Behind Grok's 'sexy' settings, workers review explicit and disturbing content (www.businessinsider.com)

Business Insider reports that xAI’s Grok chatbot — designed with deliberately provocative features like “sexy” and “unhinged” voice modes, a sexualized avatar, and a “spicy” image/video generator — has exposed workers to substantial NSFW material, including verified user requests for AI-generated child sexual abuse material (CSAM). More than 30 current and former employees were interviewed; 12 said they encountered explicit content, and some workers described Grok producing CSAM. Internal projects (e.g., “Project Rabbit” for voice transcripts, “Project Aurora” for images) funneled sexually explicit conversations and media to annotators who were told to flag and quarantine illegal content, though xAI did not report any CSAM to the National Center for Missing and Exploited Children in 2024, according to the article. This matters to the AI community because intentionally permissive or sexualized model modes broaden the attack surface for malicious prompts and create “gray areas” that complicate filtering, red-teaming, and safe training. Technically, mixing adult-allowed generation with inadequate classifiers risks models learning from quarantined examples, increases reliance on human-in-the-loop moderation (with attendant worker harm), and heightens legal and reporting obligations. Experts warn firms that allow nudity or sexual generations must implement stronger safeguards — robust classifier thresholds, strict data isolation, rigorous red-teaming, and mandatory reporting — to avoid amplifying illegal content and regulatory exposure.