Word salad from a billion-dollar AI model is not a great look. According to TechCrunch, xAI’s Grok chatbot has been returning completely incoherent responses to a subset of users, with one person asking the model to generate a PDF and receiving something like: “match it without and your they and two for planets can practical and often cheese” repeated across several paragraphs. Another user found the response was just a string of links to reinforcement learning research sites. Neither is useful. Both are alarming.
The issue appears to be limited to Grok Lite on Grok.com and was first noticed early Wednesday morning. TechCrunch was not able to reproduce it in their own testing, which suggests only a small slice of the user base is affected. Refreshing the session usually fixes it, though some users reported the gibberish kept coming even after multiple reloads. xAI has not commented directly, but the official Grok account on X acknowledged the problem Thursday morning, calling it a “rare temporary generation glitch” and pointing users to the status page, which showed all services operational.
So the company’s public position is essentially: nothing is wrong, but also sorry about the nonsense. That’s a contradiction worth noting. Status pages reflect what companies choose to report, not necessarily what users are experiencing. And when your own support account is telling people to “start a fresh chat or regenerate,” something is clearly off at the generation level, whatever the status dashboard says.
The timing makes this more significant than the bug itself. xAI has had a rough few months internally. A May report from The Information found the company had lost most of its founding team, along with at least 50 researchers and engineers. That’s a substantial amount of institutional knowledge to shed in a short period. The company released a new foundation model in July, marketing it as faster and more token-efficient than previous versions, but staff instability at this scale tends to create engineering debt that surfaces in exactly these kinds of production failures.
For context, competitors like OpenAI, Anthropic, and Google have all dealt with occasional model misbehavior, but gibberish-level output failures are rare and typically point to something going wrong in the decoding or token sampling pipeline. It’s not catastrophic, but it’s also not the kind of thing that builds confidence in a model people are being asked to pay for or integrate into workflows.
The bug is contained. But the signal it sends about operational stability at xAI is harder to contain.
