Getty Images vs. Stability AI

01. Getty Images v. Stability AI: the ruling that didn't change the game

On 4 November, the English Court dismissed most of the copyright claims brought by Getty Images against Stability AI, developer of the "Stable Diffusion" generative AI model.

• Only trademark infringement was found, and only on older versions of the model.

02. Getty's original claim

In January 2023, Getty sued in the High Court of Justice alleging that Stability AI:

• Used its database and associated text to train the Stable Diffusion model with millions of photographs without its permission.
• Generated images similar to its content, some even reproducing "Gettyimages" and "iStock" watermarks.

03. What was actually discussed

After withdrawing several main claims for lack of jurisdiction (as it could not prove that the infringing acts occurred in the UK), the Court focused the case on two key issues:

• Do the outputs infringe Getty's trademarks?
• Does Stable Diffusion constitute an "infringing copy" of works protected under UK copyright law?

04. Copyright: the wall Getty fails to break down

The Court concludes that there is no copyright infringement, because:

• It was not proven that the training of the AI model took place in the UK, a requirement for the application of the UK Copyright, Designs and Patents Act ("CDPA").
• Stability implemented locks on prompts to avoid generating images similar to particular works.

In addition, the Court rejects the "infringing copy" theory:

• The model parameters are learned mathematical patterns, not Getty photographs.
• Stable Diffusion does not store or reproduce protected works, so it does not constitute "infringing copying" under section 27(3) of the CDPA.

Conclusion: the model learns from the data, but does not reproduce it. Without reproduction, there is no "infringing copy".

05. Trademarks: single infringement... and limited in scope

The Court did find trade mark infringement, but of limited scope:

• Some older versions of the model generated images with "iStock" watermarks or similar blurred signs.
• The outputs of the model are "synthetic images", not photographs, so there is no "identity" between products.
• Nor was reputational damage or free-riding established.

06. Implications of the judgment

A ruling that does not close the debate, but redefines the limits of the conflict between AI and intellectual property.

• The judgment does not resolve the key question: does training AI models with protected works constitute an infringement?
• It reaffirms that a trained AI model is not a copy of the dataset, but a statistical result of learning.
• Could anticipate an increase in trademark litigation.
• Highlights the territorial limits of copyright vis-à-vis AI and the importance of where model training takes place.

Briefing note prepared by the TMT area of ECIJA Madrid.