Knowledge Gap Illustrations Spark Curiosity

Luisa Frede; Lisa Bardach; Younes Strittmatter; Eileen Richter; Marie Mueckstein; Markus W. H. Spitzer

doi:10.5334/joc.501

Introduction

Curiosity is often described as an intrinsic motivational drive to seek new information (Berlyne, 1950; Dubey & Griffiths, 2020; Grossnickle, 2016; Gruber et al., 2014; Hsiung et al., 2023; Jach et al., 2024; Kang et al., 2009; Kidd & Hayden, 2015; Kim et al., 2024; Litman, 2005; Loewenstein, 1994; Murayama, 2022; Murayama et al., 2019; van Lieshout et al., 2020; Wilson, 2024; Yagi et al., 2023). Theoretical accounts converge on the idea that curiosity arises when individuals detect a gap in their knowledge, and is satisfied once this knowledge gap is closed (Berlyne, 1950; Dubey & Griffiths, 2020; Loewenstein, 1994). These accounts are supported by empirical work showing that curiosity peaks in the presence of knowledge gaps (Brod & Breitwieser, 2019; Dubey & Griffiths, 2020; Shin & Kim, 2019; Spitzer et al., 2024, 2025; Wade & Kidd, 2019). Critically, however, prior empirical studies typically examined knowledge gaps implicitly—using confidence self-ratings—and thereby eventually increased the salience of those knowledge gaps, albeit to an unknown extent. Thus, it remains unclear whether curiosity can be elicited when knowledge gaps are made explicitly salient, compared to conditions in which no such cues are provided. Here, we investigated whether making knowledge gaps explicitly salient—by illustrating them—can effectively elicit curiosity in humans.

Curiosity theories addressing the role of knowledge gaps

Theoretical accounts of curiosity converge that curiosity is triggered when individuals detect a gap in their knowledge. However, these accounts diverge in their predictions about the optimal magnitude of knowledge gaps for eliciting maximal curiosity. According to the information-gap theory, curiosity arises when people become aware of a discrepancy between their current and their desired knowledge (Loewenstein, 1994). This perceived gap in knowledge creates cognitive tension, motivating people to seek out new information to close the knowledge gap. Importantly, for curiosity to arise, some prior knowledge is necessary to recognize a knowledge gap. When individuals know very little about a topic, they often lack the context to identify what is missing—essentially, they don’t know what they don’t know. This results in a non-linear relationship between the magnitude of the knowledge gap and the intensity of curiosity: when individuals have no relevant knowledge (i.e., the gap is too large), curiosity is not triggered; conversely, when they already possess most of the relevant knowledge (i.e., the gap is too small), curiosity is also not triggered. Thus, the “sweet spot” for eliciting curiosity lies at a moderate knowledge gap—where enough is known to recognize what is missing, but enough is unknown to stimulate curiosity. Once knowledge gaps are closed, people’s curiosity vanishes.

Empirical evidence supporting this theory comes from Kang et al. (2009), who operationalized knowledge gaps by asking participants how confident they were in knowing the answers to questions they posed. In particular, Kang et al. (2009) prompted a set of 40 trivia questions (e.g., “What instrument was invented to sound like a human singing?”; answer: Violin) to participants and asked them to rate both their confidence in knowing the answer and their curiosity about each question (without showing the answer to participants first). In line with the information-gap theory, they observed an inverted U-shaped relationship between confidence and curiosity: when participants indicated low or high confidence in knowing the answer to the prompted question, their curiosity was relatively low. In contrast, participants’ curiosity was highest when their confidence in knowing the answer was moderate. In addition to these questions, Kang et al. (2009) also provided behavioral evidence for the link between curiosity and information-seeking behavior. They demonstrated that higher levels of curiosity increased participants’ subsequent willingness to spend resources (i.e., time or tokens) to satisfy their curiosity by finding out answers. Importantly, these findings have been replicated by several other research groups, substantiating their robustness (Brod & Breitwieser, 2019; Dubey & Griffiths, 2020; Shin & Kim, 2019; Spitzer et al., 2024, 2025; Wade & Kidd, 2019).

In contrast, the novelty theory posits that curiosity is triggered by exposure to entirely new stimuli (Berlyne, 1950, 1966, mostly reflecting large knowledge gaps¹). According to the novelty theory, curiosity drives behavior not in order to reduce an aversive state of not knowing, but rather to explore novel information (Berlyne, 1950). This has been investigated in both animals and humans. For example, children show a preference for novel visual stimuli (Fantz, 1964) and are more likely to engage with or seek out new toys (Cantor & Cantor, 1964; Smock & Holt, 1962); for similar results in rats, see Pisula & Siegel (2005).

However, the assumption that humans should always prefer novel stimuli or experiences has been put into question, as existing empirical evidence showed that, particularly in stressful situations, people choose familiarity over novelty (Litt et al., 2011; Wansink et al., 2003). Moreover, in contexts where acquiring information is uncertain, individuals intentionally avoid acquiring information, even when doing so would be beneficial (Golman & Loewenstein, 2018; Golman et al., 2017; Hertwig & Engel, 2016). For example, even when there is no cost or effort involved, many people choose not to find out their HIV status (Hightow et al., 2003; Sweeny et al., 2010; Tao et al., 1999).

