SpaceXAI’s Grok chatbot is sending gibberish replies to a subset of users, TechCrunch reported August 20.
One user who asked the model to generate a PDF received nonsense that ran for several paragraphs, including the phrase match it without and your they and two for planets. Another user found that source links in the broken output pointed to reinforcement learning research sites.
Affected users told TechCrunch they were on Grok Lite and noticed problems as early as Wednesday morning. TechCrunch could not reproduce the issue in its own testing, and it appears to affect only a small group. Reddit’s Grok community has been flooded with complaints, and refreshing a session sometimes restores normal behavior, though some users said the problem persisted across refreshes.
The bug looks limited to direct queries on Grok.com. The Grok account on X was unaffected and confirmed the issues in replies to unhappy users, while SpaceXAI did not respond to a request for comment.
