Getty Images v Stability AI: Another Legal Win for AI Against Copyright Holders

On 4 November 2025, the High Court in London delivered judgment in Getty Images (US) Inc & Ors v Stability AI Ltd.

Summary

The dispute centred on Getty Images’ claims that Stability AI’s generative model, Stable Diffusion, had unlawfully used its copyrighted material and trade marks.

The judgment delivered mixed outcomes:

• Getty secured limited trade mark relief linked to older model versions,
• but its secondary copyright infringement claims were rejected on the ground that Stability’s AI model did not actually store or reproduce copies of the copyrighted works alleged to be infringed, and accordingly there was no secondary infringement under the relevant legislation (the Copyright, Designs and Patents Act, 1988 or CDPA).

Background and Parties

• Getty Images is a globally recognized provider of licensed visual content, offering millions of high-quality images and illustrations to diverse clients across media and business sectors.

• Stability AI Limited, headquartered in the UK, developed Stable Diffusion, an open-source latent diffusion model designed to generate images from textual prompts.

The model is trained on large datasets, including images sourced from platforms such as Getty Images and iStock, processed to capture contextual and visual attributes.

The technical architecture relies on neural networks that iteratively refine image outputs, though the final model does not retain the original training files.

With the rapid evolution of artificial intelligence and its intersection with intellectual property, this case has become a major reference point for every AI Law Firm advising clients on data training, copyright compliance, and ethical deployment of generative models. 

Procedural History

• Legal proceedings began in January 2023 with a broad range of complaints, including primary and secondary copyright infringement, database rights, trade mark violations, and passing off.
• Through case management orders issued across 2023 and 2024, the trial’s scope was significantly narrowed.
  • The focus shifted to specific versions of the Stable Diffusion model (v1.x, v2.x, SD XL) and sampled works relevant to both the copyright and trade mark claims.
• Key considerations included practical issues of licensing and contributor agreements, as well as technical evidence concerning watermark generation and the underlying operation of the AI model.
• Over the course of the actual trial hearings in June 2025, Getty was forced to acknowledge that there was no evidence that the training and development of Stable Diffusion took place in the United Kingdom, and also that Stability had already blocked the prompts that allegedly gave rise to the infringing outputs thereby already achieving the relief that Getty would’ve been entitled to.
  • As a consequence Getty had to abandon their primary copyright and database rights infringement claims.
• Ultimately, the court focused on two core issues:
  • secondary copyright infringement relating to the importation and distribution of model weights in the UK, and
  • trade mark infringement arising from AI-generated outputs that allegedly reproduced Getty’s marks in an actionable way. 

Trade Mark Infringement Claims

• Getty Images contended that Stable Diffusion was capable of generating images with marks resembling its famous “Getty Images” and “iStock” watermarks.
  • Specifically, Getty argued that such outputs mislead users and diluted the distinctiveness of its trademarks, contrary to sections 10(1), 10(2), and 10(3) of the Trade Marks Act 1994.
  • The claimant presented experiments and evidence of generated outputs containing these marks and emphasized their significance for brand identity.​
• Stability AI’s defence challenged the prevalence and authenticity of the watermark reproductions, characterizing them as rare artefacts of targeted prompts rather than normal system operation.
  • Stability also pointed to the implementation of technical safeguards following earlier model iterations, designed to filter out similar outputs and prevent recurrence.

Secondary Copyright Infringement

• Getty Images’ secondary copyright case argued that the model itself, via its trained parameters (“weights”), amounted to an “infringing copy” under section 27(3) of the CDPA.
  • The company also invoked sections 22 and 23 for actions related to the UK import, possession, and sale of the model deemed to contain infringing material.
• Stability AI countered by arguing that model weights are not themselves infringing copies under the CDPA, as they do not store or reproduce the original protected images.
• Further, Stability argued that copyright liability under UK law is territorially limited, and the model training had occurred predominantly outside the UK. 

The Judgment

1. Trade Mark Findings 

• • The court evaluated whether outputs containing Getty’s marks appeared in normal use of Stable Diffusion and distinguished between different model versions and access methods.
  • The court found that while earlier versions (notably v1.x and some v2.x) could generate outputs with Getty or iStock marks, such findings related mainly to controlled testing environments and not the general use case.
  • Trade mark infringement was established only in limited instances where outputs would likely cause confusion among professional users (especially with iStock marks), and only for certain versions and distribution channels.​
  • Claims under section 10(3), which required proof of unfair advantage or detriment to distinctive character, failed due to insufficient evidence of economic or reputational harm.
  • The court also declined to grant additional damages, noting a lack of continuing or widespread infringement. 

2. Secondary Copyright Infringement 

• • The court held that Stable Diffusion’s model weights do not qualify as “infringing copies” capable of engaging secondary liability under UK law.
  • The weight files, were found neither to embody reproductions of Getty’s works nor to operate as stored copies within the meaning of the legislation. Additionally, the territorial nature of UK copyright further insulated Stability AI from liability for any training conducted outside the country.​
  • While Getty established copyright subsistence and ownership for several sampled works and demonstrated that some contributor licences met the exclusive threshold, this had no bearing on the ultimate outcome due to the statutory construction of secondary infringement.

Summary

In summary, the High Court concluded that Getty Images achieved only a narrow and largely historic victory on trade mark infringement, limited to certain legacy uses and specific marks, while all claims of secondary copyright infringement failed, reflecting the stringent statutory requirements and the limits of traditional copyright in the context of artificial intelligence. 

Conclusion

The judgment significantly narrows potential legal exposure for AI developers who distribute models trained outside the UK and employ effective technical controls.

Creators and rights holders, meanwhile, must contend with a higher bar for both copyright and trade mark enforcement, with practical remedies likely to be limited to specific and well-evidenced outputs. Future claimants may need to marshal detailed empirical evidence of actual market confusion and economic impact, not just laboratory simulations.​

The UK High Court’s approach contrasts sharply with the more expansive treatment of similar issues in US courts. Parallel US proceedings against Stability AI are ongoing, with Getty seeking to bring findings of fact from the UK judgment before the Northern District of California.

Notably, US courts may weigh evidence of dataset composition, acts of copying, and retention practices differently, potentially leading to divergent legal outcomes. The Bartz v. Anthropic litigation is illustrative in this regard: there, the existence or retention of training data drove settlement discussions and broader legal scrutiny.

As these legal frameworks evolve, experienced technology law firm and lawyers are likely to play a crucial role in interpreting judicial boundaries, helping innovators and creators navigate the growing complexities of AI regulation and intellectual property protection.