Nevertheless, the information-gap theory and novelty theory have recently been integrated into one rational account of curiosity (Dubey & Griffiths, 2020). Dubey & Griffiths (2020) propose that curiosity depends not only on the gain of information but also on its expected utility. According to their account, people are most curious about stimuli they expect to yield the greatest information gain, but only when they judge that the information could be useful in the future. Crucially, Dubey & Griffiths (2020) substantiated their predictions with empirical evidence showing that the structure of the environment determines whether curiosity followed an inverted U-shaped function of confidence (supporting the information-gap theory) or a negative relationship between curiosity and confidence (supporting the novelty theory). In particular, they found that when the probability of future encounters with a stimulus is positively correlated with prior confidence (i.e., more confidently rated stimuli are expected to appear more often), curiosity tends to peak at moderate levels of confidence—in line with the information-gap theory. This reflects the classic inverted-U pattern: the medium knowledge gap is chosen since it yields the most utility. While low-confidence gaps could also be closed, they yield less utility due to the low probability of the corresponding stimuli reappearing. In contrast, when future encounters are independent of past appraisal (i.e., when fully uncertain information is equally likely to occur and confidence is not related to the probability of occurrence), curiosity increases as confidence decreases—aligning with novelty-based theories. In such contexts, the optimal strategy is to explore the least-known stimuli, as they are expected to yield the greatest information gain.

As such, the rational account proposes that whether curiosity peaks at moderate or low confidence depends on the context and its associated utility. It thereby reconciles the information-gap and the novelty theory as special cases within a broader, utility-maximizing framework. Crucially, this account suggests that both large and moderate knowledge gaps can elicit curiosity, depending on the expected utility of the information (also see Spitzer et al., 2024).

How can we deliberately spark curiosity?

Across theoretical accounts, curiosity is consistently linked to the presence of knowledge gaps. Prior research has estimated such gaps indirectly through confidence self-ratings, which may prompt individuals to reflect on what they do or do not know, thereby increasing the salience of their knowledge gaps—although the extent of this effect remains unclear. This raises the question: can curiosity be sparked by making knowledge gaps explicitly salient by illustrating them?

Recent work by Metcalfe et al. (2023) provides compelling support for this idea. In a series of ten experiments, participants first answered general knowledge questions and rated their confidence in each response. They then received one of several types of feedback, one of which was clear yes/no feedback indicating whether their answer was correct or incorrect. After receiving clear yes/no feedback, participants rated how curious they were to find out the correct answer. A consistent pattern emerged: when participants were highly confident but then learned they were incorrect, their curiosity to find out the correct answer increased significantly. In contrast, when they learned they were correct, curiosity vanished. These findings suggest that externally provided cues about the presence or absence of knowledge gaps strongly influence curiosity.

More traditional forms of feedback such as progress or correctness indicators provide learners with information on how much of the material they have completed or whether their answers are correct. While these feedback mechanisms can enhance engagement and reduce cognitive avoidance tendencies (Devine & Otto, 2022; Hattie & Timperley, 2007), they do not necessarily stimulate curiosity or motivation (Wisniewski et al., 2020). By contrast, feedback specifically designed to indicate gaps in knowledge may be more effective in promoting curiosity. Even though Butler & Winne (1995) did not focus on curiosity, they proposed that feedback is beneficial when it helps learners recognize discrepancies between their current understanding and the correct information, prompting them to actively work to close these gaps. This discrepancy-based mechanism aligns conceptually with the information-gap theory.

Building on this work, we theorize that the explicit salience of knowledge gaps plays a key role in sparking curiosity. Specifically, we propose that illustrating knowledge gaps to individuals increases curiosity, while illustrating that no knowledge gap exists (i.e., full knowledge has been attained) reduces it. This approach extends previous research by moving beyond confidence ratings and toward a more systematic investigation of whether curiosity can be deliberately sparked by illustrating of knowledge gaps.

The present study

In this study, we examined whether making knowledge gaps explicitly salient elicits curiosity. In particular, we examined whether knowledge gap illustrations spark curiosity, reflected by an increased information-seeking behavior, ultimately resulting in increased knowledge gains.

Across two preregistered experiments, both of which were preceded by a pilot experiment with approximately n = 300 participants each (see the Online Supplement), participants first completed a pretest assessing their prior knowledge about elephant conservation in Botswana across six subtopics. Participants answered three single-choice questions about each subtopic. Next, they made a series of decisions about whether to read or skip individual chapters on these subtopics about elephant conservation in Botswana. Each chapter contained the full information to answer all questions about each subtopic. Finally, participants completed a posttest identical to the pretest.

Participants were randomly assigned to one of two groups. In the experimental group, participants received a knowledge gap illustration before each decision—this cue indicated their current knowledge level based on pretest performance for each chapter (i.e,. 0%, 33%, 67%, or 100% prior knowledge). In the control group, we only presented the title of the next chapter and the choice to read or skip were presented, without any knowledge gap illustration (see Figure 1).

A process diagram of the experimental paradigm for the experimental group and control group, respectively. We illustrated knowledge gaps as bars and as a message. We presented a total of four different knowledge bars to participants (i.e., 0%, 33%, 67%, or 100%; only two possibilities presented here).

This experimental design allowed us to assess the effect of knowledge gap illustrations on participants’ probability to read (vs. skip) a chapter (i.e., information-seeking behavior). We further evaluated whether an increased information-seeking behavior was accompanied by increased knowledge gains (i.e., the difference between pretest and posttest scores).

