Getty Images sues Stability AI over 12M scraped photos

The legal reality of generative AI just arrived with a vengeance.

Visual media giant Getty Images has officially filed a massive copyright and trademark lawsuit against Stability AI in the US District Court for the District of Delaware. The complaint, filed on February 3 and reported on February 6, comes hot on the heels of similar litigation Getty launched against the startup in the UK last month.

This time, the scale of the accusation is staggering. Getty alleges that Stability AI scraped more than 12 million copyrighted photos without permission or a dime of compensation. Even worse, it claims Stability AI took the associated metadata and captions to train its viral text-to-image generator, Stable Diffusion.

The smoking gun is a mangled watermark

While AI developers have long hidden behind the vague defense of fair use, Getty brings some very uncomfortable receipts to court.

According to the complaint, Stability AI's meteoric success was built directly on infringing Getty's library on a massive scale. Getty claims the model benefited immensely from the detailed accuracy of its image-text pairings. In some cases, the output generated by Stable Diffusion produces images that are highly similar derivatives of Getty's catalog.

Then there is the issue of branding. Stable Diffusion outputs often retain modified, distorted versions of the iconic Getty Images watermark, or churn out lower-quality art with the logo still visible. Getty argues this clearly violates trademark law and causes trademark dilution.

To add financial insult to injury, Getty pointed out that it already runs a legitimate licensing business for AI and machine learning developers. By scraping the assets for free, Stability AI essentially undercut Getty's own market and deprived creators of fair compensation.

Getty is not just asking for cash. Alongside treble damages and statutory penalties, it is demanding a permanent injunction and the outright destruction of all versions of Stable Diffusion trained on Getty's imagery.

Stability AI did not immediately respond to requests for comment, and Getty declined to speak beyond the filing.

Why it matters

This lawsuit cuts straight to the central vulnerability of the entire generative AI boom: where did the training data actually come from?

Getty’s legal strategy is sharp. Proving copyright infringement over latent diffusion features is tricky, but showing a jury a mangled Getty watermark spat out by the AI is visceral proof that the model copied its source material. It makes the technical argument around abstraction much harder for Stability AI to win.

If the court grants Getty's demand to destroy the infringing models, the ripple effects will be catastrophic for the open-source community. You cannot easily recall an open-weights model once millions of people have downloaded it. And if Stability AI has to scrub every contaminated image and retrain from scratch, both the cost and the quality hit could be fatal.

We think the era of treating the public web as an all-you-can-eat training buffet is officially closing. Going forward, AI companies will either have to pay upfront licensing fees or risk having their flagship models legislated into digital ash.