Does Getting Corrected Help You Learn a Language? What 35 Studies Found
Updated: 1 day ago
Many people learning a language are quietly afraid of being corrected. It feels discouraging, and there is a popular idea that mistakes fix themselves if you just keep talking. Researchers in second language acquisition have argued about this for decades. One prominent paper, John Truscott's, claimed grammar correction in writing does not help. So it is a fair question with real evidence on both sides, and enough studies now exist to pool them.
The dataset: a 35-study meta-analysis
In 2020, Lim See Chen and Willy Renandya published Efficacy of Written Corrective Feedback in Writing Instruction: A Meta-Analysis in the open-access journal TESL-EJ. They gathered 38 studies of learners receiving written corrections on their second-language writing, removed 3 outliers, and analysed the remaining 35. Of those, 23 also tested learners again later with a delayed posttest, which is the number that matters if you care about whether a correction sticks.
The headline results, reported as Hedges's g (a standardised effect size where about 0.2 is small, 0.5 medium and 0.8 large by the usual Cohen benchmarks):
Overall effect: g = 0.59, 95 percent confidence interval 0.42 to 0.76
On delayed posttests (23 studies): g = 0.57, confidence interval 0.34 to 0.80
Direct corrections (the correct form is given): g = 0.76 across 16 studies
Indirect corrections (the error is flagged, not fixed): g = 0.63 across 7 studies
Focused feedback (a few error types): g = 0.63 across 16 studies
Unfocused feedback (everything marked): g = 0.45 across 13 studies
The authors stress that the direct versus indirect and focused versus unfocused differences were not statistically significant. They point in a direction; they do not settle it.
The same paper's appendix lists earlier meta-analyses for comparison, and the picture is consistent. Shaofeng Li's 2010 meta-analysis in Language Learning, covering 33 studies of mostly oral feedback, is listed at 0.61 (fixed effects) and 0.64 (random effects). Kang and Han's 2015 review of 24 written-feedback studies is listed at 0.54. Truscott's 2007 analysis, the skeptical one, found an effect of -0.155 in controlled studies, which is the main dissenting number.
The computation: what g = 0.59 means for one learner
Effect sizes are abstract, so here is a translation. An effect size expresses how far the average corrected learner moved, measured in standard deviations of the comparison group. If scores are roughly normally distributed, you can turn that into a percentile.
Take the median learner in the group that was not corrected. By definition they sit at the 50th percentile.
Shift them up by 0.59 standard deviations.
The standard normal distribution puts 72.2 percent of scores below +0.59.
So the average learner who received corrections ended up around the 72nd percentile of the uncorrected group. Using Li's 0.61, the figure is about the 73rd percentile.
Now the durability check. The delayed effect, 0.57, divided by the immediate overall effect, 0.59, is 0.964. In plain terms, about 96 percent of the effect was still there on the delayed tests. These are different subsets of studies, so it is a rough comparison, but it argues against the idea that corrections help only for a day.
Three practical lessons
Corrections help, moderately. Across 35 studies, getting told what was wrong beats not being told. The effect is not magic. It is real and it is medium-sized, which by the standards of education research is a good result.
Giving the right answer is at least as good as just flagging the error. Direct correction scored a bit higher than indirect in this review. For a learner practising alone, without a teacher to come back to, seeing the corrected sentence is the practical choice.
Fewer fixes may beat marking everything. Focused feedback scored higher than unfocused. The difference was not significant, so it is a hint, but it fits common sense: a reply covered in red ink teaches less than one clear point you can remember.
The caveats matter. These studies looked at learners in writing instruction, with teachers or researchers giving the feedback, not at people practising on their phones. And an effect averaged across studies says little about how any one person will respond.
A daily correction, by text
Language Textapp is built around those three lessons. Each day it texts you one short message in the language you are learning, pitched at your level. You reply in that language, and it sends back three things: your sentence corrected, one or two fixes explained in plain English, and a reply that keeps the conversation going. That is direct and focused feedback, every day.
App: ¿Qué comiste hoy? (What did you eat today?)
You: Yo comí tacos en la almuerzo
App: ✏️ Yo comí tacos en el almuerzo. Use 'el almuerzo' (masculine). ¡Qué rico! ¿Te gustaron? (Did you like them?)
You can also ask it how you are doing, and it tells you your days practised, how many replies you have sent and your most common slips, which turns scattered corrections into a short list of patterns to watch. You can ask what its last message meant, or how to say something you are stuck on.
It is an AI, not a person or a teacher, and it can get things wrong, including in its corrections. If a fix looks off, ask it to explain. There is no audio, so this is practice in reading and writing, not speaking.
Try it by text
Text Hi to (385) 525-5598. It asks which language you are learning and whether you are a beginner, intermediate or advanced, then sends your first message. Reply in the language and you get your first correction. After that, one message arrives each day at 9am, or at whatever time you ask for.
It works on any phone that can text, with no app and no account. Your first 20 messages are free, then $9.99/month covers every Textapp. Text STOP at any time. Message and data rates may apply. More at textapp.co/apps/language.
Comments