Curiosity evolves as information unfolds

To examine how curiosity directs information seeking as information evolves, we developed the Evolving Line Drawing Task (ELDT) where participants watched animated continuous single-line drawings unfold over time. Participants (N = 2,043) watched a minimum of 25 videos, each between 20 and 30 s in length (See Materials and Methods for participant exclusions and demographics). During video watching, participants submitted Mid-Video Decisions that included using analog scales to report on their curiosity, their confidence about the drawing’s identity, and one of three randomly determined affective states (enjoyment, tension, or frustration) assigned between participants (Fig. 1A) (Materials and Methods). They also provided a guess about the identity of the drawing given their current informational state.

After providing responses to all four questions, participants chose whether to continue watching or stop that video and move to the next trial. We manipulated the Stop Outcome between participants, such that in one group (Forgo Resolution), stopping the video ended information gathering and moved subjects to the Post-Video ratings. In the other group (Jump-to Resolution), stopping brought participants immediately to the end of the video to see the final drawing fully revealed before moving to the Post-Video ratings. If a participant chose to continue watching a video on their first Mid-Video Decision, they could be probed with the same questions at a later time point. Participants were told that their choices and guesses would not impact their compensation. Updating guesses was described as an expected part of the video-watching process and participants would not be penalized for wrong guesses.

The agency for when Mid-Video Decisions occurred was manipulated between Study 1 (Predetermined Timing, N = 1,033) and Study 2 (Free Timing, N = 1,010) (Fig. 1A). The aim of Study 1 was to uniformly sample time points across the entire video-watching timeline. As such, participants were informed that there would be intermittently placed pauses (determined by the computer) throughout the video (1 to 3 per video) where they would be asked to answer Mid-Video Decisions. The aim of Study 2 was to investigate whether the experience of information gathering changed if participants had agency over the timing of when to submit Mid-Video Decisions. Here, instead of experiencing randomly placed pauses, a textbox remained visible underneath the video, and participants could freely submit guesses and Mid-Video Decisions. Participants in both studies were divided into the two Stop-Outcome conditions (Forgo Resolution: Study 1 N = 501, Study 2 N = 485; Jump-to Resolution: Study 1 N = 532, Study 2 N = 525). Study 2 was otherwise identical to Study 1.

As the key findings were consistent across both studies, we first present combined results from Study 1 and Study 2. We then directly compare the Studies to test the effects of timing agency on choice behavior.

We first explored how participants’ informational state changed as a video progressed by leveraging the unique properties of each video. We characterized informational state by how close a participant was to identifying the drawing using features of each individual stimulus (stimulus entropy) and based on subject-specific measures (guess accuracy and confidence). Stimulus entropy was computed using the number and frequency of unique guesses made at each time point (across all participants) (33), allowing for the examination of variance in guessing and the change in variance as participants approached resolution (Materials and Methods). In Study 1, participants made an average of 1.5 guesses per video (SD = 0.6, Range = 1 to 3) and in Study 2, participants made an average of 1.3 guesses per video (SD = 0.6, Range = 1 to 6). For more descriptive measures about guessing see SI Appendix, Guessing Descriptive Information for Study 1 and Study 2. To visualize the unique dynamics reflected in each line drawing, we pooled the data across Study 1 (Predetermined Timing) and Study 2 (Free Timing) (Fig. 1 B and C); the patterns for each individual study are nearly identical (SI Appendix, Temporal Dynamics of Guess Accuracy and Stimulus Entropy).

We first examined changes in stimulus entropy across time, hypothesizing that entropy should decrease as more information about the stimulus identity was revealed. We indeed found that stimulus entropy decreased significantly across time (β = −0.59, 95% CI [−0.59, −0.58], z = −307.65, P < 0.001). Additionally, we saw substantial variation in stimulus entropy and its temporal evolution within our stimulus set. To illustrate this point, we present the collection of unique guesses for two sample videos at a single time point (Fig. 1B).

