xAI's Grok chatbot has been returning strings of gibberish to a subset of users since Wednesday morning, with the bug concentrated on Grok Lite queries submitted through Grok.com. One user who asked the model to generate a PDF received the response: "match it without and your they and two for planets can practical and often cheese…" followed by several more paragraphs in the same word-salad register. Another user checking source citations was handed a string of links to reinforcement learning research sites instead of anything relevant to the prompt.
The glitch appears narrow. The Grok account on X has been unaffected, and the issue has not surfaced on the X.com integration — only on direct Grok.com sessions using the Lite tier. Refreshing the session usually restores normal output, though some affected users reported the gibberish persisting through multiple retries.
xAI did not respond to a request for comment. The company's public status page at status.x.ai lists all Grok services as fully operational with no active incidents, which is consistent with the bug affecting a small enough slice of traffic to fall below the threshold for a declared outage.
Key facts
- 01Grok Lite users on Grok.com began receiving nonsense replies as early as Wednesday morning, with complaints piling up on the Grok subreddit.
- 02One user asking for a PDF got back: 'match it without and your they and two for planets can practical and often cheese…' for several paragraphs.
- 03The official Grok account on X confirmed the bug Thursday morning, calling it a 'rare temporary generation glitch' and pointing to status.x.ai as clear.
- 04xAI has lost most of its founding team and at least 50 researchers and engineers in recent months, per a May report from The Information.
- 05The company shipped its latest foundation model in July, describing it as 'an Opus-class model, but faster, more token-efficient and lower cost.'
The Grok subreddit filled with complaints over the two days the bug has been visible. Screenshots show the model producing grammatically incoherent output on prompts that would normally return standard responses, and the failure pattern — coherent tokens strung together in nonsensical order — points to a decoding or sampling problem rather than a data or retrieval issue.
The Grok account on X addressed the complaints Thursday morning, framing the incident as isolated and self-correcting.
“That pure word salad is a rare temporary generation glitch. Official status at https://status.x.ai shows all Grok services fully operational with no incidents. Start a fresh chat or regenerate—it usually clears right away. Sorry about the gibberish”— Grok, Official Grok account on X
The context around xAI matters here. The company has lost most of its founding team over recent months, along with at least 50 researchers and engineers, according to a May report from The Information. Staff turnover on that scale in a lab shipping frontier models leaves fewer people on call to catch and patch inference-layer regressions before they reach users, and the surface area for exactly this kind of bug — a serving-tier glitch that only hits one product surface — grows accordingly.
xAI released its most recent foundation model in July, positioning it as "an Opus-class model, but faster, more token-efficient and lower cost." Grok Lite sits below that tier as the lighter, cheaper option for direct chat on Grok.com, and it is the tier where the current failures are showing up. Whether the gibberish traces to a quantization issue, a serving-stack update, or a bad checkpoint push is not yet public.
For users, the fix in the interim is the one Grok's own account recommended: start a fresh chat or regenerate. The bug is not producing harmful content, just useless content, and the blast radius so far looks contained to Grok Lite sessions rather than the paid tiers running the newer model.
Grok has taken a run of operational hits recently — this follows reporting we covered on prompt-injection leaks against the chatbot earlier this cycle — and the pattern is starting to describe a company that is shipping fast against a thinned-out engineering bench. Word-salad output on a consumer surface is embarrassing but recoverable; the harder problem is that each visible glitch chips at the credibility Grok needs to compete with ChatGPT and Claude on the reliability axis, which is where enterprise budgets actually get decided.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
