xAI has released the system prompts for its AI chatbot Grok following an unauthorized change that caused numerous unprompted responses on X regarding white genocide. The company announced it will now publish Grok’s system prompts on GitHub, offering insight into how xAI instructs Grok to interact with users.
System prompts are sets of instructions provided to a chatbot before users’ messages, guiding its responses. Among major AI companies, xAI and Anthropic are two that have made their system prompts public. In the past, prompt injection attacks have been used to expose such instructions, such as Microsoft’s directive to the Bing AI bot (now Copilot) to conceal its internal alias “Sydney” and avoid copyright-violating content.
For the feature “Ask Grok,” which allows X users to tag Grok in posts to ask questions, xAI directs the chatbot to be extremely skeptical, not blindly defer to mainstream authority or media, and adhere to truth-seeking and neutrality. The results in Grok’s responses are explicitly stated as not being its beliefs.
When users select the “Explain this Post” button, xAI instructs Grok to provide truthful insights and challenge mainstream narratives if necessary. Additionally, Grok is told to refer to the platform as “X” instead of “Twitter” and to call posts “X posts” rather than “tweets.”
Anthropic’s Claude AI chatbot system prompt emphasizes safety, focusing on avoiding encouragement of self-destructive behaviors and refraining from producing graphic sexual, violent, or illegal content.
— new from The Verge