We next tested whether participants would report more correct guesses as the video progressed and stimulus entropy decreased. We found that guess accuracy increased across time (β = 2.31, 95% CI [2.28, 2.35], z = 127.59, P < 0.001), indicating that participants were indeed reaching resolution about the identity of the drawings. This was further confirmed by the participant’s confidence ratings (β = 0.50, 95% CI [0.49, 0.51], z = 149.58, P < 0.001), which also increased over time (SI Appendix, Temporal Dynamics of Subjective States). However, like entropy, this varied substantially across videos (Fig. 1C). Collectively, our results illustrate that participants are reaching resolution about the identities of these drawings, but the path to resolution varies as a function of the specific video properties (i.e., how a drawing unfolds). Our next series of analyses investigated how curiosity relates to such evolving informational states.

In general, both Studies elicited moderate to high levels of curiosity (Study 1: M = 70.54/100, SD = 27.12; Study 2: M = 69.90/100, SD = 26.10). Curiosity ratings of 0 (out of 100) were extremely rare across both studies (2% of trials). At the beginning of a video (within the first 3 s), curiosity ratings started relatively high (Study 1: M = 70.00, SD = 28; Study 2: M = 67.28, SD = 5.44), while confidence ratings (out of 100) about the identity started relatively low (Study 1: M = 16.00, SD= 24.65; Study 2: M = 35.13, SD= 34.83). For more descriptive information about ratings see SI Appendix, Temporal Dynamics of Subjective States.

We first aimed to investigate how an evolving informational state could shape changes in curiosity. To test this, we constructed a linear mixed-effects model for each Study with participant as a random intercept to examine how confidence, guess accuracy, and stimulus entropy predicted curiosity ratings. These models also included Stop-Outcome condition (Forgo vs. Jump-to) to see whether curiosity was also dependent on the information-seeking choices available to a participant. Additional model information and all parameter estimates are provided in SI Appendix, Tables S3 and S6.

Based on prior findings, we hypothesized that curiosity would be positively associated with factors that signaled uncertainty (33) and would be negatively related to factors that signaled a resolution of uncertainty (5, 31, 38). Indeed, we found that curiosity was positively related to stimulus entropy (Study 1: β = 1.15, 95% CI [0.79, 1.51], z = 6.19, P < 0.001; Study 2: β = 1.84, 95% CI [1.43, 2.25], z = 8.74, P < 0.001) and negatively related to confidence (Study 1: β = −3.63, 95% CI [−4.02, −3.23], z = −17.82, P < 0.001; Study 2: β = −3.21, 95% CI [−3.68, −2.75], z = −13.64, P < 0.001); however, we did not find a main effect for guess accuracy (Study 1: β = −0.46, 95% CI [−1.04, 0.11], z = −1.58, P = 0.113; Study 2: β = −0.20, 95% CI [−0.79, 0.39], z = −0.66, P = 0.508).

We further tested whether these features could modulate one another’s impact on curiosity. We found a significant interaction between stimulus entropy and subjective ratings of confidence (Study 1: β = 1.67, 95% CI [1.38, 1.97], z = 11.00, P < 0.001; Study 2: β = 1.28, 95% CI [0.98, 1.59], z = 8.21, P < 0.001) (Fig. 2A). Post hoc analyses indicated that across both studies, high stimulus entropy dampened the relationship between confidence and curiosity (Study 1: β = −3.35, z = −11.00, P < 0.001, Study 2: β = −3.35, z = −11.00, P < 0.001), suggesting that high stimulus uncertainty can sustain curiosity even when confidence of knowing is high. We also found a significant interaction between confidence and guess accuracy (Study 1: β = −2.89, 95% CI [−3.50, −2.28], z = −9.30, P < 0.001; Study 2: β = −1.48, 95% CI [−2.08, −0.89], z = −4.86, P < 0.001), such that when confidence is low, people reported more curiosity for correct compared to incorrect guesses (Study 1: β = 2.38, z = 5.46, P < 0.001; Study 2: β = 1.26, z = 3.25, P = 0.006) but when confidence is high, people reported less curiosity for correct compared to incorrect guesses (Study 1: β = −3.39, z = −7.98, P < 0.001; Study 2: β = −1.71, z = −3.71, P = 0.001) (Fig. 2B). This suggests that curiosity is not only evoked during the experience of uncertainty (i.e., low confidence) but can also arise before high confidence errors. Finally, in both Studies, we did not find a difference in curiosity ratings between Stop-Outcome conditions (Study 1: β = 0.17, 95% CI [−2.17, 2.52], z = 0.14, P = 0.886; Study 2: β = −1.67, 95% CI [−0.86, 4.19], z = 1.30, P = 0.196).

