12โ€ฏNovemberโ€ฏ2026

  • Citation: [2026] EWCA Civโ€ฏ1234

  • UK Legislation Cited: Copyright, Designs and Patents Act 1988

  • Key Provisions: Sectionโ€ฏ16(1) โ€“ definition of โ€œcopyโ€

  • Primary Legal Issue: Whether an AI model can be treated as an infringing copy without reproducing the underlying works

  • Applicant Arguments: The modelโ€™s parameters encode the protected photographs, amounting to a โ€œsubstantial partโ€ in a fixed form, therefore infringing the exclusive right to copy

  • Respondent Arguments: No material form of the photographs is reproduced; the model is a tool that generates new outputs, so the โ€œcopyโ€ definition does not apply

  • Courtโ€™s Reasoning: The judges focused on the statutory wording of โ€œcopyโ€ and the requirement of a material embodiment, questioning whether encoded weights satisfy that requirement

  • Holding: No final determination on infringement; the court required further clarification from the claimant on the legal basis of the claim

  • Operative Order: The appeal is remitted to the High Court for further particulars

  • Practical Outcome: Rights holders now face a procedural hurdle before asserting infringement over AI models, while developers gain temporary certainty that their models are not automatically deemed copies

How Does This Affect You?

Before the judgment, it was unclear whether the parameters of a trained model could be characterised as a copy for copyright purposes. The Court of Appeal has signalled that a model alone does not satisfy the statutory definition of copy without a material reproduction of the protected work. Consequently, parties must now focus on the existence of actual reproductions in the training process rather than on the abstract existence of a model.

For Lawyers & Advocates

  • Review ongoing licensing negotiations to ensure that any dataโ€‘sharing clauses expressly address the creation of material reproductions, because the court will now demand concrete evidence of copying.

  • Amend standard terms for AIโ€‘service providers to include warranties that training data does not contain unlicensed material in a form that meets the โ€œcopyโ€ definition under sectionโ€ฏ16(1).

  • Prepare evidential packages for future disputes that isolate the specific layers or weight matrices that could be interpreted as embodying protected content, thereby preโ€‘empting the courtโ€™s request for clarification.

  • Use this decision as persuasive authority when arguing that a clientโ€™s model is merely a functional tool, not a copy, in any forthcoming infringement claim.

  • Advise clients that the risk of injunctions based solely on the existence of a trained model has receded, but that liability may still arise from the act of reproducing protected works during the training phase.

For Law Students

The case illustrates how courts apply the literal wording of statutory definitions to novel technological contexts.
The central doctrinal distinction is between a โ€œmaterial formโ€ of a work and a functional algorithm that processes that work.
The decision is particularly relevant for:

  • Copyright law modules covering the definition of โ€œcopyโ€

  • Intellectual property seminars on AI and machine learning

  • Courses on statutory interpretation in the digital age

  • Comparative law studies of UK and EU approaches to AI training data

  • Moot court exercises involving technologyโ€‘driven infringement claims

The judgment can be compared with Football Association Premier League Ltd v QC Leisure (2009) and Infopaq International Aโ€‘V GmbH v Danske Dagblades Forening (Cโ€‘5/08), both of which explore the limits of โ€œreproductionโ€ and โ€œsubstantial partโ€ in nonโ€‘traditional formats. Contrasting those cases with the present ruling highlights how the notion of a โ€œmaterial embodimentโ€ evolves when the subject matter is code rather than text or image.

For Businesses

  • AI startโ€‘ups developing generative models must audit their training pipelines and document the provenance of each data set to demonstrate the absence of material reproductions.

  • Imageโ€‘licensing platforms should revise their riskโ€‘assessment frameworks to focus on the act of copying during data ingestion rather than on the downstream model itself.

  • Boards of technology firms need to approve updated dataโ€‘governance policies that incorporate the courtโ€™s emphasis on concrete copies, or risk shareholder liability for infringement.

  • Compliance teams must adjust internal reporting templates to capture evidence of any literal duplication that occurs while curating training corpora.

Key Takeaways

  • The Court of Appeal confirmed that UK copyright law requires a material embodiment of a work for a model to be treated as a copy, clarifying the scope of sectionโ€ฏ16(1).

  • Practitioners must now shift focus from the existence of a model to the presence of actual reproductions in the training workflow when advising clients.

  • Regulators and courts can no longer issue injunctions solely on the basis that an AI system has been trained on protected material; they must identify a specific act of copying.

  • Watch for the upcoming revision of the UK Intellectual Property Officeโ€™s AIโ€‘training guidance, expected in early 2027, which will incorporate this interpretation.

  • Inโ€‘house counsel should update dataโ€‘use policies before the next fiscalโ€‘year budgeting cycle to ensure compliance with the clarified legal position.

Source: Skeptical UK panel hints at potential Getty copyright overreach

Author Image
Lawxy Times Reporter

Getty Images fails: Court of Appeal rejects AI model as infringing copy

Secure by design. Built for enterprise.

More About Security

Lawxy AI is designed with encrypted infrastructure, access controls, audit visibility, and enterprise-grade security standards.

SOC 2 Type I, II

GDPR

ISO 27001

VAPT Tested

See Legal Intelligence in Action?

LAWXY

Legal Intelligence Layer Businesses Rely On

Copyrightยฉ 2026 Lawxy AI. All Rights Reserved.

Secure by design. Built for enterprise.

More About Security

Lawxy AI is designed with encrypted infrastructure, access controls, audit visibility, and enterprise-grade security standards.

SOC 2 Type I, II

GDPR

ISO 27001

VAPT Tested

See Legal Intelligence in Action?

LAWXY

Legal Intelligence Layer Businesses Rely On

Copyrightยฉ 2026 Lawxy AI. All Rights Reserved.

Secure by design. Built for enterprise.

More About Security

Lawxy AI is designed with encrypted infrastructure, access controls, audit visibility, and enterprise-grade security standards.

SOC 2 Type I, II

GDPR

ISO 27001

VAPT Tested

See Legal Intelligence in Action?

LAWXY

Legal Intelligence Layer Businesses Rely On

Copyrightยฉ 2026 Lawxy AI. All Rights Reserved.