Elon Musk's AI firm blames unauthorised change for chatbot's rant about 'white genocide' [View all]
Source: The Guardian
Fri 16 May 2025 06.32 EDT
Last modified on Fri 16 May 2025 07.29 EDT
Elon Musk’s artificial intelligence company has blamed an “unauthorised modification” for a glitch in its Grok chatbot that resulted in the tool ranting about “white genocide” in South Africa. In a post on Musk’s X platform, xAI said new measures would be brought in to ensure its employees cannot modify the bot’s behaviour without extra oversight.
The Grok bot repeatedly referred to a white genocide in South Africa – a discredited claim promoted by Donald Trump among other US populist figures – this week in responses to unrelated queries. One user of Musk’s X platform, which also hosts Grok, asked the bot to identify the location of a photo of a walking path, triggering a non sequitur swerve into “South Africa’s farm attack debate”.
xAI, the Musk-owned company that developed the chatbot, said in a post on X that the bot’s erratic behaviour was due to an unauthorised change made to the Grok bot’s system prompt, which guides a chatbot’s responses and actions. “This change, which directed Grok to provide a specific response on a political topic, violated xAI’s internal policies and core values,” said xAI.
Indicating that a rogue employee was to blame, the post added new measures would be brought in to ensure that xAI employees “can’t modify the prompt without review”. It said the code review process for prompt changes had been “circumvented” in the incident. A new 24/7 monitoring team is also being brought in to deal with answers not caught by automated systems, xAI said. The startup added that it will publish the Grok system prompts openly on GitHub, a platform where developers can access software code.
Read more: https://www.theguardian.com/technology/2025/may/16/elon-musks-ai-firm-blames-unauthorised-change-for-chatbots-rant-about-white-genocide
"It's easier to beg forgiveness than ask permission."
The desired intention has now been put out there.
= new reply since forum marked as read