University of Chicago develops 'Nightshade' to prevent AI training on artists’ digital works

University of Chicago researchers have created a tool called 'Nightshade' that enables artists to 'poison' their digital artwork to prevent its use in training artificial intelligence (AI) systems. This has been done to stop developers from using digital artwork to train their own AI image-creation models. 

The tool alters images in a way that contaminates AI training datasets with incorrect information. This means the metadata and the code of the original image are altered in such a way, that an AI training model trying to read its data will receive incorrect information thus training its model on false data.

As per MIT's Technology Review report, Nightshade modifies the pixels of a digital image to fool the AI into misinterpreting it. This could potentially harm the AI's ability to produce accurate outputs. 

For instance, the AI might incorrectly identify a cat as a dog and vice versa. The research team behind Nightshade, led by Professor Ben Zhao, plans to integrate it into their existing artist protection software, Glaze, which is currently available free for web use or download. 

Vitaly Shmatikov, a professor at Cornell University, suggested that there are currently no known robust defences against these attacks, implying that even robust models like OpenAI's ChatGPT could be vulnerable.
 

Several publishing companies have raised concerns about their data being scraped off the internet by these training models and have started to poison their data to ensure that any AI training models scraping its data would get incorrect information. While this might not be good in the long run for free AI models, this could trigger a movement where publishing companies or artwork creators ask all models that are training on its data for a premium.