Stable Diffusion Generated Images Containing Getty Images Watermarks Expose Copyright Training Data

In August 2022, shortly after Stability AI released Stable Diffusion 1.4 to the public, users began experimenting with the open-source text-to-image generation model. Within days of release, multiple users on social media platforms including Twitter and Reddit discovered that the model could generate images containing distorted versions of the distinctive Getty Images watermark. These watermarks appeared as corrupted text overlays that resembled Getty's copyright protection markings, though often garbled or partially formed. The watermark discovery provided smoking-gun evidence that Stable Diffusion had been trained on copyrighted Getty Images content without authorization. Stability AI had trained their model using the LAION-5B dataset, a massive collection of 5.8 billion image-text pairs scraped from the internet. This dataset included millions of Getty Images photographs that contained visible watermarks, which the model learned to associate with certain visual patterns and occasionally reproduced in generated outputs. The technical explanation for this phenomenon lies in how diffusion models learn visual patterns during training. When exposed to thousands of Getty watermarked images across diverse subjects and compositions, the model internalized the watermark as a visual feature that could appear in generated content. The corrupted appearance of these watermarks in generated images reflected the model's imperfect understanding of the watermark as a visual element rather than a copyright protection mechanism. This discovery became central evidence in Getty Images' February 2023 lawsuit against Stability AI in the US District Court for the District of Delaware. Getty alleged that Stability AI had unlawfully copied and processed millions of copyrighted images to train Stable Diffusion, with the watermark generation serving as direct proof of this unauthorized use. The lawsuit seeks monetary damages and injunctive relief to prevent further copyright infringement. The incident sparked broader debate about training data transparency and copyright compliance in AI development. Legal experts noted that the visible watermarks made this case particularly strong for Getty, as it provided clear evidence of copying rather than requiring complex analysis of model outputs. The case has implications for the entire AI industry's approach to training data sourcing and copyright compliance.