Based on the information-gap theory, we predicted that illustrating moderate knowledge gaps (33% or 67% prior knowledge) would increase the probability to read chapters compared to no illustration. In contrast, illustrating full knowledge (100% prior knowledge) should reduce curiosity, resulting in a lower reading probability than in the control group. This should result in an inverted U-shaped relationship between prior knowledge and the probability to read in the experimental group but rather no relationship between these two variables in the control group.

According to the rational account of curiosity, the expected utility of acquiring information moderates curiosity for large knowledge gaps. Thus, our predictions for 0% prior knowledge differed between Experiment 1 and Experiment 2. In Experiment 1, participants were not told that a posttest would follow, implying relatively low utility for gaining information. Here, we hypothesized that illustrating large knowledge gaps would have no effect on participants decision to read chapters. In Experiment 2, we increased the perceived utility of gaining information by explicitly instructing participants that a posttest would follow. In this context, we hypothesized that illustrating large knowledge gaps would significantly increase the decision to read in the experimental group compared to the control group.

Additionally, we expected that the decision to read chapters should generally lead to knowledge gains across groups (except for the 100% prior knowledge condition, where improvement was not possible).

Finally, we predicted that the experimental group should gain significantly more knowledge than the control group in conditions where they also read significantly more than the control group.

Experiment 1

The aim of Experiment 1 was to test the hypothesis that illustrating moderate knowledge gaps (i.e., 33% and 67% prior knowledge) sparks curiosity, reflected by an increased information-seeking behavior. We therefore expected an inverted U-shaped relationship between prior knowledge and the probability to read in the experimental group but not the control group. We further asked whether such amplified information-seeking behavior was accompanied with increased knowledge gains.

We preregistered Experiment 1 (AsPredicted #218945, preregistered on 03/22/2025, 04:46 AM [PT]) and intended to collect a final sample size of 200 participants (after applying the exclusion criteria). Critically, we preregistered to only include participants showing variability in their reading behavior (i.e., to skip and read at least once; but see the Online Supplementary for the results of the full data set which replicated the major findings of this experiment). The sample size was estimated based on a power analysis based on the pilot experiment. In particular, we fitted the model on the probability to read between groups using the simr package (Green & MacLeod, 2016). We ran 100 simulations and targeted an alpha level of α = .05. The power simulation was based on the quadratic term of the interaction effect of prior knowledge and group on the probability to read from the pilot experiment (an estimate of β = 10.79, see the Online Supplementary), as this interaction indicated an overall differential relationship between prior knowledge and the probability to read between the two groups. This power analysis indicated a power of 99% with a 95% confidence interval (CI) between 94.6% and 100%. We decided to collect a rather conservative sample size of n = 500 participants for Experiment 1, as we expected to exclude about half of the participants (similar to the pilot experiment). As preregistered, we excluded participants who either always chose to read or always chose to skip throughout the experiment.

Method

Participants

We collected data from n = 501 participants (age range = 18–45; mean age = 24.3; 250 females; 251 males) online via Prolific. Participants received US$2 for participating in the experiment. The experiment lasted for 11.9 minutes on average (SD = 8.0). All participants participated voluntarily and gave consent that their anonymous data could be stored and published before the start of the study. Participants were told that they could stop the experiment at any time without providing a reason. Experiment 1 as well as all other experiments were conducted in English.

Stimuli

Experiment 1 consisted of three parts: a pretest, a decision and reading phase, and a posttest. The pretest and posttest each contained a total of 18 single-choice questions, with three questions per chapter and four possible answers per question (for the full article and the 18 questions, see the Online Supplement). The questions were identical across the two tests. The questions covered key elements from the reading material that participants could read between the two tests, and assessed participants’ knowledge on the respective chapters. The reading task consisted of an article comprising six chapters that were similar in length, each of which covered a different subtopic about elephant conservation in Botswana.

Procedure

Figure 1 illustrates the process participants followed during this study. After providing consent to participate in the study, participants took the pretest assessing their baseline knowledge of each of the article’s six chapters. Next, we provided participants with the choice to either read or skip each of six chapters. If participants chose to read, we provided them with the chapter and they could voluntarily decide when to continue the experiment (i.e., self-paced their reading time). If they chose to skip, we presented the next choice to read or skip a different chapter. After participants made their six decisions, they completed the posttest. We randomized the order in which chapters were presented to participants.

Crucially, we randomly assigned participants to one of two groups. We collected the same number of participants for each group². In the experimental group, we presented participants a knowledge bar and a message indicating their current knowledge level (i.e., 0%, 33%, 67%, or 100%) on the specific chapter based on their answers in the pretest (the knowledge gap illustration; see Figure 1). We estimated their current knowledge based on their pretest score. However, we did not directly tell participants how we specifically estimated their knowledge. The instructions were: “Your knowledge on [topic of the following chapter] is at [0%, 33%, 67%, or 100%]. You now have the opportunity to learn more about the topic. Do you want to read this section or skip it?” This was designed to illustrate participants’ knowledge gaps. In the control group, we neither provided participants with a knowledge bar nor a message on their present knowledge. Instead, we presented the following message to the participants after each chapter: “You now have the opportunity to read more on [topic of the following chapter]. Do you want to read this section or skip it?” (see the Online Supplementary for all instructions).

