이미지: TechCrunch AI
Summary
- Grok Lite users have been reporting responses made up of meaningless word strings since at least Wednesday morning
- xAI addressed the issue via its official X account Thursday morning, calling it a temporary generation error and advising users to start a new conversation or regenerate the response
- The glitch surfaced as xAI continues rolling out new models following a mass exodus of founding staff and its absorption into SpaceX
- 발생 시점
- 늦어도 수요일 오전부터 이용자 보고
- 영향 이용자
- 무료 등급 그록 라이트(Grok Lite) 이용자 다수
- 증상
- 의미 없는 단어 나열, 강화학습 연구 사이트로 연결되는 출처 링크 오류
- 영향 범위
- Grok.com 직접 질의에 한정, X 내 그록 계정은 정상 작동
- xAI 공식 답변
- 목요일 오전 X 계정에서 "일시적 생성 오류"라고 설명
- 임시 해결책
- 새 대화 시작 또는 답변 재생성, 다만 일부는 반복돼도 지속됐다고 보고
- 조직 배경
- 창립 멤버 대부분·연구원·엔지니어 최소 50명 이탈(더 인포메이션 5월 보도)
"An answer mixing cheese and planets"
One user asked Grok to create a PDF file. What came back read: "match it without and your they and two for planets can practical and often cheese…" It wasn't even a sentence — just paragraphs of disconnected words strung together. Another user opened the source links attached to a response and found a string of links pointing to reinforcement-learning research sites.
Grok, the chatbot built by xAI, has been producing this kind of nonsensical output for several days now. Users who spoke with TechCrunch were on the free Grok Lite tier and said they noticed the problem starting no later than Wednesday morning. TechCrunch said it was unable to reproduce the issue in its own testing, suggesting the bug may be limited to a subset of users.
Community already in an uproar
The Grok subreddit has seen a string of similar complaints. One user posted asking how to fix the issue, while another asked for help after Grok started spewing gibberish. Refreshing the session generally restored normal behavior, though some reported that the gibberish persisted even after multiple refreshes.
The issue appeared only when querying Grok directly through the Grok website (Grok.com). Grok accounts invoked within X (formerly Twitter) were unaffected.
xAI: "Temporary generation error"
As complaints piled up, the official Grok account posted a response on X Thursday morning. It explained that the "completely garbled responses are a rare, temporary generation error." The status page at status.x.ai showed all Grok services operating normally, and the account advised that starting a new conversation or regenerating the response would generally resolve the issue immediately. xAI headquarters did not respond to TechCrunch's request for comment.
Glitch comes amid staff exodus
The error comes as xAI continues to undergo organizational upheaval. According to a May report from The Information, xAI has lost most of its founding members along with at least 50 researchers and engineers. xAI was founded by Elon Musk in 2023, acquired X in 2025, and was itself acquired by SpaceX in 2026 — meaning it is no longer an independent company but now operates as SpaceX's AI division.
When xAI released its latest foundation model this past July, it described it as "Opus-level, but faster, more token-efficient, and cheaper." Whether this gibberish bug is related to that model or stems from a previously released version remains unconfirmed.
Editor's View
Chatbots suddenly spitting out unintelligible text is not a rare occurrence in itself. What stands out here is the timing. xAI is churning out new models even as most of its founding staff have left and the company itself has been absorbed into SpaceX. Service reliability taking a back seat during periods of organizational turmoil is a pattern that has repeated across multiple AI companies. Given that even companies with large infrastructure teams occasionally suffer outages, it's not especially surprising that a glitch like this emerged at an xAI that has lost a significant share of its workforce.
Compared with other chatbot companies, Grok's response speed here isn't bad. It's rare for OpenAI or Google to have their official accounts step in directly to explain an outage within hours. Grok's official account addressed the issue not long after it surfaced and even offered a workaround — refreshing the session. Still, the fact that the status page displayed "all systems normal" while some users were receiving gibberish for days suggests that the reliability of an official status page should be viewed separately from what users actually experience.
For teams that have integrated Grok into their workflows, it's worth noting that this issue appeared particularly pronounced on the free Grok Lite tier. If Grok has been plugged into an automated pipeline where output is passed along without human review, adding some form of format validation before and after the response seems like a reasonable safeguard. It seems unlikely that xAI will offer an official explanation of the root cause within the next few weeks — the company has a track record of staying silent even on questions about its organizational restructuring.
Comments