Grok's 'word salad' glitch hits users as xAI says systems are fine

xAI’s Grok chatbot has been spitting out nonsense to a subset of users, and the company’s response has been a shrug. After one user asked the model to generate a PDF, Grok responded with: “match it without and your they and two for planets can practical and often cheese…” — a stream of gibberish that continued for several paragraphs. Another user found that checking source links returned a string of URLs pointing to reinforcement learning research sites, which makes little sense for ordinary queries.

The affected users told TechCrunch they were on Grok Lite and noticed the problems as early as Wednesday morning. TechCrunch couldn’t reproduce the issue in its own testing, and xAI didn’t respond to a request for comment. That silence is telling. The company has seen significant staff turnover in recent months — The Information reported in May that xAI lost most of its founding team and at least 50 researchers and engineers. When a model starts emitting pure word salad, you have to wonder whether institutional knowledge walked out the door with those departures.

The bug appears limited to direct queries on Grok.com. The Grok account on X.com has been unaffected, which narrows the problem to a specific deployment rather than the underlying model. Refreshing sessions sometimes fixes the nonsense, but some users reported that gibberish persisted even after multiple refreshes. The official Grok account eventually acknowledged the issue on X, calling it “a rare temporary generation glitch” and pointing to status.x.ai, which showed all systems operational.

Here’s what’s striking: xAI released its most recent foundation model in July, describing it as “an Opus-class model, but faster, more token-efficient and lower cost.” Shipping a model with that kind of claim and then having users receive paragraphs of random words doesn’t inspire confidence. The Reddit community has been flooded with complaints, and while the company frames this as a minor hiccup, the combination of staff exodus, an unresponsive PR team, and a public glitch suggests deeper turbulence than a simple refresh can fix.