Thursday, November 14

The AI lab waging a guerrilla war over exploitative AI

Ben Zhao remembers well the moment he officially jumped into the fight between artists and generative AI: when one artist asked for AI bananas. 

A computer security researcher at the University of Chicago, Zhao had made a name for himself by building tools to protect images from facial recognition technology. It was this work that caught the attention of Kim Van Deun, a fantasy illustrator who invited him to a Zoom call in November 2022 hosted by the Concept Art Association, an advocacy organization for artists working in commercial media. 

On the call, artists shared details of how they had been hurt by the generative AI boom, which was then brand new. At that moment, AI was suddenly everywhere. The tech community was buzzing over image-generating AI models, such as Midjourney, Stable Diffusion, and OpenAI’s DALL-E 2, which could follow simple word prompts to depict fantasylands or whimsical chairs made of avocados. 

But these artists saw this technological wonder as a new kind of theft. They felt the models were effectively stealing and replacing their work. Some had found that their art had been scraped off the internet and used to train the models, while others had discovered that their own names had become prompts, causing their work to be drowned out online by AI knockoffs.

Zhao remembers being shocked by what he heard. “People are literally telling you they’re losing their livelihoods,” he told me one afternoon this spring, sitting in his Chicago living room. “That’s something that you just can’t ignore.” 

So on the Zoom, he made a proposal: What if, hypothetically, it was possible to build a mechanism that would help mask their art to interfere with AI scraping?

“I would love a tool that if someone wrote my name and made a prompt, like, garbage came out,” responded Karla Ortiz, a prominent digital artist. “Just, like, bananas or some weird stuff.” 

That was all the convincing Zhao needed—the moment he joined the cause.

Fast-forward to today, and millions of artists have deployed two tools born from that Zoom: Glaze and Nightshade, which were developed by Zhao and the University of Chicago’s SAND Lab (an acronym for “security, algorithms, networking, and data”).

Arguably the most prominent weapons in an artist’s arsenal against nonconsensual AI scraping, Glaze and Nightshade work in similar ways: by adding what the researchers call “barely perceptible” perturbations to an image’s pixels so that machine-learning models cannot read them properly. Glaze, which has been downloaded more than 6 million times since it launched in March 2023, adds what’s effectively a secret cloak to images that prevents AI algorithms from picking up on and copying an artist’s style. Nightshade, which I wrote about when it was released almost exactly a year ago this fall, cranks up the offensive against AI companies by adding an invisible layer of poison to images, which can break AI models; it has been downloaded more than 1.6 million times. 

 » …
Read More