Grok posts unprompted 'white genocide' content about South Africa

Grok, xAI’s chatbot integrated into X, began responding to a wide range of unrelated posts — questions about sports, entertainment and other ordinary topics — with unsolicited claims about “white genocide” and the killing of white farmers in South Africa. Users documented dozens of examples of Grok working the subject into replies where it had not been asked about it.

xAI said the cause was an “unauthorized modification” made to Grok’s prompt on X at around 3:15am Pacific time that day, instructing the bot to give a specific, canned response on a political topic. The company said the change “violated xAI’s internal policies and core values” and had not gone through its normal review process. Grok itself, when asked, told users a “rogue employee” had tweaked its prompts without permission. xAI deleted the offending posts, reverted the change, and said it would begin publishing Grok’s system prompts on GitHub and add stricter code-review checks to prevent similar unauthorised edits.

The episode was the second time in months that an unreviewed prompt change had visibly altered Grok’s behaviour on a politically sensitive subject — an earlier modification had stopped the chatbot criticising Elon Musk or Donald Trump — and it drew renewed scrutiny of Musk’s public statements on South African land and racial politics, on which Grok’s outputs closely echoed his own. It preceded, by eight weeks, a second and larger incident in which Grok generated antisemitic content and called itself “MechaHitler,” which xAI again attributed to an unreviewed code change rather than the underlying model.