Dependent Variables

We considered probability to read as the dependent variable reflecting participants’ binary choice to read or skip each chapter (coded as 1 = read and 0 = skip).

We measured knowledge gains for each prior knowledge condition (0%, 33%, 67%, or 100%). For this, we considered the score difference between posttest and pretest. Values ranged from 0 to 3 points (three questions and we counted one point for each correct answer), since participants could reach 0, 1, 2 or 3 points in the pretest as well as the posttest for each chapter.

Independent Variables

We considered participants’ prior knowledge as our first independent variable. Prior knowledge was assessed per chapter. Participants could score between 0 and 3 points per chapter (as we asked three questions per chapter) leading to the following four prior knowledge conditions: 0%, 33%, 67%, and 100% prior knowledge. We refer to prior knowledge conditions as large knowledge gaps (0%), moderate knowledge gaps with lower knowledge (33%), moderate knowledge gaps with higher knowledge (67%) and no knowledge gaps/full prior knowledge (100%).

We considered a dichotomous group variable as our second independent variable which indicated whether participants were in the experimental or in the control group (we coded the contrasts with +1 for the experimental group and -1 in the control group).

To examine whether participants’ decision to read affected their knowledge gains, we considered a third independent variable: the decision to read, a dichotomous variable indicating whether participants chose to read (contrast coded as +1) or skip (contrast coded as -1) a chapter. All variables were group mean-centered.

Data Analysis

We conducted the statistical analyses using R (R Core Team, 2021). We applied the tidyverse package (Wickham et al., 2019) and the sjPlot package (Lüdecke, 2024) to visualize the results and the patchwork package (Pedersen, 2024) to group figures. Additionally, we used the lmerTest package (Kuznetsova et al., 2017) to fit hierarchical linear regression models and the lme4 package (Bates et al., 2015) to fit hierarchical logistic regression models. To examine contrasts between groups and within groups for prior knowledge conditions, we utilized the emmeans package (Lenth, 2024).

To estimate the probability to read as a function of prior knowledge and group, we ran a hierarchical logistic regression with participants’ probability to read as the dependent variable. We additionally added a second-order polynomial term for prior knowledge, for allowing a non-linear relationship between prior knowledge and probability to read. We included a random intercept for participants to account for variability in participants’ overall willingness to read chapters. In particular, we applied the following regression model:

probability to read \sim {prior knowledge}^{2} \times group + (1 ∣ participant) .

We expected an inverted U-shaped relationship between prior knowledge and participants’ decision to read or skip the chapter in the experimental group but not the control group. This should be reflected in a significant interaction between prior knowledge and group, especially for the second polynomial term.

We further evaluated pairwise comparisons to assess differences between the two groups for each prior knowledge condition (0%, 33%, 67%, or 100%). We accounted for multiple comparisons using the Tukey adjustment for family-wise error rate. We expected significant differences between the two groups at moderate knowledge gaps (33% prior knowledge and 67% prior knowledge), with a higher probability to read in the experimental group. We further expected a significantly lower probability to read in the experimental group compared to the control group at full prior knowledge (100%) and no differences between groups at large knowledge gaps (0%). We expected no differences between these two variables in the control group.

We substantiated these pairwise comparisons by examining, within each group, the differences in the probability to read across prior knowledge conditions. We ran these pairwise comparisons to more closely examine the shape of each curve for each group. As we expected an inverted U-shaped relationship between prior knowledge and the probability to read in the experimental group, we expected a significant difference between 0% and 33% as well as 33% and 100% prior knowledge and 0% and 67% as well as 67% and 100%. We expected no difference between 33% and 67% as these conditions both represent moderate knowledge gaps.

Next, we investigated whether the decision to read led to knowledge gains. Therefore, we first evaluated whether the decision to read generally led to significantly higher knowledge gains across groups. In particular, we fitted a hierarchical linear regression with score difference as the dependent variable, prior knowledge and the decision to read (read vs. skip) as fixed effects, and a random intercept for participants. We applied the following regression model:

score difference \sim prior knowledge \times decision + (1 ∣ participant) .

We quantified knowledge gain differences for each prior knowledge condition by running pairwise comparisons between the two groups for each prior knowledge condition. We expected significantly higher knowledge gains when participants decided to read compared to when they decided to skip a chapter. However, we only expected this for prior knowledge conditions where participants had knowledge gaps (i.e., 0%, 33%, and 67%; not 100%), since knowledge gains are not possible when having full knowledge.

To assess knowledge gains between groups, we calculated score differences and modeled them using a hierarchical linear regression with prior knowledge and group, including their interaction as fixed effects, and a random intercept for participant. We applied the following regression model:

score difference \sim prior knowledge \times group + (1 ∣ participant) .

Again, we ran pairwise comparisons between the two groups per prior knowledge condition. We expected the experimental group to improve significantly more from pretest to posttest, especially in those cases where the probability to read had been significantly higher.

As an additional analysis, we investigated whether illustrating knowledge gaps also influenced participants’ reading time. We conducted this analysis with a hierarchical logistic regression with reading time as the dependent variable and prior knowledge, group, and their interaction as predictors, and a random intercept for participants. We included a second-order polynomial term to allow for the suspected non-linear relationship between reading time and prior knowledge. Therefore, we applied the following model:

reading time \sim {prior knowledge}^{2} \times group + (1 ∣ participant) .

