New Nightshade Tool Allows Artists to Sabotage AI Training Data
A new open-source tool called Nightshade lets artists secretly alter their artwork to poison generative AI models that scrape their images without permission. These invisible pixel changes cause trained AI systems to produce chaotic and incorrect outputs, such as turning dogs into cats.
A new tool called Nightshade gives artists a way to fight back against generative AI companies that use their work without permission. The software allows creators to add invisible changes to the pixels of their art before uploading it online, which secretly corrupts the data if it is scraped for AI training.
When poisoned images are integrated into AI training sets, the resulting models break in chaotic and unpredictable ways. For example, the AI system might start generating cats when prompted for dogs, or cows when asked for cars, rendering features in tools like DALL-E, Midjourney, and Stable Diffusion essentially useless.
Ben Zhao, the University of Chicago professor who led the project, states that Nightshade aims to shift the power balance back toward artists by creating a strong deterrent against unauthorized data scraping. The team plans to integrate this data-poisoning feature into their existing style-masking tool, Glaze, and is making Nightshade open source so that its collective defensive power grows as more people use it.