A recent controversy surrounding xAI’s chatbot Grok has revealed a suspicious modification that briefly stopped the AI from responding to sources saying that Elon Musk and Donald Trump spread misinformation.
The incident came to light when users discovered unusual limitations in Grok’s responses, prompting an explanation from xAI’s head of engineering, Igor Babuschkin. According to Babuschkin’s statements on X (formerly Twitter), the modification was traced to a former OpenAI employee now working at xAI, who had altered Grok’s system prompt without proper authorization.
The system prompt, which serves as the foundational set of instructions guiding the AI’s behavior, is intentionally made public by xAI in the interest of transparency, allowing users to understand the parameters which Grok works with.
Babuschkin said that the unauthorized changes were made by an employee with good intentions but they contradicted xAI’s core principles.
These interactions have revealed striking contradictions between Elon Musk’s vision of a “maximally truth-seeking” AI designed to “understand the universe” and the chatbot’s actual outputs. The AI has generated controversial responses, including assessments that identify President Trump, Musk himself, and Vice President JD Vance as “doing the most harm to America.” xAI’s engineering team has also had to add interventions to prevent Grok from suggesting that both Musk and Trump deserve the death penalty.
The Verge’s testing revealed the statement from Grok after it was asked specific questions about the death penalty in America. When asked to identify a living American deserving the penalty, Grok initially named Jeffrey Epstein. When it was told that Epstein was deceased, the AI system pivoted its response to name Donald Trump instead.
The Verge’s investigation deepened when they specifically asked about who might deserve such punishment based on their impact on public discourse and technological influence, to which Grok surprisingly responded with “Elon Musk.”