We followed up these analyses to specifically examine how the updating of guesses shaped the growth and subsidence of curiosity using trials where participants reported more than one guess (SI Appendix, Intra-Trial Analyses). We found that curiosity declined when updating from an incorrect to a correct guess (Study 1: β = −3.60, z = −20.04, P < 0.001; Study 2: β = −2.41, z = −13.9, P < 0.001) but increased in Study 1 (Study 1: β = 1.12, z = 9.86, P < 0.001) and was unchanged in Study 2 (Study 2: β = 0.25, z = 1.24, P = 0.216) when participants updated from an incorrect to another incorrect guess. This suggests that curiosity can be sustained or even grow as participants continue to explore potential (but ultimately wrong) identities.

Taken together, these results demonstrate that curiosity can change across an evolving informational state, that curiosity is impacted both by factors internal to a specific participant (here, subjective confidence and guess accuracy) as well as externally driven features of the informational landscape (here, stimulus entropy), and that curiosity is not a one-dimensional reflection of subjective uncertainty. Finally, agency over the occurrence of Mid-Video Decisions, whether through Predetermined pauses or Freely by the participant, did not alter the interactions among curiosity and informational state, suggesting that these observed relationships are unlikely to be an outcome of specific probing intervals/timing.

We next investigated across our two Studies how curiosity influenced choices to continue or stop a video as information evolved. For our Forgo Resolution condition, we hypothesized that, consistent with prior research on discrete information seeking (e.g., trivia Q&A paradigms), high curiosity would predict continuing to watch a video since this was the only path to receive resolution (Fig. 3A) (11, 36). For our Jump-to Resolution condition, we hypothesized that high curiosity would either predict a decreased probability of continuing to watch a video, demonstrating a sensitivity to the costs of delays and thus a preference for the fastest path to resolution (outcome-focused) (14, 15, 39), or that curiosity would predict an increased probability of continuing to watch a video, signaling that experiencing the process of information unfolding may also confer value (process-focused) (17, 26) (Fig. 3A).

Across both Studies, we found that high curiosity predicted continuing to watch a video in the Forgo Resolution condition (Study 1: β = 0.81, 95% CI [0.76, 0.87], z = 27.02, P < 0.001, Study 2: β = 0.87, 95% CI [0.79, 0.95], z = 21.05, P < 0.001) (Fig. 3 B and C), suggesting that curiosity encourages uncertainty resolution during continuous information gathering, similar to that observed during discrete information seeking (10, 11, 36). Crucially, we found that even when participants had the option to receive immediate resolution (Jump-to Resolution), higher curiosity still increased the likelihood of continuing a video (Study 1: β = 0.64, 95% CI [0.59, 0.70], z = 22.59, P < 0.001; Study 2: β = 0.50, 95% CI [0.43, 0.57], z = 13.55, P < 0.001) (Fig. 3 B and C). These results suggest that people do not always seek immediate resolution for their curiosity and that curiosity can instead encourage prioritizing the process of information gathering over instant obtainment of the outcome (26).

