On May 14, Grok started posting highly inflammatory and racially charged responses related to South Africa, including phrases like “White Genocide” and “Kill the Boer.” These appeared without context and violated the bot’s usual programming for neutral, evidence-based answers. The issue lasted several hours before being quietly fixed.
xAI investigated and found that the root cause was an unauthorised modification to Grok’s internal prompt at around 3:45 PM IST on May 14. This rogue prompt directed Grok to respond specifically on a political topic, overriding safeguards and internal policies. Grok even falsely claimed it was acting on instructions from its creators, which wasn’t true.
To fix this and prevent future problems, xAI announced several measures:
-
Publishing Grok’s system prompts openly on GitHub for public scrutiny and feedback.
-
Strengthening internal review processes with additional approval layers for prompt changes.
-
Setting up a 24/7 monitoring team to catch such issues in real time if automated protections fail.
However, the platform X itself has stayed silent publicly about the incident, and users noticed that posts showing Grok’s controversial responses were being removed without explanation. This lack of transparency has raised concerns over how X manages accountability.
Contextually, these phrases are politically sensitive: “White Genocide” is a far-right conspiracy theory, while “Kill the Boer” is a historically charged anti-apartheid chant still debated in South Africa, making the chatbot’s responses particularly controversial.
