An online art platform built to resist unauthorized AI training has endured three major scraping incidents in less than two weeks. The first person to copy its library has now apologized and is helping its founder develop a tool that could alert artists when their work appears in public AI datasets.
Cara hosts portfolios for roughly 1.5 million artists and filters out AI-generated images. Yet its public design makes complete prevention of scraping nearly impossible, leaving a small volunteer team to absorb technical costs, legal uncertainty and anger from users.
The collaboration offers a practical response, but it does not restore control over every copied image or make Cara immune to another attack.
Three scrapes exposed the platform’s limits
The first incident began on August 13. A Reddit user later identified by the screen name Heft copied a 12-terabyte archive containing about 12 million publicly available works—almost Cara’s full image library—and posted it online. He said the operation cost less than $10.
Cara founder Jingna Zhang learned about the archive when users flagged a Reddit post that appeared to invite others to use the dataset. The scrape increased server costs and alarmed artists who had joined Cara specifically to avoid having their work taken for AI training.
Copycat incidents followed. A second scraper collected approximately 8.5 million links plus usernames, titles and tags, then uploaded the material to Hugging Face. The platform agreed to ask for removal of personal metadata but said it would not remove links pointing to images artists had published publicly on Cara.
On August 22, a third scraper obtained 123,000 images as well as posts and user biographies containing personal information. Zhang opened a legal fundraiser with a $120,000 target and had collected more than $100,000 by the time Wired reported the story.
The first scraper reversed course
Heft, a North American student with software experience, told Wired that he initially treated the scrape as a technical and archival project. He said he had not planned to publish the data but chose to provoke users on Reddit and underestimated the distress that followed.
After seeing artists report panic attacks and delete their portfolios, he apologized, removed his dataset and began speaking directly with creators. Zhang said those conversations changed his understanding of how personally artists viewed their work.
Heft then joined Cara’s Discord community as a technical troubleshooter. His knowledge of scraping helps the team assess proposed defenses and identify controls that can be bypassed quickly.
This reversal does not erase the original copying or the other datasets. It is a reported change in one participant’s behavior, not evidence that hostile scraping has stopped.
Lantern focuses on detection after collection
Zhang and Heft are now collaborating on an open-source project called Lantern. The planned tool creates a one-way fingerprint for an artist’s image without retaining the original work on its own platform.
Lantern is designed to scan newly available AI datasets. If it detects a matching fingerprint, the artist would receive a notification and a link to the dataset, allowing a removal request or takedown notice.
The approach accepts an uncomfortable technical premise: a public website cannot guarantee that its images will never be copied. Instead, it tries to give creators earlier visibility after collection occurs.
Lantern is still at an early stage, and Wired describes it as an imperfect response to limited legal protection. Detection does not guarantee removal, particularly when hosts distinguish between copied files, links and metadata. Fingerprinting also cannot prevent the bandwidth costs and disruption created by the scrape itself.
Cara has added temporary measures such as login gates, but Zhang said stronger barriers can make the platform unpleasant to use without solving an internet-wide problem. The new project therefore shifts the goal from perfect exclusion to traceability. It may help artists discover where their work travels, even when technology and law cannot reliably stop the initial copy.
