Commentary On The Getty v. Stability Judgment

Getty v. Stability

New technologies are increasingly colliding with established legal principles, forcing courts and legislators to reconcile principles drafted for a different era with today's technological reality.

Perhaps nowhere is this tension clearer than in the intersection between Artificial Intelligence ("AI") and Intellectual Property ("IP"), where traditional concepts of ownership and originality are being redefined. The legal frameworks governing AI and IP remain nascent, and precedent is still catching up.

The case of Getty Images v. Stability AI marks a foundational step in that process, a landmark case attempting to draw the line between training and copying, and to define how far traditional IP law can stretch to accommodate a technology that "learns" ... but does not necessarily "remember".

Background

Generative AI systems are trained on vast datasets, often containing existing works, raising complex questions of legality. In essence, an AI's outputs are built on the inputs. This raises a simple yet profound question: when an AI learns from protected material... has it copied it?

It is this quintessential question that brought Getty Images into conflict against Stability AI Ltd, the developer of Stable Diffusion, alleging that the company had used Getty's extensive catalogue of photographs and captions to train its generative AI model without authorisation.

Getty's claim initially spanned copyright, database rights, and trademarks. Albeit at a later stage, Getty honed its claims towards two central questions:

1. Whether Stable Diffusion itself constituted an infringing copy of Getty's works; and
2. Whether any of the model's outputs infringed Getty's registered trademarks by reproducing or imitating the Getty Images and iStock watermarks.

Copyright: A Gap between 'Process' and 'Product'

Getty's copyright claim forced the UK's High Court of Justice (the "Court") to confront a new question under the UK Copyright, Designs and Patents Act 1988, namely could a trained AI model, derived from protected works, itself be an "infringing copy"?

Getty argued that by incorporating patterns and relationships derived from its images, the model embodied a reproduction of its works. Stability countered that Stable Diffusion did not store or reproduce any of the training images but merely encoded statistical relationships between pixels and text.

The Court accepted that intangible 'articles' can exist under UK law and that training involved reproductions of Getty's works yet held that Stable Diffusion itself was not an infringing copy.

The Court upheld the view that Stable Diffusion itself was not an infringing copy as its final 'model weights'¹ contained no stored or retrievable images. The model was the product of a process that involved copying but did not itself reproduce any works.

The Trademark Question: Real-World infringement and developer liability

Getty's remaining claim focused on the appearance of its watermarks in images generated by early versions of Stable Diffusion. Getty alleged that these outputs amounted to trademark infringement under the UK Trade Marks Act 1994.

The Court accepted that in the earliest releases some synthetic images displayed iStock watermarks. Stability attributed this to 'overfitting', where a model memorises artefacts from its training data. The Court found limited infringement confined to early versions. The more serious allegation, that Stability had taken unfair advantage or caused dilution failed as Getty could not prove any change in consumer behaviour or reputational harm.

Developer Liability

Stability argued that any infringing "use" of Getty's marks occurred at the user level, that it merely provided a neutral technical tool, comparable to a search engine or hosting platform. The Court firmly rejected this argument.

The Court held that Stability was engaged in "active behaviour" by training, distributing and ultimately maintaining their model. The End-User could not prevent watermarks from appearing, nor did they request or control their presence.

This finding is far from peripheral as it represents one of the first acknowledgements that AI developers ("Providers" within the EU's sphere) can incur direct liability for infringing outputs, where the act of "making available" or "maintaining" a model may itself constitute 'use'

Concluding Remarks

Getty achieved a narrow technical win but lost on principle: no copyright infringement, no ongoing trademark misuse, and no damages awarded.

If anything, this case highlighted the immense evidentiary difficulty of proving infringement in an AI context.

For practitioners across the EU, the decision highlights a critical divergence. Unlike the UK, EU law recognises text-and-data-mining ("TDM") exceptions under the DSM Directive, albeit whether AI training falls within those exceptions remains unresolved within the EU's legal landscape. Some Member States (notably Germany, see Robert Kneschke v. LAION e.V²) have provisionally accepted that dataset creation may constitute TDM, while the CJEU is expected to address the issue in pending references (read more here and here).

Article 53(1)(c) of the AI Act introduces opt-out mechanisms for rightsholders, but Getty v. Stability exposes a regulatory gap, whereas the primary remedies remain copyright and trademark law... both possibly ill-suited for the generative AI era.

Footnotes

1. The Court held that Model Weights are "purely the product of patterns and features learnt during the training process". It is the training process that utilised the images in question.

2. Not binding on other courts.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.