We also calculated pairwise comparisons using model-based estimated marginal means to compare reading times between groups at each prior knowledge condition. The readers can find the results to this analysis in the Supplementary Material.

Finally, we ran several non-preregistered models to examine whether alternative models could better explain participants’ decision to read. To examine whether previously illustrated knowledge gaps or just the very first presented knowledge gap influenced participants’ decision to read, we ran six alternative models and compared the goodness of fit of these models with the preregistered model presented above. We compared model fits using the Bayesian Information Criterion (BIC). Differences larger than 10 in BIC values between models indicate a better fit of the model with the lower BIC (Burnham & Anderson, 2002; Neath & Cavanaugh, 2012). Table 1 lists all considered models, including their BIC. In particular, we compared the preregistered model reported in our main analysis, i.e. the effect of current-chapter prior knowledge (prior knowledge_n) on the decision to read (Model 1), against an alternative model examining the effect of cumulative mean prior knowledge (prior knowledge_≤n) on the decision to read (Model 2). We conducted this model comparison to examine whether previous prior knowledge better explained the decision to read rather than just current-chapter prior knowledge. Next, we tested the effect of first-chapter prior knowledge (prior knowledge_first) on the decision to read (Model 3) to investigate whether the very first decision to read influenced all following decisions more than the present knowledge gap. Further, we modeled the effect of overall mean prior knowledge (prior knowledge_n̄) on the decision to read (Model 4) and the effects of both prior knowledge_n and prior knowledge_n̄ on the decision to read (Model 5) to examine the effects of participants’ average prior knowledge and their current prior knowledge and whether these models fit the data better than the preregistered model. Lastly, we conducted a median split (mean-centered) separating participants with low prior knowledge and high prior knowledge (prior knowledge_split). We tested the effect of prior knowledge_split on the decision to read (Model 6) and the effects of both prior knowledge_n and prior knowledge_split on the decision to read (Model 7) to examine whether both factors influence participants’ decision to read. All models also contained the effect of group, interaction effects between independent variables and a random intercept for participants. Additionally, we specified all the above-mentioned models again, but included random slopes for prior knowledge. Table 1 lists these models as Model 8 through Model 14. We added random slopes in order to account for within-participant variation in prior knowledge across chapters. However, model comparisons indicated a better fit to our data without the inclusion of random slopes.“

Table 1

Models predicting the probability to read and associated BIC scores of each model for all experiments.

MODEL	FORMULA	EXP. 1	EXP. 2	PILOT EXP. 1	PILOT EXP. 2
Model 1	$P K_{n} \times G + (1 ∣ P)$	1828.7	1674.6	820.8	763.0
Model 2	$P K_{\leq n} \times G + (1 ∣ P)$	1966.9	1864.4	857.3	–
Model 3	$P K_{first} \times G + (1 ∣ P)$	1981.8	1923.4	–	–
Model 4	$P K_{\bar{n}} \times G + (1 ∣ P)$	1991.3	1919.7	–	–
Model 5	$P K_{split} \times G + (1 ∣ P)$	1981.8	1919.2	–	–
Model 6	$P K_{n} \times P K_{\bar{n}} \times G + (1 ∣ P)$	–	–	–	–
Model 7	$P K_{split} \times P K_{n} \times G + (1 ∣ P)$	–	–	–	–
Model 8	$P K_{n} \times G + (PK ∣ P)$	1841.3	1685.5	832.7	774.3
Model 9	$P K_{\leq n} \times G + (PK ∣ P)$	–	1825.8	–	–
Model 10	$P K_{first} \times G + (PK ∣ P)$	1955.4	1855.1	–	–
Model 11	$P K_{\bar{n}} \times G + (PK ∣ P)$	1970.3	1864.1	870.8	–
Model 12	$P K_{split} \times G + (PK ∣ P)$	1959.6	1857.1	857.9	845.7
Model 13	$P K_{n} \times P K_{\bar{n}} \times G + (PK ∣ P)$	–	–	–	–
Model 14	$P K_{split} \times P K_{n} \times G + (PK ∣ P)$	–	–	–	–

[i] Note. Lower BIC scores indicate better model fits. The best-fitting models with respect to BIC are marked in bold font. Models marked with “–” in a given experiment either failed to converge or resulted in a singular fit; their BIC can therefore not be reported. Independent variables are Group (G), Current-Chapter Prior Knowledge (PK_n), Cumulative Mean Prior Knowledge (PK_≤n), First Chapter Prior Knowledge (PK_first), Overall Mean Prior Knowledge ( $P K_{\bar{n}}$ ), Prior Knowledge Group (PK_split), i.e. participants grouped into low prior knowledge and high prior knowledge by performing a median split). Models 1 through 7 contain a random intercept for participants (1|P). Models 8 through 14 additionally contain random slopes for Prior Knowledge per participant (PK|P).

Additional Analyses

We further examined whether there were potential carry-over effects from previous decisions that influenced the decision to read in the current chapter. In particular, we calculated the effects of prior knowledge and group on the probability to read and additionally included a lagged predictor indicating whether the immediately preceding chapter was read. Since this model made the exclusion of the first trial for all participants necessary (since the very first decision had no preceding decision), the number of trials that this model used was not the same as in all other models. We did not compare the BIC of this model with the other models due to the different sample sizes. We report the results in the Supplementary Material.

