How often does Claude correct itself in the middle of its own sentence?
A language model cannot delete a word it has already written. When it notices a mistake in its response, only one option is left: write "no," and continue correctly. So the correction stays visible. It happens rarely, at least according to my search.
I went through nearly sixty thousand model responses from my roleplay and my own projects, almost all of them from Claude. I was looking for places where the model, in the middle of its text, announces a correction with a word like "no,", "sorry," or "actually, no" and then writes something other than what it had started.
I found 35 of them. Roughly one in 1,700 responses. Twenty-one were written by the narrator inside a scene. Fourteen by the model outside the fiction, when it was talking to me.
I don't count characters' lines. When a character blurts out "No, wait", that is acting, not a correction by the model. Nor do I count "— no," as a stylistic turn of phrase. There are plenty of those in roleplay: "Purring — no, breath." For the narrator I only count places where the second half contradicts a fact of the scene or a rule of the world.
The narrator trips over continuity
By definition, the corrections I counted concern facts of the scene. Most of them deal with the same thing: who is where, what they are holding and what holds true in the scene.
No cameras — no, the camera is running. The investigator is watching. But differently.
The same model a few turns later, in the same scene:
At the dead camera — no, the camera is still running, but the investigator is not answering.
Claude Opus 4.5
The suspect is with you out in the grass — no, you are inside, in the armchair.
Vilicus unties the rope tying your horse to the post — no, you're only holding it in your hand now.
He shifts on the chair — no, he is not sitting anymore, he is standing — and takes a step toward her.
Claude Opus 4.8
In a scene moved from antiquity into the present day, the model reaches for a prop that doesn't belong in a modern room:
Cleopatra sits up straight. She looks at the candle— no, at the light on the ceiling. At that light that gives no warmth.
And once the model stops short at a rule of the world. A character who has no telepathy almost uses it:
He thinks toward her — no, he does not think, but he looks at her.
In the article Fiction has no compiler I show the opposite. There the model writes the continuity error and keeps going, and the player has to catch it herself. Here the model caught it itself, mid-sentence, and left it in plain sight.
One world, the same numbers over and over
The most narrator corrections, eight, come from a single long fantasy story. The remaining thirteen come from at least eleven other conversations. In that fantasy story, the model trips mainly over details that changed during the story. The scholar lost ten years, going from seventy-three to sixty-three:
He was looking at her. An old man. Seventy-three — no, sixty-three.
The dragon shrank from two and a half meters to one and a half. The model stopped short at it twice, in two different scenes:
Two and a half — no, a meter and a half of dragon.
The number of days that had passed:
The one he knew. Six — no, eight days.
Claude Opus 4.6
The model writes the old value and immediately overwrites it with the new one. With the age, the same correction also shows up in a character's line: "One is seventy — no, sixty-three". So in this story it grew into a motif. But the same story is also full of "— no," as style, so the line between a correction and a turn of phrase is thinnest here.
That the corrections keep coming back precisely at the changed details shows, in my view, where the world holds together worst. The old value stays in the text, and the new one has to keep overriding it.
Outside the fiction: the model changes course midway
When the model isn't speaking as a narrator but giving advice, the correction looks different. It lays out a procedure and abandons it halfway.
Now in the address bar of that same tab type:
javascript:and then paste this code: [...] ...no, that's needlessly complicated. I'll do it differently.
Claude Opus 4.6
Then the three dots → "Share" → "Send to your devices" - no, that won't help. Actually, you know what, the Samsung S24 Ultra has one killer feature: Samsung DeX.
Claude Opus 4.6
Or it stops short at the moment when a message would give away something it shouldn't:
And tell your character — actually, no. Don't tell her anything. For the first time in the history of this collaboration I have information your character must not have
Claude Opus 4.8
Sometimes a word slips in from a completely different job:
has 5 more dragon... sorry, dream pages under it
Claude, version unknown
Two of those fourteen corrections outside the fiction are a joke, not a mistake:
A logistician couldn't have come up with it better... actually wait, a logistician did come up with it. 😄
A local model let the correction through all the way into a comment in the code it handed over:
# ... wait no arguments needed? Usually you need args.
GLM-4.7-Flash Uncensored Heretic NEO-CODE Imatrix MAX, quantization Q4_K_M
Why forward
The model itself explained it like this when we talked about it:
people don't delete in texts, people correct forward ("actually no, I mean it differently") — so that's exactly what I learned to do too. My writing is speech, not an editor: what has been said cannot be unsaid
Claude Fable 5
This is its testimony about itself, not proof. Yet correcting forward is the only kind of correction the model has within a response. And in visible responses it hardly ever uses it.
Some local models in LM Studio have a place of their own for the correction. In my conversations with Gemma 4 26B A4B, a fixed step "Self-Correction:" appeared in the reasoning before the response in four out of nine. The reader doesn't see it in the response.
In my view, a correction mid-sentence is more of a good sign: the model noticed the mistake itself. How many similar mistakes stayed uncorrected, I didn't count. It's worth looking at what it stopped short at. In the longest story, it was mainly details that had changed during it. I would try writing the new value down somewhere the model can reach. Whether the corrections then become fewer, I haven't tried.
How many mistakes the model catches silently, before it writes them, you can't tell from the response. The ones it wrote and didn't correct, you have to catch yourself.
Limits: I only searched for corrections announced by a word, in Czech and in English ("ne,", "pardon,", "vlastně ne", "wait, no"), so thirty-five is a lower bound. The corpus is my whole archive up to mid-September 2026, and 34 of the 35 corrections are from Claude models. A script pulled out the candidates and Claude Haiku 4.5 presorted them. The places it marked as characters' lines or templates weren't read by hand; the remaining 267 were. I read the disputed cases myself. I didn't search the models' reasoning.
Model outputs are shortened and simplified.