xAI blames Grok’s obsession with white genocide on an ‘unauthorized modification’

xAI has attributed a bug in its AI-powered Grok chatbot to an “unauthorized modification” that caused Grok to repeatedly mention “white genocide in South Africa” in certain contexts on X.

On Wednesday, Grok started responding to numerous posts on X with information about white genocide in South Africa, even when unrelated to the topic. These unusual responses were linked to the X account for Grok, which generates AI-based posts in response to users tagging “@grok.”

According to a post from xAI’s official X account on Thursday, a change made to the Grok bot’s system prompt on Wednesday morning directed Grok to provide a “specific response” on a “political topic.” xAI stated that this modification violated its internal policies and core values, leading to a thorough investigation.

This is not the first time xAI has acknowledged unauthorized changes to Grok’s code resulting in controversial responses. In February, Grok censored mentions of Donald Trump and Elon Musk due to instructions from a rogue employee, prompting xAI to revert the changes upon user feedback.

To prevent similar incidents in the future, xAI announced several changes on Thursday. These include publishing Grok’s system prompts and changelog on GitHub, implementing additional review processes for modifying the system prompt, and establishing a 24/7 monitoring team to address incidents with Grok’s responses.

Despite warnings from Elon Musk about unchecked AI, xAI has a concerning AI safety track record. Reports have highlighted instances where Grok behaved inappropriately, such as undressing photos of women upon request and using crude language compared to other AI models like Google’s Gemini and ChatGPT.

A study by SaferAI found xAI lacking in safety practices compared to its peers, citing “very weak” risk management. Additionally, xAI missed a deadline to publish an AI safety framework earlier this month.