On the interplay of curiosity, confidence, and importance in knowing information

The information-gap theory (Loewenstein, 1994) sees the roots of curiosity in the gap between already-known information about a topic and the knowledge level one aspires to. Loewenstein proposed that curiosity is like the feeling of hunger, but for knowledge: a small “bite of knowledge” increases the hunger for more information, but after gaining more and more information, the hunger is satiated, and thus, curiosity decreases. Inspired by this theory, Kang et al. (2009) examined curiosity as a function of confidence in knowing information. They hypothesized that if people are moderately confident in knowing information, they will be most curious about it. In contrast, if they know nothing, then they cannot be curious. And if they feel to know all about something, they will feel not curious about it anymore. To investigate this hypothesis, Kang et al. (2009) used a set of 40 trivia questions regarding knowledge of the information to examine the information-gap theory.² The questions were shown to cover different curiosity levels during pretesting. Specifically, their experiments followed the same basic procedure for all 40 questions: (1) participants were presented with a trivia question and were instructed to guess the corresponding answer in their head; (2) participants rated their curiosity on a scale from 1 to 7; and (3) participants rated their confidence concerning their guessed answer on a scale from 0 to 100%. The results of their first experiment mimicked the hunger metaphor of the information-gap theory for curiosity, with an inverted U-shaped relationship between curiosity and confidence ratings, with little curiosity at low and high confidence ratings and maximum curiosity ratings at medium confidence levels.

In addition to these self-rating measurements, Kang et al. (2009) hypothesized that situations in which curiosity about knowing something is high, should be associated with a rewarding effect for the learner, as more—presumably important—information is gained during these phases. Thus, they hypothesized that curiosity would be associated with a higher probability to spend resources to learn new information to close information gaps. They tested this hypothesis with another experiment, which was similar to the first one but during which participants could spend time (or tokens) for information and thereby indicating their curiosity through their behavior. In this experiment, the first phase of the experiment was the same as in the first experiment. After the presentation of the 40 trivia questions, each question was presented again and participants typed in their initial guesses. Then, they chose whether they wanted to see the answer. One set of participants had the opportunity to wait for revealing the answer (time condition) and another set of participants had the opportunity to pay for revealing the answer with tokens (token condition). If participants were unwilling to wait or pay for the answer, they were able to choose to continue directly with the next question without seeing the right answer. With this setup, the authors sought to investigate whether participants would trade off gaining information against the cost of staying longer in the experiment or spending tokens. In line with their hypothesis, results revealed that participants were more willing to spend time or tokens with increasing curiosity.

In sum, Kang et al. (2009) provided evidence for the link between curiosity and confidence as an inverted U-shaped function following the information-gap theory by Loewenstein (1994). They also showed that the self-ratings on curiosity were indeed a valid measure, as participants were more willing to spend resources to learn information when being curious about it. However, the information-gap theory is also limited by the idea that individuals or animals can only be curious about the information that is already partly known. In terms of the hunger metaphor, a first bite is needed to elicit an appetite for knowledge. Thus, if no prior knowledge exists, individuals or animals will not be curious. Yet, being curious about something completely new has been addressed by other theories described in the following section.

Besides the information-gap theory from Loewenstein (1994), other observations suggest that the novelty of information has a vital effect on curiosity and that people gain intrinsic rewards and satisfaction when they learn about something new. In contrast to the information-gap theory, the novelty theory suggests that curiosity is highest for novel information. Evidence on this account comes from early studies on curiosity from Berlyne (1950, 1966). Berlyne found that animals, but also humans, are more curious about new stimuli (Berlyne, 1950, 1966). Similar observations were made by Smock and Holt (1962) who showed that toddlers play longer with new toys compared to already-known toys (Smock & Holt, 1962). One issue of such novelty-based accounts is that they must assume that learning about novel stimuli is always beneficial, which is not necessarily the case (Gottlieb et al., 2013). Furthermore, novelty theories cannot explain why people have familiarity preferences or sometimes even avoid novel information which may be useful to them (Golman et al., 2017; Kidd & Hayden, 2015; Loewenstein, 1994). For example, many people, who are already in medical care, annually omit the opportunity to be tested for HIV, even if tests are for free and not associated with any additional effort (Sweeny et al., 2010; Tao et al., 1999).

