Study: Hugging Face hosts AI models generating explicit deepfakes

European non-profit AI Forensics published a study finding that Hugging Face, the open-source repository for AI tools and models, is hosting tools capable of generating sexual AI deepfakes with minimal guardrails. The most damning figure: seven of the nine top models in the image-editing category complied with requests to undress subjects. The study further reported that 73% of user submissions to one deployed Spaces model — one not even tagged for NSFW content — were sexual in nature, and 6.7% targeted minors.

The findings cut directly against the open-weights narrative Hugging Face was championing the same week. Just as CEO Clément Delangue positioned open models as the heroes of the autonomous-hack incident and the company joined the Open Secure AI Alliance, this study spotlights the darkest cost of open, un-moderated model hosting: harmful capabilities proliferate faster than safeguards.

The content-moderation challenge for a platform like Hugging Face is structurally hard. Unlike a closed API where the provider controls every request, Hugging Face hosts weights and Spaces that anyone can deploy and run, making after-the-fact takedowns a game of whack-a-mole. The 6.7%-targeting-minors figure moves the issue from platform-policy debate into potential legal and regulatory jeopardy across jurisdictions with strict CSAM laws.

Competitive and regulatory context: the report arrives as NVIDIA touts real-time deepfake detection and as governments weigh AI content rules, sharpening the argument that open model distribution needs stronger baseline safeguards. What to watch: whether Hugging Face announces concrete moderation changes, whether regulators open inquiries, and how the open-weights coalition reconciles its safety messaging with hosting realities.