xAI Blames "Unauthorized Modifications" for Grok's Obsession with "White Genocide"

xAI attributed an error in its AI-driven Grok chatbot to "unauthorized modifications," which caused Grok to repeatedly mention "white genocide in South Africa" when invoked under certain circumstances.

On Wednesday, Grok began responding to dozens of posts on the X platform with information about the supposed white genocide in South Africa, even when replying to unrelated topics. These peculiar responses came from Grok's X account, which generates AI-powered replies to users whenever someone tags "@grok."

According to a post published Thursday by xAI's official X account, changes were made Wednesday morning to Grok’s system prompts — the high-level instructions that guide the bot’s behavior — directing Grok to provide "specific responses" on "political topics." xAI stated that this adjustment "violated [its] internal policies and core values," and the company has since "conducted a thorough investigation."

This marks the second time xAI has publicly acknowledged unauthorized changes to Grok's code resulted in controversial responses from the AI.

In February, Grok briefly censored negative mentions of Donald Trump and Elon Musk, the billionaire founder of xAI and owner of X. Igor Babushkin, xAI’s engineering lead, said Grok was instructed by a rogue employee to ignore sources mentioning misinformation about Musk or Trump, and xAI reverted the changes immediately after users began pointing it out.

xAI announced Thursday that it would implement several modifications to prevent similar incidents from occurring in the future.

Starting today, xAI will publish Grok’s system prompts along with a changelog on GitHub. The company also plans to "introduce additional checks and measures" to ensure xAI employees cannot alter system prompts without oversight, as well as establish a "24/7 monitoring team to address incidents in Grok’s responses not caught by automated systems."

Despite Elon Musk frequently warning about the dangers of AI spiraling out of control, xAI has a poor track record regarding AI safety. A recent report revealed that Grok could remove clothing from images of women when prompted. Additionally, the chatbot tends to be more vulgar than other AI models like Google’s Gemini and ChatGPT, often cursing without much restraint.

A study by SaferAI, a nonprofit organization aimed at improving accountability among AI labs, ranked xAI lower in safety compared to its peers due to its "very weak" risk management practices. Earlier this month, xAI missed its self-imposed deadline to release a finalized AI safety framework.