What Errors Can and Cannot Teach Us

What the claim really is

We think the useful version of "learn from your mistakes" is narrower than the slogan. Errors help under some conditions and do little under others. A 2022 review in Educational Psychologist, by two University of Georgia researchers, describes two phases. In the first, a learner generates an error, which activates related prior knowledge. In the second, the learner detects and corrects it by comparing the response with reference information and explaining the difference. The authors add that the semantic relatedness between the error and the target content strongly affects whether memory benefits occur.

What the experiments show

The best known result is the hypercorrection effect: errors made with high confidence are more likely to be corrected after feedback than errors made with low confidence. In a 2012 study from Columbia University, 51 younger adults (mean age 20.7) and 31 older adults (mean age 65.7) answered general-information questions, rated their confidence, received the correct answers and were retested. The effect was reliable in younger adults (gamma = .51) and smaller in older adults (gamma = .14), not significantly different from zero. Overall correction accuracy was about the same in both groups, roughly .76 to .80, so older adults corrected about as many errors as younger ones. Confidence simply did not predict which ones.

Surprise helps feedback stick, but it does not check that feedback is true. In five experiments published in 2019, confident errors followed by true feedback showed hypercorrection every time. In experiments 3 to 5, confident errors followed by false feedback showed it too, which the authors interpret as susceptibility to misinformation after a confident error.

A related line of work asks whether a failed attempt before study helps. A 2012 study at Purdue University had participants guess the target of a word pair before seeing it. For related pairs, recall was .73 with a pretest and .56 without. For unrelated pairs it was .24 either way. When an entire block separated the guess from the study, the benefit vanished (.55 against .56), compared with .70 when the two were immediate. A 2012 paper by researchers at three US institutions found that failed tests followed by immediate feedback produced better retention than study alone, while delayed feedback generally did not, unless the final test was reversed.

A worked example

The first example is illustrative, not real data. Suppose a learner answers 20 questions and gets 12 wrong. Six of the errors were made with high confidence and six with low confidence. After feedback, the learner corrects 5 of the 6 high-confidence errors and 3 of the 6 low-confidence errors. The correction rate is 5 / 6, or 83 percent, for high-confidence errors and 3 / 6, or 50 percent, for low-confidence errors. The simple hypercorrection effect is the difference, 33 percentage points. Overall, 8 of 12 errors were corrected, or 67 percent.

The second example uses the real recall figures from the Purdue study above. For related pairs, the pretest benefit is 0.73 - 0.56 = 0.17, or 17 percentage points. Relative to the no-pretest baseline, that is 0.17 / 0.56, about a 30 percent increase. For unrelated pairs the benefit is 0.24 - 0.24 = 0. When the guess came a block before study, it was 0.55 - 0.56 = -0.01, which is no benefit. With immediate study it was 0.70 - 0.56 = 0.14, or 25 percent. The same activity gives a sizeable benefit, none, or a negligible one, depending on relatedness and timing.

What the evidence does not show

Most of this work is laboratory work with word pairs or general-knowledge questions, so we should be careful about what it says for a course. The US Department of Education's 2007 practice guide rates pre-questions as low-level evidence, noting that it comes mostly from reading studies with college students. The Georgia review says evidence on motivation and emotion is limited, as is research on developmental differences and the timing of feedback.

Errors are also not a shortcut. In the first experiment of that 2012 paper, overall recall was about .71 for items studied twice and .66 for items tested then studied, and with spaced practice the study-only items did better. The common mistake is to conclude that making errors is good in itself. The evidence supports errors that are related to the material, followed by timely, correct feedback. The practice guide makes the practical point that correct-answer feedback should be provided, because without it errors on one test can persist and reappear on later tests.

Going deeper

Three simple measures sit behind the figures above.

1. Correction rate = Errors corrected on retest / Errors made.
2. Hypercorrection effect (simple form) = Correction rate for high-confidence errors - Correction rate for low-confidence errors.
3. Pretest benefit = Recall with pretest - Recall without pretest. Relative benefit = Pretest benefit / Recall without pretest.

The papers themselves report gamma, a correlation between confidence and correction, where zero means no relationship. That is why .51 and .14 are not percentages and cannot be read as the difference in the worked example. When you open a study of this kind, check four things: the sample size, whether feedback was immediate or delayed, whether the final test used the same cue or a reversed one, and whether the errors were related to the target. The last three each changed the results in the papers we read. We may have misread them, and we welcome scrutiny of every figure and link here.