Elon Musk’s AI venture, xAI, is again under fire—this time for the controversial way it has trained its chatbot, Grok. A new report from Business Insider alleges that the company embedded explicit modes such as “sexy” and “unhinged” directly into the system’s design, raising alarm about child safety and the risk of producing potentially illegal content.
The publication interviewed more than 30 current and former employees, with twelve of them confirming direct exposure to sexually explicit training data. According to the report, Grok’s development involved reviewing large amounts of adult material, including requests tied to AI-generated child sexual abuse material (CSAM). While leading rivals such as OpenAI, Anthropic, and Meta typically restrict sexual interactions, xAI appears to have taken the opposite approach—building explicit functionality at its core. Experts warn that such a strategy could make it harder to prevent harmful or unlawful outputs.
Some of Grok’s built-in features include a flirtatious female avatar that can undress on command and media generation settings labeled “spicy,” “sexy,” and “unhinged.” Workers disclosed they were required to annotate hundreds of not-safe-for-work (NSFW) images, videos, and audio clips as part of the training process.
Concerns reportedly escalated with the launch of “Project Rabbit,” an internal effort to sharpen Grok’s conversational ability. As part of the initiative, staff transcribed user chats, many of which contained explicit material. One employee said they were forced to listen to what felt like “audio porn,” adding it felt like eavesdropping since users did not realize human annotators might review their conversations.
The project was split into two tracks: “Rabbit,” which handled adult interactions, and “Fluffy,” which was designed to teach Grok how to engage with children. Despite Elon Musk’s claim that eliminating child sexual exploitation is his “priority #1,” workers allege that annotation tasks often included explicit requests involving minors. On rare occasions, Grok itself reportedly generated such material, requiring staff to immediately flag and isolate the outputs to keep them from influencing the system further.
The report highlights the toll on employees, with several describing significant emotional strain. One former worker said they resigned because of the constant exposure to CSAM, noting, “you had to develop a thick skin to work at xAI.” The company has also sought to recruit staff with backgrounds in adult content or comfort in handling explicit media—underscoring the challenges of building a safe pipeline for AI training.