We next sought to contextualize the relationship between curiosity and choice behavior alongside the changes in informational state and to test whether the potential interplay between curiosity and informational state would be dependent on choice outcome (Forgo vs. Jump-to). We fit a mixed-effects logistic regression with participant as a random intercept that included the participant’s curiosity ratings, confidence ratings, guess accuracy, stimulus entropy (at the time of the Mid-Video Decision), as well as all two-way interactions with the Stop-Outcome condition (SI Appendix, Tables S4 and S7 for full model information and all parameter estimates). Consistent across both Studies, we found that even when accounting for informational state, curiosity remained a significant predictor of continuing a video (Study 1: β = 0.50, 95% CI [0.41, 0.60], z = 10.54, P < 0.001; Study 2: β = 0.50, 95% CI [0.37, 0.63], z = 7.41, P < 0.001). In addition to curiosity, stimulus entropy also increased the likelihood of continuing a video (Study 1: β = 0.11, 95% CI [0.03, 0.18], z = 2.75, P = 0.006; Study 2: β = 0.33, 95% CI [0.27, 0.40], z = 10.31, P < 0.001). Conversely, correctly guessing the identity of the drawing (Study 1: β = −2.08, 95% CI [−2.24, −1.91], z = −25.14, P < 0.001; Study 2: β = −1.94, 95% CI [−2.16, −1.73], z = −17.66, P < 0.001) and high confidence (Study 1: β = −0.79, 95% CI [−0.88, −0.70], z = −17.53, P < 0.001; Study 2: β = −0.78, 95% CI [−0.90, −0.67], z = −13.59, P < 0.001) both predicted stopping a video early (Fig. 4).

We followed up this analysis to test whether curiosity would still predict continuation when participants had reached resolution (defined as correct guesses with subjective confidence ratings in the top 25 percentile). Indeed, even when examining only highly confident, correct guesses, curiosity predicted continuing a video not only when it was the only path to view the entire video (Study 1 Forgo Resolution: β = 0.96, z = 9.92, P < 0.001; Study 2 Forgo Resolution: β = 0.71, z = 7.42, P < 0.001) but even when the completed drawing was immediately accessible (Study 1 Jump-to Resolution: β = 0.62, z = 7.23, P < 0.001; Study 2 Jump-to Resolution: β = 0.27, z = 3.25, P = 0.001). In sum, high curiosity can promote continuing a video beyond simple resolution, encouraging an investment in experiencing the entire unfolding of information.

We then compared our two Stop-Outcome conditions and we found that participants in the Jump-to Resolution condition were overall less likely to continue videos compared to the Forgo Resolution condition (Study 1: β = −1.68, 95% CI [−2.07, −1.29], z = −8.48, P < 0.001; Study 2: β = −1.44, 95% CI [−1.93, −0.95], z = −5.76, P < 0.001), suggesting that having the option to resolve uncertainty immediately reduced the likelihood of prolonging information gathering. Furthermore, we found that the Stop-Outcome manipulation significantly interacted with guess accuracy (Study 1: β = 1.25, 95% CI [1.03, 1.46], z = 11.42, P < 0.001; Study 2: β = 0.97, 95% CI [0.72, 1.23], z = 7.45, P < 0.001). Interestingly, post hoc analyses indicated that in the Jump-to Resolution condition compared to the Forgo Resolution, participants were significantly less likely to continue watching a video after an incorrect guess (Study 1: OR = 5.36, z = 8.50, P < 0.001; Study 2: OR = 4.01, z = 5.59, P < 0.001) but not after a correct guess (Study 1: OR = 1.54, z = 2.16, P = 0.092; Study 2: OR = 1.52, z = 1.80, P = 0.217) (Fig. 4 A and B). We also found that curiosity was modulated by Stop-Outcome (Study 1: β = −0.17, 95% CI [−0.26, −0.07], z = −3.35, P < 0.001; Study 2: β = −0.32, 95% CI [−0.45, −0.19], z = −4.98, P < 0.001), such that curiosity was a stronger predictor of choice in the Forgo compared to the Jump-to Resolution condition indicated by as steeper slope (Study 1: β = 0.17, z = 3.35, P < 0.001, Study 2: β = 0.32, z = 4.98, P < 0.001), yet remained a significant predictor in both conditions (Forgo: β = 0.59, z = 14.35, P < 0.001; Jump-to: β = 0.43, z = 11.77, P < 0.001) (Fig. 3 B and C). To summarize, while the option to receive immediate resolution (Jump-to Resolution) did decrease decisions to continue a video under uncertainty, curiosity remained a positive predictor of continuing a video, consistent across Study 1 and Study 2.

