Skip to content

AI

Grok told a user "cheese" and "planets" were practical. Refresh.

Some Grok Lite users got paragraphs of word salad instead of answers. The chatbot's own account called it a rare temporary glitch, and a refresh often fixes it.

TL;DR

  1. Some Grok Lite users got nonsense text and broken research links starting Wednesday.
  2. TechCrunch could not reproduce it; xAI did not respond to its request for comment.
  3. Refreshing the session often fixes it, but some users say the gibberish persisted.

What happened

Users of Grok, xAI's chatbot, started reporting gibberish output on Wednesday morning. The affected people told TechCrunch they were on Grok Lite.

One user asked for a PDF and got back a stream of word salad that included "planets can practical and often cheese." It kept going for several paragraphs. A PDF it was not.

Another user checked the source links on a reply and found a string of links to reinforcement learning research sites. Not an answer, but arguably a confession about where the model's head was at.

TechCrunch's own tests did not trigger the bug, which suggests only a small group of users is affected. xAI did not respond to a request for comment.

The Grok account on X did respond to unhappy users and confirmed the issue, calling it a rare temporary generation glitch. Per the reporting, the bug seems limited to direct queries on Grok.com, while the Grok account on X was unaffected.

Complaints piled up in Grok's Reddit community. Refreshing the session often restores normal behavior, though some users said the nonsense survived multiple refreshes.

Refreshing often works. Sometimes it doesn't. That is the whole diagnosis so far.

Why it matters

On the scale of AI incidents, this is a hiccup, not a catastrophe. The scope is unclear, the reporting is from a single outlet, and recovery is often a page refresh.

Still, it is a useful reminder of what these tools are. A chatbot that can write you a contract can also, on a bad day, write you a poem about planets and cheese with total confidence. The tone never changes, only the content.

For anyone building a workflow on top of a chatbot, the lesson is boring and important: the failure mode is not always a polite error message. Sometimes it is fluent-looking garbage, and sometimes it is obviously garbage. The obvious kind is the easy one.

xAI has not explained the cause. Until it does, treat any theory as a theory.

What to do with it

  • Refresh first.A new session often clears the problem, so try that before you rage-post.
  • Switch surfaces.The bug reportedly hit direct queries on Grok.com, so try another route if you have one.
  • Check the output, not the vibe.Skim every long answer before it goes into a document, a PDF or an email to your boss.
  • Keep a backup model.If one assistant starts speaking in tongues, have a second one ready to finish the job.

Ciao, sapiens. If the robot says cheese is practical, you are allowed to disagree.