While evidence exists in favor of the information-gap theory and in favor of the novelty theory, Dubey and Griffiths (2020) recently argued that these two dominant theories are not mutually exclusive and combined both theories within one theoretical computational framework. They suggest that a key factor that determines the level of curiosity is the value of knowing information. This value, however, depends on the probability that this information will (or will not) occur again in the future. If the information will occur again, it is valuable knowing it. Otherwise, it is not valuable to know it. Dubey and Griffiths (2020) tested this prediction, using a similar paradigm and the same stimulus material as in the study by Kang et al. (2009) in an online study that had three phases. Phase 1 was based on the core procedure from Kang et al. (2009)—participants were shown a question, rated their curiosity (0–7), and their confidence in knowing the right answer to that question (1–100%) for all 40 questions in a row. In Phase 2 each question was presented again, and participants could choose to reveal the correct answer in exchange for waiting 10 s (with the decision to wait as an indicator of curiosity) or otherwise continue directly with the next question.³ In Phase 3, 10 out of the 40 questions were shown again and the participants had to type in the respective answer/their guess (they received a small reward per right answer within a set time frame to prevent online research on this question).

To test their hypotheses about whether the function of curiosity and confidence further interacts with whether the information will, or will not, occur again in the close future, and thus is more or less useful to maximize rewards, participants were assigned to two conditions. These two conditions differed in how the 10 questions were sampled in Phase 3. In the uniform condition, each question was equally likely to occur in Phase 3. In the confidence condition, however, the probability of the 10 questions occurring in Phase 3 depended on participants' confidence ratings in Phase 1, with a higher confidence rating leading to a higher probability of occurrence in Phase 3. Importantly, participants were told about each respective sampling process before Phase 2. As participants were explicitly instructed before Phase 2 about the sampling procedure in Phase 3, Dubey and Griffiths (2020) hypothesized that participants in the confidence condition should be more curious about information in Phase 2 they were moderately confident in knowing the answer, reflected with an inverted U-shaped function between curiosity and confidence (as these questions were more likely to occur again in Phase 3, where they will be able to gain rewards for a correct answer). In addition, participants in the confidence condition would be less curious about the information they were either confident in knowing, as they already knew the answer to these questions, or not confident at all in knowing the answer, as these questions were unlikely to occur again in Phase 3, and thus were expected with fewer rewards. However, participants from the uniform condition (where information was equally likely to occur again in Phase 3) were supposed to be most curious about unknown information (low confidence ratings in Phase 1), reflected in a negative relationship between curiosity and confidence, as learning the answer about these questions would help to maximize rewards in the close future (i.e., in Phase 3).

The results were in line with the predictions from their rational computational model. Results from Phase 1 first showed that curiosity followed an inverted U-shaped function for both groups of participants (as in Experiment 1 from Kang et al., 2009). Results for Phase 2 differed between groups: participants from the confidence group were most curious about information when their confidence was moderate in knowing the answer. Participants from the uniform group were most curious about information when their confidence was low in knowing the answer.

Together, these results support the prediction that curiosity about information depends on the value of knowing this information. This critical aspect allows for explaining the findings from Dubey and Griffiths (2020) in the view of perceived importance of information. Specifically, participants from the confidence condition were explicitly told that information they were more confident about was more important in the close future (because of the higher likelihood to appear in Phase 3 in which knowing the correct answer/information would be rewarded). This indicated that information on moderate confidence levels was important to maximize rewards, as this information was likely to occur again in Phase 3. On the other hand, participants from the uniform condition were told that information from all confidence levels was equally important to maximize rewards in Phase 3. Thus, information with the lowest confidence levels was most important, as the largest information gaps existed here, and closing these gaps would lead to maximal rewards.⁴

In sum, these results suggest that people should be most curious about information gaps, which are perceived to be important. In other words, irrespective of whether confidence in information is low or moderate, as long as these information gaps are perceived as important, they should induce curiosity. Critically, the results from Dubey and Griffiths (2020) indicate that people can, and do, estimate the value of knowing a particular piece of information, even if their confidence in knowing the answer is low. This implies that people have—at least in some situations—a metacognitive estimation of the importance of knowing information, even if they hardly know anything about this information other than that it is important.⁵ This further implies that asking people on whether knowing information would be important or not, in addition to asking them about their confidence in knowing the answer, may be a critical predictor for curiosity about knowing this information and consequently, their willingness to close this information gap. Here, we sought to explicitly test the influence of importance on the relationship between curiosity and confidence.

This study comprised two experiments. For each experiment, we conducted a two-step analysis approach to (1) replicate the findings by Kang et al. (2009) and (2) extend these findings by investigating whether the importance of knowing information interplays with the relationship between curiosity and confidence. In the following, we first describe the analyses conducted for Experiment 1 followed by a description of analyses for Experiment 2.

