Being corrected without being crushed

Ask your partner to stop you and make you say it again, rather than to repeat your sentence back correctly. In the largest classroom study of correction, that gentle restatement was 55% of all corrections and produced a self-correction only 18% of the time. Being prompted to fix it yourself worked far better.

Updated September 26, 2026 · 8 min read

"Just correct me whenever" is the wrong request

It is the request everybody makes, and it is well meant. The problem is that it does not tell your partner which of six quite different things to do, and the one they will default to is the one that mostly does nothing.

In 1997 Roy Lyster and Leila Ranta transcribed 18.3 hours of classroom interaction across four French immersion classrooms and coded every single error, every correction, and what the learner did next. Their six-way taxonomy is still the one the field uses, and their central result is uncomfortable: the most popular correction is close to the least effective.

The six things "correcting you" can mean

Type	What your partner does	Responding to “I go to Busan yesterday”
Recast	Repeats your sentence back, fixed, and carries on	“Oh, you went to Busan yesterday? Nice.”
Explicit correction	States the correct form and says it is a correction	“You should say went, not go.”
Clarification request	Signals that something did not parse	“Sorry — what?”
Metalinguistic feedback	Comments on the form without giving the answer	“That’s the present tense.”
Elicitation	Stops and pulls the correct form out of youThe one with the best numbers.	“Yesterday you…?”
Repetition	Repeats your error back, usually with the stress on it	“You go to Busan yesterday?”

Lyster & Ranta's taxonomy, with a worked example of each responding to the same error. Only the last four require you to do anything, which turns out to be the whole point.

What teachers do, against what works

Here is the finding, drawn from Lyster and Ranta's own tables. The left bar is how often each type was used. The right bar is how often that type actually produced a repair — the learner correcting themselves.

Share of all corrections givenOf those, share that produced a repair

Recastthe polite restatement
55%18%
Elicitationstopping and pulling it out of you
14%46%
Clarification request“sorry — what?”
11%28%
Metalinguistic feedbacknaming the problem, not the answer
8%45%
Explicit correction“say X, not Y”
7%36%
Repetitionyour error, echoed back
5%31%

Two percentages with different denominators, both on a 0–100 scale: the first is each type's share of the 618 corrections observed, the second is the proportion of that type which led the learner to produce the correct form. Recasts dominate and under-deliver; elicitation is the inverse. From Lyster & Ranta (1997), four French immersion classrooms, 18.3 hours of transcript.

Recasts were 55% of all corrections and produced repair 18% of the time. Sixty-nine per cent of recasts produced no learner response at all — the conversation simply moved on. Elicitation was used a quarter as often and worked two and a half times as well.

The reason is not mysterious. A recast is indistinguishable from ordinary conversational agreement. When somebody says "Oh, you went to Busan yesterday?" they might be correcting you or they might just be interested, and in a friendly conversation you will assume the second. Nothing about it requires you to notice, and mostly you do not.

The funnel, which is worse than any single number

Lyster and Ranta also followed errors all the way through, and the attrition is severe. Of all the errors learners produced, 62% got any feedback at all. Of that feedback, just over half produced any learner response. Only 27% of feedback turns produced a repair.

From the total number of errors, 17% eventually led to a repair. Five in six errors, in a classroom explicitly designed to teach the language, went by uncorrected in any way the learner acted on.

100%
Errors the learners produced
The starting population. Every ungrammatical or non-target utterance in 18.3 hours of transcript.
62%
Got any feedback at all
The other 38% were followed by a topic continuation — the teacher or another student simply carried on. Not noticed, or noticed and let go.
55%
Of that feedback, produced any learner response
Just over half. The rest sank without trace, and recasts are heavily represented here: 69% of them produced nothing.
17%
Of all errors, ended in a repair
The learner producing the correct form themselves. One error in six. This is the number that should change what you ask for.

The attrition from error to repair, as Lyster & Ranta measured it. Each step is a place a correction stops being one.

