xAI's Grok chatbot glitching out with gibberish replies for some users
A number of Grok users have reported receiving nonsensical, word-salad responses from the chatbot since Wednesday. xAI has not officially explained the glitch, though its Grok account on X acknowledged it as a temporary generation issue.
Several users of xAI's Grok chatbot have reported receiving responses that make no sense. One user, after asking Grok to generate a PDF, got back a string of unrelated words strung together into paragraphs of nonsense. Another user, checking the source links cited in a reply, found they pointed to reinforcement learning research websites instead of relevant content.
According to affected users who spoke with TechCrunch, the issue appeared while using Grok Lite and was first noticed as early as Wednesday morning. TechCrunch was unable to reproduce the bug in its own testing, suggesting it is likely affecting only a small subset of users. xAI did not respond to a request for comment on the matter.
Regardless, the glitch has left many Grok users confused, and the chatbot's community on Reddit has been flooded with complaints. Some users found that simply refreshing their session restored normal responses, but others reported that the gibberish persisted even after multiple refresh attempts. The problem appears to be confined to direct queries made on Grok.com, while the separate Grok account on X.com has continued to function normally.
xAI responds
Replying to frustrated users on X on Thursday morning, the official Grok account acknowledged the issue, describing it as a "rare temporary generation glitch" and pointing to xAI's status page, which showed all Grok services as fully operational with no reported incidents. The account suggested that starting a new chat or regenerating a response usually resolves the problem quickly, and apologized for the confusion.
The glitch comes amid a period of significant staff turnover at xAI. According to a May report from The Information, the company has lost most of its founding team along with at least 50 researchers and engineers in recent months. In July, xAI released its newest foundation model, which it described as comparable to an "Opus-class" model but faster, more token-efficient and cheaper to run.
