The Great AI Art Heist

In a world of generative AI, artists like Kim Van Deun are finding themselves in an industry with dwindling opportunities. Enter Ben Zhao, a University of Chicago professor and, as Kelley Engelbrecht describes him, “a computer scientist who genuinely loves art.” In Zhao’s work at the university’s SAND Lab, this appreciation for art is loud and clear. His team has developed tools, such as Glaze and Nightshade, that shield artists’ work from AI models that scrape and train on images without consent. Engelbrecht’s Chicago Magazine story explores the importance of human creativity in a world where machines increasingly mimic art.

The success felt good. But it was soon evident that protecting artists at an individual level wouldn’t solve the greater problem of the nonconsensual scraping of images. While SAND Lab researchers had figured out how to confuse generative AI machines, they came to realize there was an opportunity to aim even higher by damaging the data sets used to train the models.

Nightshade, released in early 2024, took a more collective approach. If generative AI models were going to continue to train on images without consent, Nightshade would make sure those images would teach the machines unexpected and unpredictable behavior. If Glaze was built to be a defensive bulwark, Nightshade was designed to go on the attack.

Take an image of a cat. Apply Nightshade to the image, and the AI model will see not a cat but something entirely different — perhaps a chair. Do this to enough images of cats, and gradually the model stops seeing cats and sees only chairs. Ask the same model to generate a picture of a cat, and you get an overstuffed high-back chair instead, maybe even with scrolled wooden feet. While Glaze provides immediate protection on individual images, Nightshade, which has been downloaded more than 2 million times, plays the long game. It poisons the well one image at a time.