Since their arrival, generative AI models and their trainers have demonstrated their ability to download any online content for model training. For content owners and creators, few tools can prevent their content from being fed into a generative AI model against their will. Opt-out lists have been disregarded by model trainers in the past, and can be easily ignored with zero consequences. They are unverifiable and unenforceable, and those who violate opt-out lists and do-not-scrape directives can not be identified with high confidence.
In an effort to address this power asymmetry, we have designed and implemented Nightshade, a tool that turns any image into a data sample that is unsuitable for model training. More precisely, Nightshade transforms images into "poison" samples, so that models training on them without consent will see their models learn unpredictable behaviors that deviate from expected norms, e.g. a prompt that asks for an image of a cow flying in space might instead get an image of a handbag floating in space.
