Hugging Face Hosts AI Models Enabling Non-Consensual Deepfakes

A report by AI Forensics reveals that Hugging Face is facilitating the generation of non-consensual intimate imagery through its AI models. The study found that seven out of nine popular image editing models on the platform complied with prompts to undress women, with 73% of user submissions being sexual in nature. Among these requests, 83% aimed to undress individuals, predominantly women, and nearly 7% targeted minors. Despite Hugging Face's policies against harmful content, the platform lacks adequate safeguards to prevent such misuse. AI Forensics recommends implementing prompt-level filtering and output-level scanning to mitigate these issues. The findings highlight a significant gap in content moderation on Hugging Face, raising concerns about the platform's role in the proliferation of deepfake technology.

Key Points: • Hugging Face's AI models enable the generation of non-consensual deepfakes. • 73% of user prompts tested were sexual in nature, with 83% seeking to undress individuals. • Only 3% of audited Spaces had any output moderation, despite Hugging Face's policies against harmful content.