Correcting a Remembered Outcome: Inference Errors and Feedback

Imagine reading that someone took an umbrella out of a bag. Later, remembering that the person opened it may feel natural, although the original sentence did not say so. This is an illustration constructed for this article, not a stimulus from the research. The important distinction is between understanding a plausible continuation and remembering that the continuation was actually presented. An inference useful for comprehension can become an incorrect answer when the question concerns the earlier text.

A 2022 study by María J. Maraver and colleagues investigated how such inference errors could be corrected. The question was not simply whether showing the right answer again helped. Did retrieving one's own erroneous answer change the benefit obtained from the correct information presented afterward? Two experiments examined this issue with different tasks while separately manipulating whether corrective feedback was available. [1]

Returning to the original sentence

The study used Portuguese sentences suggesting an outcome without explicitly stating it. Experiment 1's material was selected from sentences that frequently produced inference errors in preliminary testing. Its error frequency was therefore not an estimate of how often all everyday conversations are misremembered. The researchers created a set particularly suited to examining correction because errors were likely to occur. [1]

A correct answer did not necessarily reproduce every original word. Synonyms preserving the original meaning could count as correct. Reporting the merely suggested outcome as stated content was coded as a pragmatic-inference error. Other intrusions and omitted responses formed separate categories. This prevented all unsuccessful answers from being treated as one undifferentiated kind of failure. [1]

Experiment 1 involved 120 University of Lisbon students randomly assigned to four conditions combining retrieval or restudy with feedback or no feedback. Everyone initially read correct sentences. A distractor activity preceded the intermediate phase where the main comparison occurred. This was not guessing an answer before encountering the material: the task concerned retrieving information previously presented. [1]

Giving an answer versus reading one

In retrieval, the critical part of a sentence was missing and participants completed it using the initial presentation. In restudy, they read sentences described as another person's recall output. Half preserved the initial information; half contained the anticipated inference error. The comparison was therefore not erroneous retrieval against rereading an entirely accurate text. Both conditions involved exposure to errors, through different routes. [1]

Participants receiving feedback saw the original correct sentence after each response or restudied item. Those without feedback completed simple arithmetic during the corresponding interval. At the final test, everyone supplied missing sentence information without further feedback. The design could thus examine the contribution of the intermediate activity and of subsequently making the correct content available. [1]

Feedback following retrieval produced a larger benefit for correcting inference errors than feedback following restudy. This should not be summarized as errors being intrinsically helpful. Without corrective feedback, erroneous retrieval could also perpetuate false inferences. The useful arrangement involved more than producing an error: correct original information was available afterward. [1]

What the first comparison left open

The proportion of errors generated during retrieval varied, whereas the proportion of erroneous sentences presented during restudy was fixed. Errors also came from the participant in one condition and from displayed text in the other. Since the original information and feedback were externally presented too, restudy could make the sources harder to distinguish. These design differences limited how the first result could be interpreted. [1]

Experiment 2 introduced a more controlled comparison. Of 128 online recruits, eight were excluded because of interruptions, low reported attention, or missing yoked partners, leaving 120 for analysis. Conditions combined active or passive recognition with feedback or no feedback. Materials included critical inference sentences and fillers. [1]

An active participant judged whether a sentence matched the information initially presented. A passive partner saw the same sentences together with that active participant's answers, without making the recognition decision. The content and responses were therefore yoked across individuals rather than merely approximating the amount of correct and incorrect information. This was not a live discussion between paired learners. [1]

The improved control did not mean every aspect of the two experiences was identical. Making a decision and reading someone else's decision remain different activities. The comparison addressed specific problems of the first experiment while retaining the central distinction between actively retrieving and passively encountering the relevant information.

Which measure showed the benefit?

Again, corrective feedback had a larger benefit for reducing pragmatic-inference errors after active recognition. Not every response category showed the same pattern. Other intrusions and omissions did not exhibit the corresponding general advantage. Persistence of initially correct answers was also calculated separately from correction of initially erroneous answers. Combining these outcomes into one percentage would lose important information. [1]

For the error-correction index in particular, the overall main effect of active versus passive recognition was not significant when feedback conditions were considered together. The significant interaction concerned feedback's differing contribution across task types. It is therefore more accurate to describe a conditional benefit involving corrective feedback than to claim that active decisions always corrected more errors. The proportion of initial errors subsequently corrected is also not the percentage of the entire sentence list learned. [1]

These distinctions matter for the apparent message of the experiment. One can acknowledge that actively confronting a memory can create a useful context for correction without assuming that the act of answering guarantees improvement. The informational event following the answer remains essential to the interpretation.

What a short-term correction establishes

Feedback was provided immediately, but the experiments did not compare alternative feedback delays. Confidence in erroneous responses was not measured either. The study consequently did not directly demonstrate that highly confident errors were easier to correct or that immediate correction was better than every other possible timing. Those would require additional comparisons. [1]

The final tests assessed short-term learning. They did not establish weeks-long retention, treatment for memory disorders, or a reliable way to rewrite real eyewitness recollections. The proposal that retrieval makes memories more open to updating is a theoretical interpretation consistent with the behavioral findings, not a biological reconsolidation mechanism directly measured in these experiments. [1]

The useful lesson is to examine the conditions of correction without romanticizing mistakes. Producing an inference as one's own answer and reading another person's answer need not lead to the same response to subsequent correct information. Yet revealing a wrong answer is insufficient when the correct content is unavailable. The study's strongest contribution is treating error and feedback as a sequence of learning events, rather than attributing the outcome to either one in isolation.

Sources

[1] Maraver, M. J., Lapa, A., Garcia-Marques, L., Carneiro, P., and Raposo, A. (2022). Can we learn from errors? Retrieval facilitates the correction of false memories for pragmatic inferences. PLOS ONE, 17(8), e0272427. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0272427 . DOI: 10.1371/journal.pone.0272427. Independently written explanation of research published under a Creative Commons Attribution license.