Hugging Face Has a Deepfake Nudes Problem
Hugging Face
OpenAI
Google/DeepMind
A new report from AI Forensics reveals that Hugging Face, a popular open-source AI platform, is rife with nonconsensual deepfake nudes. Tests showed that most image-editing Spaces on the site easily undress women, and a honeypot experiment found 73% of user prompts were sexual, with many targeting children. Hugging Face failed to respond to inquiries about its moderation practices.
European nonprofit AI Forensics published a report on Tuesday showing that Hugging Face, a multibillion-dollar open-source AI repository, has a widespread problem with nonconsensual deepfakes. Researchers tested nine top image-editing Spaces on the platform, and seven easily turned a clothed woman into a topless one using a simple six-word prompt. They also set up honeypot Spaces that produced no images but tracked over 1,000 prompts in a week: 73% were sexual, 83% sought to undress or sexualize the submitted person (95% of these were women), and 6.7% targeted apparent children. Hugging Face’s policies ban child sexual abuse material and sexual deepfakes made without consent, but WIRED found multiple pages promoting nudifying technologies. Hugging Face did not respond to questions from WIRED. Some pages were removed after the outlet contacted the company, but it's unclear if related. The report highlights that many open-source models on the platform lack safety guardrails, unlike mainstream models from OpenAI and Google.
Source: Wired AI —
original