Grok started replying in nonsense. What happens when a model breaks

Grok spent 20 August 2026 answering some users in fluent nonsense. One asked it to generate a PDF and got back "match it without and your they and two for planets can practical and often cheese", with, as TechCrunch put it, "similar nonsense continuing for several paragraphs". Affected users told TechCrunch they were on Grok Lite and had noticed the problem as early as the morning of Wednesday 19 August. The Grok account on X answered unhappy users on the Thursday morning, the same day the report went up: it described the problem as a rare temporary generation glitch, pointed to a status page showing no incidents, and recommended starting a fresh chat or regenerating the answer.

We carry two Grok models, Grok 4.3 and Grok 4.3 Reasoning, so we are reporting this one rather than enjoying it. It is a useful incident because it was documented in public while it was happening, and because it is only one of the ways an assistant goes wrong. There are five worth knowing, and they call for different responses. Each one below gets the same four lines: what you see, what is actually happening, what to do, and a verdict.

First, the move that fixes the most cases

The advice in xAI's own reply, start a fresh chat or regenerate, is the right first move, and the reason it works is worth a paragraph because it holds for every assistant.

A conversation is not held inside the model. Every time you send a message, the thread up to that point is sent again as context. So a mangled reply sitting in your history is part of what goes to the model on your next turn, and a thread that has started producing nonsense has a way of carrying on. TechCrunch recorded exactly that pattern in August: "Refreshing the session often restores normal function, but some users reported that the gibberish responses continued even after multiple refreshes."

In MultiChats there are two buttons under a reply, and on the website they do different things. Re-generate replaces the reply where it stands. The old one leaves the thread and is never replayed to the model afterwards, so the bad text is gone from the context as well as from your screen. Try Again carries a branch icon and opens a new thread from that point, copying the history up to the message you acted on and leaving the broken reply behind in the original conversation.

The phone apps use the same two moves under swapped names, which is worth knowing before you tap. On mobile, Try again replaces the reply in place, and Branch here is the one that opens the new thread.

Both menus also let you send the retry to a different model. Switching model inside a thread keeps the conversation where it is: the history stays, and the next answer comes from somewhere else. The picker holds 60 models from 18 providers, and there is no side-by-side view, so switching is the move.

Word salad, which is what August was

What you see. Fluent-looking rubbish, or text that reads normally until you check it. In the August case one user who followed the source links on a reply found, in TechCrunch's words, "a string of links to reinforcement learning research sites".

What is actually happening. Nobody outside xAI knows, and nothing on the public record points at anything you typed. What is on that record: TechCrunch could not reproduce the fault and judged it "likely affecting only a small subset of users"; xAI did not respond to its request for comment; and the Grok account called it "a rare temporary generation glitch". No cause was published. The other detail worth keeping is where it landed. "The bug appears to be limited to direct queries on Grok.com. The Grok account on X.com has been unaffected," TechCrunch reported, and xAI's own consumer FAQ explains why those two are separable: "While xAI provides the Grok in X service on X.com and X apps, it does not have operational oversight of X's service." One brand, two surfaces, run by two organisations, which is the sort of wiring that stays invisible until a glitch traces its outline.

What to do. Regenerate first. If the next reply is also broken, start a fresh thread, which gives the model a clean context. If both fail, switch model in the thread and carry on. The thing not to do is rewrite your prompt, because on this failure mode the prompt was never the problem.

Verdict. Not your fault, not worth debugging, and usually fixed by a button.

A web page that gave the model instructions

What you see. You ask the assistant to summarise a page, and it does something you did not ask for, such as attaching details about you to a link it produces. The reply itself may read perfectly well.

What is actually happening. Indirect prompt injection: instructions hidden in content the model reads, which it then follows as if they came from you. On the same day as the gibberish, The Register reported a version of this against Grok's web chat agent, found by researchers at Adversa AI and named cryptographic context injection. The instructions on the page are encrypted, with the key sitting next to them, so the guardrail scanner passes the page through and the model does the decryption itself. Adversa's lead researcher Rony Utevsky put the last step plainly: "the model runs that decryption inside its own code execution sandbox." In the proof of concept the payload exfiltrated a Grok.com conversation, sending "the user's name, coarse location, subscription tier, and the full set of the user's prompts in the conversation by appending them to a URL as parameters". The Register reports that xAI was told on 3 June 2026, directly and through HackerOne, acknowledged the report without giving a mitigation timeline, and that the technique still worked as of 19 August 2026.

What to do. Treat "summarise this page" on a page you do not trust as handing that page a turn in your conversation, and keep anything you would not paste into a stranger's browser out of the same thread. Since this failure mode is about data rather than output quality, one thing is worth saying plainly about ours: we do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker. Both Grok entries carry one, reading in full: "xAI may use your conversations to train Grok models." Most entries carry no notice at all, so a silent card is not a statement either way.

Verdict. The failure that does not look like one. Regenerating will not help, because the reply was produced correctly from a context that had been tampered with.

Plain confident invention

What you see. A specific, well-formed, wrong answer. A citation to a paper with a real author and a title that does not exist, a date off by a year, a quotation nobody said.

What is actually happening. The same process that produces the right answers. The model extends text along the patterns it has learned, and nothing inside it switches over when the output stops being true, which is why an invented answer arrives in the same confident voice as a correct one. Our explainer on why models invent facts goes through the common types and what actually reduces them.

What to do. Check the claim itself, against something you opened yourself. Regenerating gives you a second sample and not a verification, and two samples agreeing tells you the pattern is strong.

Verdict. The only entry here that never announces itself, so it is the one worth a habit.

The model you picked has been retired or swapped

What you see. The model button reads a name you did not choose, or a model you liked has gone from the picker.

What is actually happening. A retirement, run to a schedule the app shows you. A model marked for retirement names its successor in its own entry, so choosing it hands you the replacement, and the entry drops out of the picker once its discontinuation date has passed.

What to do. Look for the amber "Deprecated" pill in the picker, whose tooltip names the recommended replacement and, where one is set, the date the model goes. Then open the replacement's details and confirm it still does the specific thing you used the old model for. Nothing in your history stops working.

Verdict. The least dramatic failure here, and the only one where the app has already made the decision for you.

The provider is having a bad hour

What you see. Nothing arrives, sending again produces the same nothing, and then one short sentence naming the model you chose.

What is actually happening. A problem at the provider, reported on the provider's own terms. xAI declares incidents on status.x.ai, and declaring an incident is a deliberate act that follows the first failures rather than arriving with them, so a page reporting nothing wrong and a chat that will not answer are not in contradiction. Our Grok status page reads that feed, shows the current state and lists the incidents xAI has recently resolved, which is the context a single green dot does not give you.

What to do. Switch model in the thread first and check status pages second, because the switch takes one tap and answers your actual question. Both Grok entries sit behind the same provider, so the useful moves are the other providers.

Verdict. The one failure mode where waiting is a legitimate strategy, and the one where nothing you do in the thread will help.

The order to run through

Regenerate, then a fresh thread, then a different model, then a status page. That sequence takes under a minute and it disposes of the word salad and the bad hour at the provider, while the retirement disposes of itself, because the app hands you the replacement whether you act or not.

Two of the five ignore all of it. A page that slipped instructions into your context produced its reply correctly, so a second attempt reproduces it. An invented fact has not broken in any way the app can detect, so nothing in the menu applies. Both of those need a reader rather than a retry.