First, these data show that it is not the case that the correction of misinformation is indistinguishable from misinforming people at the expense of previously held correct beliefs. In all five experiments, participants hypercorrected when they had made an error and were given correct feedback to remember. In none of the five experiments did participants hyper'correct' when they had originally been correct and were given false feedback to remember. These results are the first to show empirically that individuals respond differentially, regardless of confidence, as a function of whether they were initially correct or incorrect. They also suggest that people may be able to use second-order knowledge concerning their own metacognitive confidence judgments (Buratti et al., 2013; Dunlosky, Serra, Matvey, & Rawson, 2005).
Second, although these data support the idea that people have some knowledge that differentiates correct and incorrect responses—and results in differential processing of correct responses overwriting errors than for misinformation overwriting correct answers, despite being equally confident in the correctness of these correct and incorrect responses initially—the pattern of the differences was only slightly more exaggerated and clear when people were provided with explicit knowledge of whether they were initially correct or incorrect, and whether the feedback they received in response to their answers was true or false. In the experiments in which participants were given this explicit knowledge (Experiments 2 and 5), compared to those in which such knowledge was not provided (Experiments 1, 3, and 4), similar trends emerged.
Third, when participants had made errors initially, not only did they hypercorrect when given correct answers, but they also hyper'corrected' when they were given false feedback. The fact that participants did not distinguish true and false feedback following an error is remarkable. When people were wrong with high confidence, they were particularly vulnerable to further misinformation.
The locus of differences in memory for true or false feedback
Evaluation of the feedback
An obvious source of the differential processing of feedback following correct answers and errors may be that participants are able to evaluate the truth of the feedback that they were given, and they use this evaluation, along with the additional information that the feedback brings to mind, to bias updating in memory. That telling participants whether the feedback was true or false made almost no difference to their processing suggests that they may have already made this evaluation on their own.
A rapid online feature-matching model illustrates this idea, in which memory is updated as a function of the participants’ initial response—correct or incorrect—and the feedback given—true or false (Fig. 5). Here, bars indicate the accuracy of relevant potential features about answers to the question “What is the largest state that’s east of the Mississippi River?” that a participant may use to update memory. In Fig. 5(a), the participant incorrectly responds “Minnesota,” a large state that borders the Mississippi River. However, possible disconfirming information is not considered. The feedback “Georgia” makes this participant realize that Minnesota is too far west. Crucially, this discriminating information (see Gigerenzer, Hoffrage, & Kleinbölting, 1991; Koriat, 2012) is triggered by the feedback, but was not considered during the initial evaluation of their answer to the question. Realizing that they were wrong, they hypercorrect this true feedback.
In Fig. 5(b), the participant is confident that Georgia is one of the biggest states east of the Mississippi, and also they have coded that Georgia is east of the Mississippi. When they receive the false feedback “Texas,” they know immediately that this cannot be a correct response because Texas is west of Louisiana, which borders the Mississippi to the east, and so they do not hyper'correct' to the false feedback.
The last case, illustrated in Fig. 5(c), is potentially the most interesting, from the perspective of understanding why misinformation is adopted. Here, the participant responds “Minnesota” and receives the false feedback “Texas.” If this person was unsure of where the Mississippi River is geographically (which in this hypothetical example is clear, as their initial response was Minnesota, a state that is west of the Mississippi), but does know that Texas is the largest state in the continental U.S., they may hyper'correct'. They realize that their high-confidence response may be wrong and they get no information from the feedback that is disconfirming, and so they are willing to accept the feedback regardless of whether it is true or false and shift their belief towards this false feedback. This interpretation is supported by the fact that in all three of the experiments (3–5) in which false feedback was provided to incorrect responses, participants showed a hyper'correction' effect.
The speed of this postulated evaluative process and the ensuing enhanced processing (for a correct answer following an error) or the inhibition of the enhanced processing (for erroneous feedback following a correct response) is striking. Event-related potential (ERP) evidence indicates that an exaggerated P3a component related to the feedback to high-confidence errors occurs roughly 400 ms post-feedback (Metcalfe et al., 2015). Given this, it is likely that people have some additional knowledge—above and beyond their stated confidence—that they bring to bear even before the feedback is presented.
