13 detectors split after editing as Cara's scraper co-builds a protection tool
A preprint published on arXiv on 29 August reported that 13 AI detectors showed false-positive rates from 0.0 percent to 100.0 percent after professional editing. Cara, run by photographer Jingna Zhang, faced three major scrapes from 13 August; the first scraper said he regretted it and agreed to work with Zhang on an open-source tool to protect artists. Content authenticity was contested in measurement tools and on artist platforms in the same week.
Artificial Intelligence··Midday
Editing shifted detector verdicts
Three researchers compared 135,389 pairs of documents from a professional editing service covering 2018 to 2025, each pair a manuscript by a writer whose first language is not English and its edited version, and ran 13 detectors over both. False-positive rates on human-written text ranged from 0.0 percent to 100.0 percent depending on the detector. The same edits pushed a document's AI score up on some detectors and down on others, and how far a score moved tracked how much editing the text had received. The authors conclude that the style of professional editing acts as a confounding variable in these tools and raises fairness questions when they are used on academic work. The paper is a preprint on arXiv and was presented at EMNLP 2026, and the authors argue that a single detector verdict cannot be read as a stable label once professional copy-editing enters the workflow when the tools disagree with one another as sharply as they disagree with human judgment on unedited drafts.[1]
Cara faced three scraping waves
Cara, run since 2023 by photographer Jingna Zhang and a crew of volunteers and now home to about 1.5 million artists, was hit by three major scrapes starting on 13 August. The first scraper announced on a Reddit forum that he had pulled a 12-terabyte archive of 12 million works, close to the site's entire public library, for under 10 dollars, and later deleted the post. WIRED reports that the same person came to regret it and agreed to work with Zhang on a new open-source tool to protect artists. The second scrape took about 8.5 million links plus metadata such as usernames, titles and tags and uploaded them to Hugging Face; when takedown requests arrived, Hugging Face said it would notify the user to remove personal metadata but would not do the same for the links, because it hosts no copies of the artworks. In the third, on 22 August, 123,000 images along with text posts and user bios containing personal information were shared on a torrent site.[2]
Authenticity debate meets measurement and platforms
The detector preprint shows the line between human authorship and professional editing may not be reliably drawn by automated tools; disagreement among 13 detectors leaves open which verdict to trust. On the Cara side the problem runs the other way: artists' works were copied in bulk at low cost and personal metadata moved to third parties; Zhang opened a fundraiser for legal costs with a 120,000 dollar goal and had passed 100,000 dollars as of Thursday. The scraper agreed to work with Zhang on a protection tool, an unexpected collaboration between platform and scraper after Hugging Face said it would not remove link lists that point off-site. The second and third scrapes carried metadata and images into repositories and torrents where takedown paths differ. For readers the shared frame is that whether content is human- or machine-sourced remains unresolved at both measurement and ownership levels in the AI era, and both stories arrived in the same reporting week.[1], [2]
