Getty Images (US) Inc & Ors v Stability AI Limited: Implications for AI and IP

Getty achieved limited success on trade mark grounds.  The judge found that excessive exposure to Getty and iStock watermarks during training  caused memorisation and subsequent reproduction in synthetic outputs (known as “overfitting”).

The judge found infringement under section 10(1) (double identity of registered mark and sign used by Stable Diffusion) and section 10(2) (likelihood of confusion of users) of the Trade Marks Act 1994 (TMA), but only for specific images presented as evidence.  The judge declined to make a broader ruling on this occurring more widely due to insufficient real-world evidence from later model versions and the variable appearance of watermarks in outputs (varying from identical reproduction to similarity that was insufficient for a finding of trade mark infringement).

Getty's section 10(3) claim (reputation and tarnishment) failed entirely. The judge required robust evidence of changes in economic behaviour and real-world harm, which Getty could not demonstrate. This sets a high evidential bar.

Key findings:

• Precise specifications and clear pleadings are crucial:  AI-generated content services did not exist when Getty's marks were registered. The judge declined to interpret terms in the specification broadly, instead examining whether new variants fall within existing specifications. Claimants must clearly plead why articles or services fall within registered specifications. In light of the Skykick case (see our previous article here), overly broad specifications of goods and services can be deemed to be indicative of bad faith and so cannot be used to ‘futureproof’ trade mark registrations.  This leads to a delicate balancing act.  Taking time to ensure the specification is well-drafted could avoid litigation further down the line.
• Stability was responsible for using the sign in the course of trade as it had control: The judge held that whilst users control prompts, they lack complete control over outputs, which depend on training data over which Stability exercised “absolute control and responsibility”.  Average users would not consider themselves responsible for watermark reproduction, even when using out-of-scope prompts like “news photo”.  This finding prevents AI providers from escaping liability by attributing outputs solely to user actions, a potentially significant precedent for content owners.
• Users believing Getty content was licensed to Stability to train the model would create the impression that there is a material link in the course of trade: This is sufficient for the likelihood of confusion required by section 10(2) of the TMA. 

Passing off: abandoned claim

Getty conceded that its passing off claim stood or fell with trade mark infringement, relying on identical evidence. Upon receiving the draft judgment, Getty declined further submissions, and the judge did not address passing off separately. This tactical decision raises questions about whether maintaining independent arguments for each cause of action might preserve claims even when related theories fail.

Licensing formalities

A significant subsidiary issue concerned whether Getty held exclusive licences satisfying section 92 CDPA, which requires licences to be “in writing signed by or on behalf of the copyright owner” and granted to a single licensee.  Several agreements failed this test due to multi-entity “Getty Images” definitions that defeated single-licensee requirements, meaning Getty did not have standing to bring claims in relation to them.

The judgment highlighted jurisdictional complexities: whilst New York law governed the interpretation and construction of some licences, English law determined whether they qualified as “exclusive” under the CDPA.  Some agreements seemed to suggest different positions on exclusivity in different clauses, something to be carefully avoided. 

Additionally, Getty lacked adequate evidence that certain licences were properly executed electronically (via tick boxes or DocuSign), underscoring the importance of keeping comprehensive execution records. The judge confirmed that electronic signatures can satisfy section 92 if properly evidenced and authenticated, but documentation proving execution is essential.

Practical Implications

• For AI developers: The judgment provides qualified reassurance that diffusion models that do not store training images avoid secondary copyright infringement but leaves primary infringement during training unaddressed.  Developers should avoid overfitting and implement robust filtering to remove trade marks, document non-storage of training data by model weights and models and maintain comprehensive safety systems. The effectiveness of Stability's later filtering efforts demonstrates that technical mitigation can significantly reduce, if not eliminate, the reproduction of trade marks and their infringement.  However, the judge noted users can remove safeguards in open-source software, highlighting ongoing challenges for open-source AI governance.
• For content owners: Copyright enforcement faces significant hurdles under current law, but trade mark protection offers a possible alternative where registered marks are reproduced in outputs.  Those wanting to enforce rights they are using should ensure either that they own the underlying IP rights (and can prove chain of title) or that any exclusive licences granted to them are in writing, signed and clearly name one exclusive licensee.  Comprehensive execution records should be maintained and any evidence of clear reproduction or consumer confusion in the UK should be kept and documented.
• For platforms: The distinction between third-party repositories (lower liability) and actively curated hosted services (higher liability based on control) suggests architectural choices affect legal exposure. Investment in filtering systems serves both risk mitigation and evidential purposes.
• For users: Whilst the judgment confirms that AI providers have “control” over outputs, users should not assume this will absolve them of all liability.  The case turned on its facts and even if a user’s use is not found to have infringed the IP rights of third parties, it may fall outside the terms of use, potentially creating contractual liability to the provider or host.

Conclusion

Getty Images v Stability AI provides important but incomplete guidance on AI liability under English IP law.  The rejection of secondary copyright infringement for non-storing models offers AI developers qualified comfort, whilst limited trade mark liability for watermark reproduction confirms that technical safeguards and comprehensive filtering remain essential. The case emphasises the importance of technical evidence, precise trade mark specifications, rigorous licensing formalities, and comprehensive documentation in AI-related IP disputes.  Whilst binding only in England and Wales, the judgment may prove persuasive in other common law jurisdictions facing similar AI and IP challenges, particularly those with comparable statutory frameworks.  At the time of writing, Getty has not indicated that it intends to appeal.

However, the judgment's narrow scope, resulting from Getty's abandoned claims, means fundamental questions about AI training remain. This underscores the need for legislative clarity on AI training and copyright.  The UK government is still considering the responses received to its consultation on AI and copyright some nine months after it closed, illustrating the complexity of the issues and the competing demands that need to be addressed.  We wait with anticipation to see what developments the next few months bring in this space.