Pre-feedback knowledge
There is considerable evidence that people may have access to other knowledge that is related to their confidence judgments. A number of studies have indicated that a particularly dense semantic landscape is related to highly confident erroneous responses (Eich et al., 2013; Metcalfe & Finn, 2012, 2013; Metcalfe & Miele, 2014; Sitzman & Rhodes, 2010; Sitzman et al., 2015) and that the density of knowledge may facilitate further learning. Butterfield and Mangels (2003) and Butterfield and Metcalfe (2006) provide evidence that people are more familiar with high- than low-confidence questions, as well as with the responses, both correct and incorrect, associated with them—a result which has been borne out by LSA of the relation of the errors to the correct answers (Eich et al., 2012; Metcalfe & Finn, 2012; Metcalfe & Miele, 2014). Metcalfe and Finn (2012, 2013) showed that once people are provided with the correct answer following a highly confident compared to a low-confidence error, they frequently say they “knew it all along.” Thus, even without feedback, participants are more likely to produce the correct answer on a second guess, to choose it on a multiple-choice test, and to guess it when given clues. Given that confidence is related to the density of the semantic landscape, and that the density of the semantic landscape facilitates learning, it is possible that the density of the semantic landscape is responsible for the hypercorrection effect. This view is bolstered by the fact that hyper'correction', in the present experiments, was observed when people were wrong with high confidence and incorrect answers were provided to them to learn. However, presumably the density of the semantic neighborhood increases with confidence in correct (Tunney, 2010) as well as incorrect answers. Thus, this underlying knowledge density factor should facilitate both the updating of correct as well as incorrect responses. However, in the present studies, a hyper'correction' effect was not observed for correct answers, suggesting that some other factor is involved.
It is, of course, possible that participants were disingenuous about their initial confidence ratings, or that there was no equivalency between ratings made for incorrect versus correct responses at the time that the rating was made. In other words, a participant who rates their confidence as 90% and gives the correct response might not mean the same thing as when they give a 90% confidence rating and gives an incorrect response. While this may seem, at first, unlikely, given that confidence ratings were made before feedback was given, this possibility is not entirely implausible.
In other paradigms, people sometimes appear to have knowledge that is not captured by simple metacognitive ratings. For example, when people are in a tip-of-the-tongue state they typically express very high feeling-of-knowing ratings (see, Schwartz & Brown, 2014). However, they also sometimes know—and can report if asked—when they are experiencing a blocker that they are certain is wrong (see, Kornell & Metcalfe, 2006). Similarly, Lindsay and Johnson (1989) showed that although people will confidently affirm that they saw a misleading event in one context, if they are pressed more directly about the context of the event, they can reveal that they know that the experience did not occur in the queried context after all. They based their original judgment upon confirmatory fluency information, but can further scrutinize their memory for disconfirming information that alters the simpler assessment. Son (2010) observed that even when participants indicated that they were 100% confident that they had fully learned a particular item, when asked a subsequent question of whether they wanted to restudy the item, they would often opt to do so. Presumably, an item that was fully learned with absolute certainty would not need to be restudied. Thus, in this case, the metacognitive measure indicated that they believed they knew the answer, but their subsequent choices indicated that they knew they did not.
Furthermore, several studies have shown that asking a metacognitive question in different ways can produce different results. Koriat and Bjork (2005) showed that when people were asked to give their confidence ratings about how likely it was that they would remember an item, their subjective expected probability of recall was greater than when the experimenters reframed the question to enquire about how likely it was that they would forget the item. Presumably—while technically asking for the same information—the question about remembering evoked confirmatory evidence while the forgetting question provoked disconfirming evidence, producing different results. Finn (2015) showed that this difference in judgments had consequences. It played out in study choices, wherein people chose to study more in the latter than in the former case. Perhaps, if subjects in the present study had been asked how confident they were that their responses had been false rather than true, the results would have been different. People may have more knowledge about the truth of their answers, even prior to receiving feedback, than a single simple query about confidence taps.