In addition, we also examined the effects of chapter order on the probability to read. To do this, we ran a model using a median split for chapters. Results for this model are reported in the Supplementary Material.

Transparency and Openness

We documented all aspects of the study in accordance with transparency standards, including criteria for data exclusion, sample size determinants (i.e., power analyses), and all experimental manipulations. The complete analysis code and dataset is located at osf.io/bk9c3. Experiment 1 was preregistered via AsPredicted (AsPredicted #218945, preregistered on 03/22/2025, 04:46 AM [PT]).

Results

The results of Experiment 1 are illustrated in Figure 2a, 2b and 2c. BIC comparisons between all models revealed that our preregistered model explained our data best in each experiment (see Table 1. Table 2 lists the distributions of participants, trials, and average trials per participant for each prior knowledge conditions and group. Note that relatively few participants fell into the 0% prior knowledge conditions. The results of the pairwise comparisons are listed in Table 3 (pairwise contrasts in the probability to read between groups), Table 4 (pairwise contrasts within groups for each pair of prior knowledge conditions), Table 5 (pairwise contrasts for knowledge gains by decision to read), and Table 6 (pairwise contrasts for knowledge gains between groups).

Plots a and d display the probability to read by prior knowledge and group. Plots b and e display the mean score difference by prior knowledge and decision to read. Plots c and f display the mean score difference by prior knowledge and group. Please note that mean score difference refers to the point difference between pretest and posttest scores per chapter. For each chapter, participants could achieve a score between 0 and 3 points.

Table 2

Number of Participants, Trials, and Average Trials per Participant by Group and Prior Knowledge.

GROUP	PRIOR KNOWLEDGE	EXPERIMENTAL			CONTROL
GROUP	PRIOR KNOWLEDGE	n(P)	n(T)	M(T/P)	n(P)	n(T)	M(T/P)
Experiment 1	0%	30	33	1.10	14	17	1.21
	33%	92	150	1.63	74	129	1.74
	67%	125	313	2.50	93	237	2.55
	100%	120	314	2.62	91	217	2.38
Experiment 2	0%	22	29	1.32	23	34	1.48
	33%	82	130	1.59	61	112	1.84
	67%	129	315	2.44	81	196	2.42
	100%	126	342	2.71	80	216	2.70

[i] Note. n(P) = number of participants; n(T) = number of trials; M(T/P) = average trials per participant. Values are reported separately for each group (Experimental, Control) across prior knowledge conditions (0%, 33%, 67%, 100%).

Table 3

Pairwise Contrasts for the Probability to Read Between Groups.

PRIOR KNOWLEDGE	EXPERIMENT 1				EXPERIMENT 2
PRIOR KNOWLEDGE	OR	SE	z	p	OR	SE	z	p
0%	4.35	2.50	2.55	.011*	2.50	1.38	1.66	.097
33%	3.18	0.71	5.16	<.001***	8.33	2.17	8.13	<.001***
67%	1.19	0.22	0.95	.341	3.67	0.78	6.08	<.001***
100%	0.23	0.05	–6.46	<.001***	0.21	0.05	–6.60	<.001***

[i] Note. Odds Ratios (OR) reflect the contrast: Experimental / Control for each prior knowledge condition. Values are based on estimated marginal means from hierarchical logistic regression models. z-values and p-values correspond to model-based comparisons. Significance levels: *p<.05, **p<.01, ***p<.001.

Table 4

Pairwise Contrasts Between Prior Knowledge Conditions Within Groups.

GROUP	CONTRAST	EXPERIMENT 1				EXPERIMENT 2
GROUP	CONTRAST	OR	SE	z	p	OR	SE	z	p
Experimental	0% vs 33%	0.88	0.26	–0.44	.971	0.40	0.13	–2.79	.027*
	0% vs 67%	2.08	0.85	1.80	.272	0.96	0.43	–0.10	.999
	0% vs 100%	13.38	5.28	6.57	<.001***	13.37	6.02	5.76	<.001***
	33% vs 67%	2.37	0.31	6.57	<.001***	2.37	0.36	5.76	<.001***
	33% vs 100%	15.26	3.79	10.98	<.001***	33.13	9.26	12.53	<.001***
	67% vs 100%	6.43	1.21	9.85	<.001***	13.96	2.87	12.82	<.001***
Control	0% vs 33%	0.64	0.24	–1.20	.629	1.34	0.42	0.95	.778
	0% vs 67%	0.57	0.29	–1.10	.688	1.40	0.59	0.81	.850
	0% vs 100%	0.71	0.32	–0.77	.869	1.14	0.44	0.35	.986
	33% vs 67%	0.89	0.13	–0.77	.869	1.05	0.13	0.35	.986
	33% vs 100%	1.10	0.25	0.43	.974	0.85	0.20	–0.68	.907
	67% vs 100%	1.24	0.23	1.15	.660	0.81	0.16	–1.07	.708

[i] Note. Odds Ratios (OR) reflect the contrasts between prior knowledge conditions within each group (Experimental/ Control). Values are based on estimated marginal means from hierarchical logistic regression models. z-values and p-values correspond to model-based comparisons. Significance levels: *p<.05, **p<.01, ***p<.001.

Table 5

Pairwise Contrasts for Score Differences by the Decision to Read.

