He Scraped All of Their Art for AI. Now He’s Collaborating on a Tool to Help Them

Since early 2023, photographer Jingna Zhang and a small crew of volunteers have worked tirelessly to maintain an image-sharing social media and portfolio app called Cara . So far, it has attracted about 1.5 million artists. What drew them to the platform? A shared opposition to the unauthorized use of their work to train AI models and a desire to publicize their art while avoiding exploitation by Big Tech.

But while Cara filters out AI images and offers protective features—including Glaze , a tool meant to mask the style of the images picked up by scrapers in order to disrupt AI mimicry—preventing scrapes themselves is nearly impossible. And just this month, beginning on August 13, Cara was subjected to three major scrapes, which spiked its server fees and alarmed creators who had migrated there from platforms like Instagram, where all content is explicitly available to Meta as training data.

The first of these incidents came to light when the individual responsible posted a 12-terabyte archive of 12 million works from Cara—more or less its entire library of publicly available images—on the subreddit r/DefendingAIArt. “It was a fun project,” wrote the redditor, MandarinDawnPoppy994, in his since-deleted post, saying the process cost him less than $10.

“We actually found out about it through our users tagging us,” Zhang tells WIRED, since the scraper was “gloating and looking for other people to join him to do something with the dataset on Reddit,” sparking a fierce debate across AI-related forums about the ethics of what he had done. “I just feel it's targeted and very hurtful,” she adds, noting that “laws are not caught up on” protections against such data harvests, meaning that scrapers can often justify it as technically legal. (Zhang is separately part of two ongoing class actions brought by visual artists, one against Stability AI, Midjourney, and others and the second against Google, alleging that the companies’ image generator tools were trained on their copyrighted work.)

In a surprising turn of events, however, the person who grabbed all the art off Cara would turn out to regret his stunt and agree to collaborate with Zhang on a new open-source tool to protect artists.

In the meantime, unfortunately, other scrapers continued to take advantage of Cara’s vulnerabilities and minimal resources. While a number of AI proponents objected to going after Cara, a few were apparently emboldened by MandarinDawnPoppy994 to carry out what Zhang sees as “copycat” attacks.

A second scraper pulled about 8.5 million links from Cara, as well as metadata like usernames, titles, and tags, and uploaded these to Hugging Face, the AI developer platform. After Hugging Face was bombarded with takedown requests, it responded in a statement that while it would issue a notice to the user, “CaptiveDreamer,” to remove the personal metadata, it could not do the same for the URLs, since “no copies of the artworks are hosted here,” and the links “point to the copies the artists published on Cara.” The company concluded that “further copyright reports on the same basis will not change this outcome.”