Remembering the Old Estimate After Learning the Answer

Learning an answer and accurately remembering what you thought before learning it are different achievements. First we offer an estimate, then receive new information. Later, when asked for our original estimate, present knowledge can enter the reconstruction of the past. We may remember having been closer to the answer than we actually were. Kosuke Kaida and Naoko Kaida's 2023 study examined this shift by varying the amount of information participants had to learn. Its question concerned the separation of new answers from old estimates in memory, rather than whether participants were honest people. [1]

Asking for the earlier response, not the answer

The researchers used numerical questions to investigate hindsight bias. Participants supplied initial responses, learned the correct information provided by the researchers, and eventually tried to recall their own earlier responses. The final request was not for the newly learned answer. As a purely explanatory example, imagine an initial estimate of twenty, a supplied answer of sixty and a later recollection of thirty. These are not a participant's recorded values. The remembered estimate has moved toward the new information. Learning the answer has not preserved an accurate memory of the earlier estimate.

That distinction also prevents every altered response from being called a movement toward hindsight. If the estimate in the example were remembered as ten, it would have moved away from the supplied answer. The study separated distortions toward the correct information from distortions in the opposite direction. A changed number did not automatically count as a change toward greater apparent prior knowledge. Nor was this a test of actual foresight. Measuring how an estimate is recalled once an answer is known cannot establish that the answer was genuinely predictable before it was supplied. [1]

Why the shared twenty questions mattered

The study involved sixty-three participants. Thirty-one were assigned to a low-load condition containing twenty questions, and thirty-two to a high-load condition containing fifty. Both groups had mean ages of approximately twenty-three. The longer list included the same twenty questions used in the shorter list plus thirty additional questions. The main comparison used responses to the twenty shared items. This was intended to reduce the possibility that an apparent load effect merely reflected one group answering a different collection of questions. The overlap made it possible to examine common material within different amounts of surrounding material.

Memory load was defined as the number of correct answers to be learned. It was not a simultaneous dual task in which participants held digits in mind while reading a story. The longer list should therefore not be described as a direct measurement of working-memory capacity. An individual's capacity and the total quantity of information assigned are separate concepts. Here, the experimental contrast concerned amounts of learning material followed by recollection of earlier responses. Sharing twenty questions made that comparison concrete without turning the task into a measurement of every aspect of memory. [1]

Repeating information until it was learned

Participants arrived in the morning and answered verbally presented questions with numerical estimates. At this point they were not informed that they would later be asked to recall those initial estimates. Correct information was then supplied immediately and participants were asked to learn it. Items that could not yet be answered correctly were repeated with the correct information; successfully learned items were not repeated. The procedure continued until participants could provide the correct information for all questions. Exposure to a single presentation of an answer was therefore not the endpoint of the learning stage.

The authors described learning as taking approximately ten minutes in the low-load condition and twenty-five in the high-load condition. A three-and-a-half-hour retention interval followed, including lunch and simple computer games. Participants then recalled their own original responses. This sequence placed time between learning the new answers and retrieving the old estimates. But learning more material also involved more study time and repetition. The findings should not be presented as if every aspect of the experience apart from item count had been held perfectly identical. That qualification matters when interpreting what the longer-list condition changed. [1]

How many responses changed, and by how much?

Two measures addressed different aspects of distortion. A magnitude index related the change from the original estimate to the distance between that estimate and the supplied answer. Another measure counted the proportion of items with altered recollections. These measures cannot substitute for each other. Many estimates can shift slightly, while a smaller number can shift substantially. Knowing the count of altered answers does not fully specify their displacement, and an average displacement does not fully specify how often it occurred. Separating frequency and magnitude gives the result more structure than a single label of poor memory.

The published results reported greater hindsight-bias magnitude and, overall, more distorted responses in the high-load condition. Shifts toward the supplied information were more prominent than opposite-direction shifts. However, the comparison of opposite-direction distorted-item proportions was not supported consistently across analyses: the parametric and nonparametric results differed. It would therefore be inaccurate to describe every submeasure as showing an equally secure load effect. The principal reported pattern can be communicated while preserving uncertainty in the less consistent contrast. [1]

The paper also describes excluding extreme index values from parametric analysis. Observations with absolute values beyond one hundred per cent were removed, and nonparametric checks using the full data were reported. In addition, the mean learning-repetition count for the low-load group is not given consistently between the results and discussion. This article does not select one conflicting value as definitively verified, or reproduce unusually reported test statistics as independently recalculated findings. Reporting the qualitative comparison does not require treating every numerical detail in the publication as settled. The underlying supplementary data have not been reanalysed here.

One experiment, several possible processes

The authors discussed weakened memory for the original response, confusion with the supplied information and reconstruction using currently available knowledge. These processes could interact. The experiment did not directly isolate one mechanism from all the others, and it did not separately measure working memory. Attributing the high-load group's distortion to a specific inhibitory deficiency in every participant would exceed what was measured. The paper's pilot status and sample-size limitation also matter for judging where similar results might occur. They invite additional tests rather than a universal rule about the amount of information a person can safely learn.

The result is not an argument that acquiring correct knowledge is itself a mistake. New knowledge is useful. But evaluating what we knew previously requires distinguishing its source from the earlier response. In this experiment, the condition involving more answers to learn was associated with greater reported distortion when original estimates were recalled. That does not establish a safe item count for every occupation or a tested intervention for preventing professional errors. It identifies a task-level contrast and a problem worth investigating beyond the laboratory, not a guarantee of practical protection.

The most grounded question is simple: am I supplying what I now know, or remembering what I actually answered then? Those responses can look similar while drawing on different information. Recording an original estimate before feedback makes the distinction observable in a way that confidence alone cannot. The study illustrates why learning outcomes and retrospective self-assessment deserve separate measurements. A person can correctly learn an answer and still misremember the distance between that answer and their earlier knowledge. Keeping both records visible makes it possible to examine that change rather than merely feeling that the answer was obvious all along.

Sources

[1] Kaida, K., and Kaida, N. (2023). Memory load of information encoded amplifies the magnitude of hindsight bias. PLOS ONE, 18(4), e0283969. CC BY open-access primary pilot research. Methods, results, analysis limitations and discussion directly read October 14, 2026. Supplementary data not reanalysed; numerical reporting inconsistencies noted above. Independently written prose; no quotations or images used.