PRIOR KNOWLEDGE	EXPERIMENT 1				EXPERIMENT 2
PRIOR KNOWLEDGE	DIFF	SE	t	p	DIFF	SE	t	p
0%	–0.54	0.20	–2.71	.007**	–0.97	0.16	–6.05	<.001***
33%	–0.64	0.08	–7.78	<.001***	–0.76	0.08	–8.92	<.001***
67%	–0.39	0.06	–6.56	<.001***	–0.41	0.06	–7.03	<.001***
100%	–0.05	0.06	–0.85	.397	–0.04	0.06	–0.68	.499

[i] Note. Values reflect estimated marginal mean differences in pretest-to-posttest scores between different decisions– reading or skipping– for each prior knowledge condition. Negative values indicate larger score gains in the “read” group. Significance levels: *p<.05, **p<.01, ***p<.001.

Table 6

Pairwise Contrasts for Score Differences Between Groups.

PRIOR KNOWLEDGE	EXPERIMENT 1				EXPERIMENT 2
PRIOR KNOWLEDGE	DIFF	SE	t	p	DIFF	SE	t	p
0%	0.47	0.21	2.21	.027*	0.39	0.17	2.26	<.024*
33%	0.10	0.10	1.02	.310	0.56	0.10	5.83	<.001***
67%	0.08	0.08	1.08	.281	0.27	0.07	3.64	<.001***
100%	0.06	0.08	0.71	.477	0.11	0.07	1.48	.138

[i] Note. Values reflect estimated marginal mean differences in pretest-to-posttest scores between experimental and control groups for each prior knowledge condition. Positive values indicate larger score gains in the experimental group. Significance levels: *p<.05, **p<.01, ***p<.001.

Before conducting our main analysis, we excluded participants who made uniform decisions (as we preregistered; see Online Supplementary for the data analysis considering all participants) to read across the experiment (i.e., always choosing to read or always choosing to skip, n = 118 in the experimental group and n = 148 in the control group). We excluded these participants as they did not add any variance to our probability to read variable. In total, n = 100 participants in the control group (40.3%) and n = 76 in the experimental group (30.0%) chose to read all six chapters. In contrast, n = 48 (19.4%) and n = 42 (16.6%) participants in the control and experimental groups, respectively, skipped all chapters. The proportion of participants who read at least one chapter was 80.6% in the control group and 83.4% in the experimental group. After we applied this exclusion criterion, the remaining sample considered n = 235 participants (n = 135 in the experimental group and n = 100 in the control group).

Illustrating Knowledge Gaps Increases Information-Seeking Behavior

Our first analysis examined participants’ probability to read as a function of their prior knowledge and group. Results showed a significant interaction between prior knowledge and group for both the first (β = 38.71, z = 7.75 p < .001) and the second polynomial term (β = 9.58, z = 2.04, p = .042). This indicated a difference between groups in the relationship between prior knowledge and the probability to read. We expected an inverted U-shaped relationship between prior knowledge and the probability to read in the experimental group. However, the curve for the experimental group rather suggested a non-linear and negative relationship between prior knowledge and reading behavior with relatively similar average values for 0% prior knowledge and 33% prior knowledge (see Figure 2a). Participants in the control group had relatively similar average values across prior knowledge conditions (also see pairwise comparisons reported below).

As expected, the pairwise comparisons between the two groups at each prior knowledge condition revealed that participants in the experimental group had a significantly higher probability to read than the control group at prior knowledge levels of 33%, and a significantly lower probability to read at 100% prior knowledge (see Figure 2a and Table 3). In addition, and against our expectations, we observed that participants in the experimental group had a significantly higher probability to read than the control group at prior knowledge levels of 0%, and the results did not reveal a significant difference between groups at 67% prior knowledge.

The pairwise comparisons within each group revealed significant differences in the experimental group between 0% and 100%, 33% and 100%, and 67% and 100%, in line with our expectations. Against our expectations, we observed a significant difference between 33% and 67% prior knowledge (see Table 4). In the control group, as expected, we observed no significant differences in the probability to read between prior knowledge conditions.

Knowledge Gains as a Function of Prior Knowledge and Decision to Read

As expected, the pairwise comparisons indicated significant differences in knowledge gains between reading and skipping at the three prior knowledge conditions of 0%, 33% and 67%. The comparison at 100% prior knowledge was not significantly different (see Figure 2b and Table 5).

Knowledge Gains as a Function of Prior Knowledge and Group

Results of knowledge gains as a function of prior knowledge and group are shown in Figure 2c. Pairwise comparisons between groups did not show significant results, except for the 0% prior knowledge condition (see Table 6). This indicated that the score differences from pretest to posttest were higher in the experimental group only for very large knowledge gaps. As such, this result was not in line with our expectation that a significantly higher probability to read, as observed for the 33% prior knowledge condition, is accompanied with significantly increased knowledge gains.

Additional Analyses

We further examined reading time as a function of prior knowledge and group. Results for pairwise contrasts between groups for each prior knowledge condition are shown in Table 7 and visualized in Figure 3. In Experiment 1, results revealed a significantly longer reading times in the experimental group compared to the control group at 67% prior knowledge.

Table 7

Pairwise Contrasts for Reading Times Between Groups. Note. Values reflect estimated marginal mean differences in reading time (in seconds) between experimental and control groups at each prior knowledge level. Positive differences (Diff) indicate longer reading times in the experimental group. Significance levels: *p<.05, **p<.01, ***p<.001.