The meta-analysis agrees, and adds a caveat about you

A single study is a single setting. Thirteen years later Lyster and Saito pooled fifteen classroom studies covering 827 learners and asked whether any of this holds up. It does: corrective feedback had "significant and durable effects on target language development," and "the effects were larger for prompts than recasts."

"Prompts" is their umbrella for the negotiation-of-form group — clarification requests, repetition, elicitation, metalinguistic clues. The same four that outperformed recasts on uptake in 1997 outperform them on actual learning outcomes in the pooled data.

The caveat is one an article aimed at adults has to print: the same meta-analysis found "effects for age, with younger learners benefiting from CF more than older learners." Most of the people reading this are the older learners.

How to ask for it, in one sentence

The ask that follows from all of this is specific enough to say out loud at the start of a call, and short enough that your partner will actually remember it.

Just correct me whenever I make a mistake!
Leaves the choice open, and the default is the recast — 55% of corrections, 18% repair.
When I get something wrong, don't say it for me — just stop me and let me try again.
That is elicitation. Used 14% of the time in the study; produced repair 46% of the time.
OK. So — yesterday you…?
One unfinished sentence. That is the whole technique.

The second version tells your partner which of six behaviours you want, and picks the one with the numbers behind it.

And one more line, for the thing that actually destroys conversations: say when. Correction during a story kills the story. The workable split is errors that break understanding get stopped immediately, and everything else waits for a gap.

You said something wrong. Your partner has to decide what to do.

They genuinely did not understand you: Stop now. Clarification request: “Sorry — what?”This is not correction, it is repair of the conversation. It has to happen immediately or the next minute is built on a misunderstanding.
It is the same mistake you have made three times tonight: Stop now. Elicitation: repeat up to the error and wait.A repeated error is a habit forming. This is the case where interrupting earns its cost.
You are mid-story and they understood fine: Let it go. Note it, and raise it at the end.Interrupting a story to fix a particle trades a minute of fluency for a correction that is 18% likely to land.
It is a slip you would catch yourself: Let it go entirely.Correcting everything is how a conversation becomes a lesson, and how a partner becomes a teacher neither of you asked for.

A decision your partner is making several times a minute, usually without a rule. Giving them one is more useful than giving them permission.

What this does not settle

These numbers are from a classroom, not a language exchange. Four French immersion classrooms of children, with trained teachers, in the 1990s. The taxonomy transfers cleanly — the six things are the six things — but the percentages describe that setting. Nobody has run the equivalent study on adults on a video call, and we are not going to pretend otherwise.

Recasts are not useless. They are unobtrusive, they keep a conversation moving, and there is a real literature arguing they work better for some structures and some learners than the uptake figures suggest. What the data supports is that they are over-used relative to how often they land, not that they should never happen.

Correction is not the main ingredient. Lyster and Saito found real and durable effects, but a conversation where you were corrected forty times and spoke for four minutes is worse than one where you were corrected twice and spoke for forty. The correction literature is about a seasoning, not the meal.

And your partner is not a teacher. Elicitation is the cheapest of the six for an untrained person to do — it is one unfinished sentence — which is precisely why it is the one to ask for. Asking for metalinguistic feedback means asking them to explain their own grammar, which most native speakers cannot do and should not be made to feel bad about.

Sources

Lyster, R., & Ranta, L. (1997) Corrective Feedback and Learner Uptake: Negotiation of Form in Communicative Classrooms. Studies in Second Language Acquisition, 19(1), 37–66.
The source of the taxonomy and of every percentage in the chart and the funnel. 18.3 hours of transcript from four French immersion classrooms. Note for anyone chasing the free PDF: its running head misprints the volume as 20 — Cambridge's own record is volume 19, issue 1, March 1997.
Checked 2026-09-27
Lyster, R., & Saito, K. (2010) Oral Feedback in Classroom SLA: A Meta-Analysis. Studies in Second Language Acquisition, 32(2), 265–302.
Fifteen classroom studies, 827 learners. Found corrective feedback had significant and durable effects, larger for prompts than recasts, and larger for younger learners than older ones.
Checked 2026-09-27

Start my free session

Your first session is free, up to 30 minutes.