TL;DR
Some Grok Lite users received long, incoherent responses on Grok.com beginning around August 19, 2026. Grok’s account acknowledged a temporary generation glitch, but xAI did not explain the cause or confirm a permanent fix.
xAI’s Grok chatbot began sending long strings of incoherent text to some Grok Lite users on Grok.com, according to affected users who spoke with TechCrunch. Grok’s official account acknowledged a temporary generation glitch, but xAI did not provide a technical explanation or confirm how many people were affected.
The malfunction appeared during ordinary requests, including one user’s attempt to have Grok generate a PDF. Instead of completing the request, the chatbot produced several paragraphs of disconnected words, beginning with a sequence about planets and cheese. Another user who inspected the response’s sources encountered links to reinforcement-learning research sites, though the relationship between those links and the garbled output was not established.
Affected users told TechCrunch that they were using Grok Lite and had seen the problem as early as Wednesday morning, August 19. TechCrunch said it could not reproduce the behavior during its own testing and judged that the problem was probably limited to a small portion of users. That estimate has not been supported by public incident figures from xAI.
Reports indicated that starting a fresh conversation or regenerating a response often restored normal chatbot output. Some users, however, said the gibberish continued after several refreshes. The malfunction appeared limited to direct requests on Grok.com; Grok’s account on X continued responding normally during the reported period.
Grok’s answers turned into word salad
Some Grok Lite users received long, incoherent responses on Grok.com beginning around August 19, 2026. Grok’s official account acknowledged a temporary generation glitch, but xAI did not explain the cause, disclose the scope, or confirm a permanent fix.
A visible failure with an invisible cause
Ordinary prompts sometimes produced paragraphs of disconnected words instead of useful answers. The most striking example appeared during a request to create a PDF and began with an unrelated sequence involving planets and cheese.
Incoherent generation
Users received long strings of grammatically disconnected terms rather than an answer that matched the request.
Credible-looking links
One user found links to reinforcement-learning research sites, but no relationship between those sources and the malformed response was established.
Probably limited
TechCrunch could not reproduce the behavior and judged that it likely affected a small subset of users. xAI released no public incident figures.
Normal request submitted
A Grok Lite user asks for an ordinary task, including document generation.
Gibberish returned
The service responds, but the text is long, disconnected, and unusable.
Glitch acknowledged
Grok’s account describes the behavior as a rare temporary generation issue.
Explanation remains absent
No root cause, affected-user count, rollback, or permanent fix is confirmed.
What the evidence supports
The central distinction is between confirmed user reports and plausible explanations. Model changes, routing failures, infrastructure updates, and staffing turnover were not tied to the incident by published evidence.
| Question | Evidence status | What is known |
|---|---|---|
| Did malformed answers occur? | ✓ Confirmed | Affected users supplied examples of disconnected output. |
| Was Grok.com involved? | ✓ Confirmed | Reports concerned direct requests made on Grok.com. |
| Was Grok on X affected? | ✗ Not observed | The account on X continued responding normally during the reported period. |
| Was there a broad outage? | ✗ Not confirmed | TechCrunch could not reproduce the behavior; xAI published no scope data. |
| Did restarting help? | ~ Sometimes | A fresh chat or regenerated answer restored normal output for some users, but not all. |
| Was a permanent fix deployed? | ~ Unknown | xAI did not confirm a fix, rollback, or corrective action. |
| Did a recent model cause it? | ~ Unestablished | No reporting connected the malfunction to a model, routing, or infrastructure change. |
Online does not always mean usable
xAI’s public status page reportedly showed services operating normally while users were receiving defective answers. Conventional uptime monitoring can confirm that a system responds without measuring whether its output is coherent, relevant, or trustworthy.
Evidence visibility
Bar lengths visualize the relative completeness of public information, not measured incident percentages.
Practical user response
Start a fresh conversation
This was the workaround recommended by Grok’s account.
Regenerate the answer
Some sessions recovered after a new generation attempt.
Inspect generated files
Verify that documents are complete, coherent, and responsive to the original request.
Verify every source
Credible-looking research links do not prove that an answer is accurate or logically grounded.
How a generation-quality failure travels
The incident shows why reliability must be assessed across the complete path from a user’s request to the action taken with the answer.
Ordinary request
A user asks for a document, summary, or research task.
Broken generation
The service remains responsive but produces incoherent text.
Plausible artifacts
Links or formatting may still create an appearance of credibility.
Independent review
The user validates meaning, sources, and generated files.
Use or reject
Only verified output should influence consequential work.
The open question is not whether it happened, but why.
A meaningful resolution would require an xAI incident report identifying the affected service, duration, user scope, root cause, and corrective action. Until then, successful regeneration is a session-level workaround—not evidence that the underlying fault has been permanently removed.
Reliability Questions for Grok Users
The episode matters because users may rely on AI assistants to create documents, summarize information or provide research support. A response made of random words is easy to identify as defective, but the attached research links show how a broken answer can still contain credible-looking elements. Users handling consequential work need to verify both generated documents and any sources supplied by the chatbot.
The incident also exposes a gap between service availability and the quality of a model’s output. Grok’s account said xAI’s status page showed all services operating normally while users were reporting unusable answers. A system can remain online and responsive while its answers are defective, meaning conventional uptime reporting may not capture generation-quality failures.
As an affiliate, we earn on qualifying purchases.
Grok’s Recent Model Changes
xAI released its most recent foundation model in July 2026, describing it as an “Opus-class” system that was faster, more token-efficient and less expensive. The available reporting did not establish whether Grok Lite used that model when the malformed responses appeared, or whether a recent model, routing or infrastructure change played any part.
The company has also experienced substantial staff turnover. The Information reported in May that xAI had lost most of its founding team and at least 50 researchers and engineers. No evidence cited by TechCrunch connected those departures to this glitch, so treating staffing changes as its cause would go beyond the confirmed facts.
“That pure word salad is a rare temporary generation glitch.”
— Grok’s official account on X
As an affiliate, we earn on qualifying purchases.
Cause and Scope Stay Unknown
It is not yet clear what caused the malformed responses, whether the fault was tied to a particular model version, or how many accounts encountered it. xAI did not respond to TechCrunch’s request for comment, and its public status page reportedly listed no active incident. The company also did not say whether it had deployed a fix, rolled back a change or merely expected affected sessions to recover when restarted. Reports that refreshing helped some users do not establish that the underlying fault was removed.
AI chatbot response quality monitor
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Users Await a Technical Explanation
The next meaningful development would be an xAI incident report identifying the affected service, the duration of the problem and any corrective action. Until then, users who encounter gibberish can follow the account’s advice to start a new chat or regenerate, while checking important output independently. Continued complaints would indicate that the session-level workaround is incomplete; a sustained return to normal responses would still leave the original cause unexplained.
As an affiliate, we earn on qualifying purchases.
Key Questions
What happened to Grok?
Some Grok Lite users received paragraphs of disconnected words instead of answers to ordinary requests. The behavior was reported on Grok.com.
Was every Grok user affected?
No widespread outage was confirmed. TechCrunch could not reproduce the problem and said it probably affected a small subset of users, but xAI released no figures.
Did the problem affect Grok on X?
The available reporting said the glitch appeared limited to direct Grok.com queries. Grok’s X account appeared unaffected during the incident.
Has xAI fixed the gibberish problem?
A permanent fix was not publicly confirmed. Grok’s account called the behavior temporary and recommended starting a fresh chat or regenerating the answer, but xAI did not explain the cause.
Source: xAI
Source: xAI
