This new data poisoning tool lets artists fight back against generative AI

Researchers led by Ben Zhao, a University of Chicago professor, have released Nightshade, a tool that lets artists add invisible pixel changes to their images so that if they are scraped into an AI training set, the resulting model can break in unpredictable ways. The team says the tool is meant to deter AI companies from using artists' work without permission. According to the research, 50 poisoned images of dogs made Stable Diffusion generate odd, multi-limbed creatures, and 300 poisoned samples let an attacker make it produce dogs that look like cats. The team plans to integrate Nightshade into its Glaze tool and to make it open source.