A number of studies have asked people to make responses and confidence judgments multiple times to the same question (Koriat, 2012). These studies follow from the finding that if several people with different viewpoints are asked a question, the answer that the group as a whole produces is more closely aligned with the truth than are the answers from individual participants (Galton, 1907; Wallsten, Budescu, Erev, & Diederich, 1997; Yaniv & Milyavsky, 2007), a phenomenon called the “wisdom of crowds” (Hertwig, 2012). Similarly, if a single individual is tasked with retrieving and making corresponding confidence judgments several times, rather than just once, the answer obtained by combining the different retrieval events can be more accurate than any of the single retrieval events, as Herzog and Hertwig (2014) have demonstrated. Fraundorf and Benjamin (2014) estimated that the advantage gained from the inner crowd is about 1/10 that of using different people, and the effects can sometimes be small (Ariely et al., 2000). Even so, the inner crowd findings demonstrate that people may have untapped knowledge about the accuracy of their retrievals. Additional research is needed to determine whether such additional knowledge might be responsible for the confidence-related differences we observed in the experiments presented here.
Finally, perhaps people’s confidence judgments were systematically inaccurate. Perhaps it really is the case that 90% confidence in one context (for example, after making an error) is not the same as 90% confidence in a different context (for example, after a making a correct response). It is possible that people sometimes impulsively and incorrectly express high confidence, and they are later able to spot when they did so. This possibility could be checked by asking subjects to make metacognitive judgments about the accuracy of their confidence ratings. We were able to find only four studies in which people were asked to make such second-order metacognitive judgments—judgments about their original metacognitive judgments (Buratti et al., 2013; Buratti & Allwood, 2012a, b; Dunlosky et al., 2005). One of these, though—the study by Buratti et al. (2013)—is directly relevant to the present results. In this study, people answered semantic memory questions and gave their confidence about their answers. Later, participants were asked if they wanted to change any of their confidence ratings. Very few ratings were changed. Indeed, so few were altered that the overall accuracy of the relation between confidence and accuracy was unaffected. Even so, when the authors investigated which particular ratings were changed, the results were systematic. The confidence ratings associated with errors that had been assigned high-confidence ratings and those associated with correct answers that had been assigned low-confidence ratings were chosen. The changes people made to these ratings were in the direction of greater metacognitive accuracy. These findings favor the idea that people do have some access to the truth of their responses above and beyond what their initial confidence ratings indicate.
It would appear that all of these lines of research bolster the idea that people have some knowledge beyond what their initial metacognitive ratings indicate. It is notable, though, that in all the above cases, people had to be explicitly asked to make the finer-grained judgment or to reassess what they had said before. Only then did the additional knowledge surface. In our experiments, people were never asked to make such a reassessment. If this additional knowledge is responsible for the differences between the patterns of updating following responses that were correct and those that were incorrect, it emerged spontaneously in the experiments described here. Whether this second-order metacognitive knowledge is sufficient to account for our findings, then, remains an empirical question.
In the present experiments, people answered factual questions and were asked, in a straightforward manner, to provide confidence judgments indicating how sure they were that their answers were true. The data presented here provide evidence that people either have access to or gain access to knowledge beyond their original confidence judgments about the correctness of their own retrieval. The present experiments cannot definitively answer whether this knowledge was gained through evaluation of the feedback that was given in response to the initial answers, or whether it was available before the feedback was provided. When this additional information about the truth of their answers emerges, and how it may be put to use, are questions for further empirical research. However, confidence has different memorial consequences, depending on whether the answer was true or false and whether the feedback was true or false. It appears, from the research presented here, that meta-metaknowledge of the truth of an individual’s own answers—knowledge that is not expressed in first-order confidence judgments—differentially impacts cognitive processing and memory.