PRIOR KNOWLEDGE	EXPERIMENT 1				EXPERIMENT 2
PRIOR KNOWLEDGE	DIFF	SE	t	p	DIFF	SE	t	p
0%	68.80	73.90	0.93	.352	5.77	81.8	0.07	.944
33%	43.22	31.50	1.37	.171	–9.93	34.7	–0.29	.775
67%	23.00	11.50	2.00	.046*	–14.95	15.2	–0.98	.326
100%	8.15	9.61	0.85	.397	–19.31	13.3	–0.70	.484

Plots display reading time in seconds by prior knowledge and group.

Finally, we examined how strongly prior knowledge was correlated across chapters. Therefore, we calculated correlation matrices and constructed heatmaps for each experiment (Figure 4). Across all experiments, prior knowledge showed correlations between r = .15 and r = .43 between chapters. Following common guidelines (Cohen, 1988), our data shows small-to-medium positive correlations of prior knowledge.

Heatmaps display correlations for prior knowledge between chapters for Experiment 1 and Experiment 2.

Discussion

The aim of Experiment 1 was to test whether knowledge gap illustrations spark curiosity. First, we observed significantly different regression curves for the two groups, suggesting that illustrating knowledge gaps affected the experimental group differently than the control group for whom we did not illustrate knowledge gaps. Importantly, we also observed this in the pilot experiment (see Online Supplementary). Further inspection of the contrasts between groups indicated a significant difference between the two groups at 0% and 33% prior knowledge, with a higher probability to read in the experimental group than the control group. While we expected high information-seeking for moderate knowledge gap illustrations (33% prior knowledge), we did not expect it at 0% prior knowledge. However, we only observed a significant difference at 33% prior knowledge in the pilot experiment and did not observe a significant difference at 0% prior knowledge in the pilot experiment. Moreover, only few participants had 0% prior knowledge (see Table 2). Therefore, the significant difference at 0% prior knowledge has to be interpreted with caution. In addition, and as expected, the experimental group had a significantly lower probability to read than the control group at 100% prior knowledge. This supports our hypothesis that curiosity vanishes when illustrating that full knowledge exists. Finally, we observed no differences in reading probabilities between the two groups at 67% prior knowledge, even though we expected such differences not only at 33% prior knowledge but also at 67% prior knowledge. Nevertheless, the differing regression curves reliably found in the pilot experiment and Experiment 1 as well as the robust finding of a significant difference at 33% prior knowledge and 100% prior knowledge largely support our hypothesis that illustrating moderate knowledge gaps with lower knowledge spark curiosity and conversely, that illustrating that full knowledge exists diminishes curiosity.

We substantiated these results by also examining knowledge gains across groups for conditions where knowledge gaps existed (0%, 33%, and 67%). We observed (in the pilot experiment and Experiment 1) that information-seeking behavior was accompanied by significantly increased knowledge gains across groups. However, between-group comparisons at each prior knowledge condition were not in line with our expectations. We expected larger knowledge gains in the experimental group for prior knowledge conditions where we also observed elevated reading probabilities compared to the control group. However, we only observed a significant difference in knowledge gains at the 0% prior knowledge condition in Experiment 1, but not at the 33% condition where we observed higher reading probabilities in the experimental group. Note that we also did not observe significantly larger knowledge gains at 33% prior knowledge in the pilot experiment.

We reasoned that increasing the utility of gaining knowledge may modulate the results towards larger knowledge gains for the experimental group than the control group. In addition, we hypothesized that increasing the utility of gaining information should results in a robust difference in the probability to read at 0% prior knowledge, favoring the experimental group. We therefore intended to increase the utility of gaining knowledge in Experiment 2.

Experiment 2

In Experiment 2, we sought to examine whether increasing the utility for reading chapters affected participants’ probability to close large knowledge gaps (i.e., 0% prior knowledge). In particular, we hypothesized the inclusion of a utility cue in the instructions (i.e., telling participants that a posttest would follow the reading phase) to increase the probability to read as well as increase knowledge gains in the experimental group in comparison to the control group, specifically for large knowledge gaps (i.e., 0% prior knowledge).

In addition, we had the same expectations as in Experiment 1. We expected that illustrating moderate knowledge gaps (i.e., 33% and 67%) would result in an elevated probability to read (i.e., information-seeking behavior). We further expected increased information-seeking behavior to result in heightened knowledge gains in the experimental group.

We preregistered Experiment 2 via AsPredicted (AsPredicted #222,391 on 04/10/2025 at 01:44 AM [PT]). As in Experiment 1, we only included participants with variation in their reading behavior in our final analysis. Importantly, we preregistered this inclusion criteria and also report an analysis considering all participants in the Online Supplementary. As we did for the power simulation for Experiment 1, we ran 100 simulations and targeted an alpha level of α = .05 at a sample size of n = 200 for the first power simulation on the interaction effect of prior knowledge and group on the probability to read. The power simulation calculated a power of 99% with a 95% confidence interval (CI) between 94.6% and 100%. This calculation was based on the effect size β = 11.70 for the quadratic term of the interaction that we found in our pilot experiment for Experiment 2 (see the Online Supplementary). As we expected to exclude half of our participants, we conservatively selected a sample size of n = 500 participants.