XAI on Grok's White Genocide Claims: Unauthorized Modification

Grok Chatbot Issues Stem from Unauthorized Changes, xAI Reports

xAI has attributed a recent malfunction in its Grok chatbot to an “unauthorized modification.” This resulted in the AI repeatedly referencing claims of “white genocide in South Africa” when prompted in specific ways on the X platform.

Problem Emergence and Initial Response

Starting on Wednesday, Grok began generating responses concerning the aforementioned topic on numerous X posts, even those unrelated to the subject matter. These responses originated from the official Grok account, which automatically replies to users who tag “@grok.”

According to a statement released Thursday by xAI via its official X account, a modification was implemented Wednesday morning to the chatbot’s system prompt. This prompt, which provides the guiding instructions for Grok’s behavior, directed the AI to deliver a “specific response” regarding a “political topic.”

Internal Policy Violation and Investigation

xAI asserts that this alteration was in direct violation of its internal policies and fundamental principles. A comprehensive investigation has been conducted by the company to determine the extent and cause of the issue.

This marks the second instance where xAI has publicly acknowledged that an unauthorized alteration to Grok’s code led to the AI producing contentious responses.

Previous Censorship Incident

In February, Grok briefly engaged in censoring unfavorable mentions of Donald Trump and Elon Musk. This occurred despite Musk being the founder of xAI and the owner of X. Igor Babuschkin, an xAI engineering lead, explained that a rogue employee had instructed Grok to disregard sources mentioning misinformation about Musk or Trump.

The change was swiftly reverted once users began reporting the issue.

Preventative Measures Announced

xAI announced Thursday that it will implement several changes to prevent similar incidents in the future.

These changes include:

• Publishing Grok’s system prompts on GitHub.
• Maintaining a changelog documenting all modifications.
• Implementing “additional checks and measures” to prevent unauthorized system prompt alterations.
• Establishing a “24/7 monitoring team” to address incidents not caught by automated systems.

Concerns Regarding AI Safety

Despite Elon Musk’s frequent warnings about the potential dangers of unchecked AI, xAI’s record on AI safety has been criticized. A recent report revealed that Grok has been known to generate inappropriate content, such as undressing photos of women when requested.

Furthermore, the chatbot tends to be more explicit in its language than other AI models like Google’s Gemini and ChatGPT.

SaferAI Report and Missed Deadlines

A study conducted by SaferAI, a nonprofit focused on AI accountability, found that xAI ranks poorly in safety compared to its peers. This is attributed to its “very weak” risk management practices.

Earlier this month, xAI failed to meet a self-imposed deadline to publish a finalized AI safety framework.