Nightshade: The AI Poisoning Tool Empowering Artists Against Copyright Theft

Artists now have a new weapon in their fight against generative AI image software that uses their works for training without permission. Nightshade, developed by University of Chicago professor Ben Zhao and his team, allows creators to add invisible pixel modifications to digital artworks, corrupting how AI models interpret them. For example, an image of a table could be read as a cat, disrupting the model's ability to generate coherent images. According to MIT Technology Review, Zhao hopes this will "overturn the current power balance, taking it away from AI companies that have taken copyrighted data to train their models." The tool targets text-to-image models like DALL-E, Stable Diffusion, and Midjourney, potentially destabilizing their general features with a moderate number of attacks, as stated in the research paper cited by The Verge. Nightshade will soon become open source. A companion tool, Glaze, masks artistic style (e.g., turning an illustration into a 3D photo or a realistic work into cubist). Developers may integrate Nightshade into Glaze, giving users the choice between style obfuscation or outright data poisoning, pending a better solution like royalty payments.