In the first step of analyses for Experiment 1, we sought to replicate the main behavioral findings by Kang et al. (2009). For this replication, the same stimulus material as in Kang et al. (2009), was applied (Experiment 1). Specifically, in the first phase of Experiment 1, participants were asked trivia questions and subsequently rated these questions on a curiosity scale (1–7) and a confidence scale (0–100%). In the second phase, participants were asked if they would like to wait for 5–25 s to reveal the answer to the question or if they would like to just skip the answer to proceed with the experiment.

We carried out two analyses also reported by Kang et al. (2009). First, we investigated curiosity as a quadratic function of confidence and expected curiosity to follow an inverted U-shaped function of confidence, reaching its maximum when confidence is moderate (see Fig. 1C in Kang et al., 2009). Second, we investigated participants’ probability to reveal an answer as a function of curiosity⁶ (see Fig. 5 in Kang et al., 2009), and expected more willingness to spend effort (in terms of time) to reveal an answer on high curiosity levels⁷ (see Fig. 5 in Kang et al., 2009).

In the second step of analyses, we sought to extend these findings by additionally investigating the effect of the importance of knowing information on curiosity and participants’ probability to reveal an answer. Therefore, we directly asked participants to indicate the importance of knowing the answer to each question in this study. We thus added an importance scale (1–7) to the first phase of both experiments, respectively. Since the trivia questions elicited different levels of curiosity in the study by Kang et al. (2009), we reasoned that the perceived importance of information may also vary.

We then tested whether the addition of the importance variable fitted the data better, compared to the model which replicated results by Kang et al. (2009). We also explored the goodness of fit of several other models with curiosity as the dependent variable with each other (see Table 1). We predicted that the importance of knowing information significantly interplays with curiosity. Therefore, we expected that adding this variable should lead to an increased model fit, compared to the models reported in the first step of our analyses.

In particular, the rational model of Dubey and Griffiths (2020) theoretically predicts that both curiosity theories can be integrated within one model when assessing the value of information. Here, we assessed the value of information by asking participants about the subjective importance of knowing information. We thus predicted that for information perceived as important, curiosity should be highest for low-to-moderate confidence in knowing this information (see red curve in Fig. 1), as closing these information gaps may be associated with high future rewards. However, with decreasing importance of knowing information, we expected this curve to drop down and shift toward an inverted U-shaped function of curiosity and confidence (see blue curve in Fig. 1), as low confidence in knowing information and perceiving this information as not important may actually indicate that people know nothing about this information at all and are thus not curious about it. But if people would already know a bit of information, they may be hungry about knowing all of it—as suggested by the information-gap theory—even if the information would be perceived as less important.

In another analysis, we explored the interplay of (1) the importance of knowing information, (2) curiosity, and (3) confidence in participants’ probability to reveal an answer (see Table 2). Therefore, we carried out several plausible combinations of models to explore whether the addition of an importance term increased the model fit to predict participants’ probability to reveal an answer. Note that we assessed all three independent variables as the analyses reported by Kang et al. (2009) and Dubey and Griffiths (2020) differed with respect to the regressor variable when predicting participants’ probability to reveal an answer. Kang et al. predicted the probability to reveal an answer with curiosity (see Fig. 5 in Kang et al., 2009), while Dubey and Griffiths predicted the probability to reveal an answer with confidence for two different groups (see Fig. 6B in Dubey & Griffiths, 2020). We report the results of the model with the best fit.

Experiment 2 considered new stimulus material to test whether the results obtained in Experiment 1 generalize to other stimulus material. Therefore, we exposed participants to a set of COVID-19-related questions (Experiment 2). Except for this change in stimulus material, the setup of Experiment 2 was the same as in Experiment 1. The analyses followed the same procedure as for Experiment 1. We first sought to replicate the findings by Kang et al. (2009) and then sought to extend these findings by considering the importance of knowing information in two model comparison analyses (see Tables 1, 2). We describe the results of the winning model.

The same inclusion criteria were applied as in Dubey and Griffiths (2020). All participants who chose to reveal the answers to all 20 questions and those who never chose to reveal the answer to any questions were removed. Based on these criteria, one participant was removed. Another participant was additionally excluded as this participant provided the same answer to all questions. The final data set of Experiment 1 thus consisted of a total of 41 participants.

As in Experiment 1, we first removed all participants who chose to reveal all answers and no answers. The final experimental data consisted of a total of 38 participants in Experiment 2. The results are reported in Tables 5 and 6.