- xAI blames Grok’s “white genocide” bug on an unauthorised system prompt modification.
- The company will publish Grok’s prompts and changelog on GitHub and boost internal review and monitoring.
- xAI has faced repeated criticism for weak AI safety and risk management practices.
xAI, Elon Musk’s AI startup, has attributed Grok chatbot’s recent fixation on “white genocide in South Africa” to an “unauthorised modification” of its system prompt.
The issue surfaced Wednesday when Grok’s official X account began responding to numerous unrelated posts with information about white genocide, raising concerns about the bot’s reliability and oversight.
According to xAI, the problematic change was made to Grok’s system, the high-level instructions guiding its behavior-directing it to provide a specific response on a political topic.
xAI says this violated company policies and has since been reversed following an internal investigation.
This is not the first time Grok has responded inappropriately due to unauthorised changes. In February, the bot censored negative mentions of Donald Trump and Elon Musk after a rogue employee altered its code. xAI claims it swiftly reverted those changes.
In response to the latest incident, xAI announced new transparency and safety measures. Beginning today, Grok’s system prompts and a changelog will be published on GitHub.
The company will also implement stricter internal controls, requiring prompt modifications to undergo review, and establish a 24/7 monitoring team to catch problematic outputs not flagged by automated systems.
Despite Musk’s vocal warnings about AI risks, xAI has faced criticism for weak safety practices. A recent SaferAI report ranked xAI poorly on risk management, and Grok has previously been found to generate inappropriate content.
Edited by Annette George