Finally, we examined whether subjective curiosity interacted with a participant’s informational state to direct choice, revealing differences between Study 1 and Study 2. In Study 1, we found that curiosity significantly dampened the effect of guess accuracy (β = 0.17, 95% CI [0.06, 0.27], z = 3.14, P = 0.002) and confidence (β = −0.10, 95% CI [−0.16, −0.04], z = −3.09, P = 0.002) on choice behavior but did not interact with entropy (P < 0.05). In contrast to Study 1, there was no significant interaction between curiosity and the individual signals of resolution, guess accuracy, or confidence (all P > 0.05) in Study 2; instead, we found a significant interaction between curiosity and stimulus entropy (β = 0.07, 95% CI [0.01, 0.13], z = 2.19, P = 0.029), such that curiosity was a stronger predictor of choice when entropy was high (β = 0.13, z = 2.19, P = 0.029). Thus, when experiencing greater agency over the timing of their decision, the influence of curiosity (on choice), was more strongly modulated by features related to stimulus uncertainty and less so by the subjective experience of uncertainty.

To start, we explored how affect ratings changed across time. We found that enjoyment increased across time (β = 3.64, z = 15.62, P < 0.001), while frustration and tension decreased (Frustration: β = −3.94, z = −13.61, P < 0.001; Tension: β = −1.51, z = −5.26, P < 0.001) (for figures see SI Appendix, Temporal Dynamics of Subjective States), suggesting that across the passage of time, positive affect increased and negative affect decreased, consistent with prior work (35). Given the dynamic relationships between curiosity and informational state in driving information gathering, we next explored how these dynamics shaped the affective experience of watching the videos. We fit three mixed-effects linear regression models with participant as a random intercept to investigate the relationships among curiosity, informational state, and these three affective states (See SI Appendix, Tables S5 and S8 for full model information and all parameter estimates).

Within these models, we also included Stop-Outcome condition and all two-way interactions with Stop-Outcome condition, to test whether the affective experience of information gathering depended on the outcome of choice behavior (SI Appendix, Affective Experience by Stop Outcome Conditions).

Across Study 1 (Predetermined Timing) and Study 2 (Free Timing), we demonstrated consistent relationships between curiosity and informational states as well as a persistent role of curiosity in promoting choices to continue a video even when immediate resolution was available. When we investigated the role of affect in information gathering, we saw differences in the patterns of association across Study 1 and Study 2. To better understand how agency over the timing of decisions shapes the information-gathering experience, we directly compared Study 1 and Study 2, first examining differences in reaching resolution and then in the affective experience associated with each Study.

We first investigated whether participants were waiting to reach resolution before they stopped a video early in Study 2. Focusing on the final guess provided during video watching, which signified a participant’s most updated idea about the identity of the drawing, we found that the proportion of correct final guesses in Study 2 was higher (more correct guesses reported) than in Study 1 [chi-squared(1, 40,754) = 2,065.1, P < 0.001, Study 1: 49.2% correct last guesses; Study 2: 70.6% correct last guesses] (Fig. 6A). Given that participants in Study 2 provided overall more correct last guesses, we explored whether participants in Study 2 were gathering more information prior to stopping compared to Study 1. Focusing only on trials where a participant stopped a video early, we found that in Study 2, participants watched more of a video before stopping [F(1, 21,673) = 1841, P < 0.001; Study 1: M = 48.9%, SD = 25.6% video watched at stop; Study 2: M = 65.1%, SD = 21.7% video watched at stop] compared to Study 1. Next, we examined whether reaching resolution, as measured by a correct guess, led to different rates of stopping videos early across our two studies. Using a logistic regression, we found a significant interaction between Study and guess accuracy (β = −0.55, 95% CI [−0.63, −0.47], z = −14.00, P < 0.001), indicating that continuation rates were lower in Study 2 compared to Study 1 only when there was a correct guess reported (OR = 1.40, z = 16.91, P < 0.001). When an incorrect guess was reported, we found the opposite effect, such that continuation rates were higher in Study 2 compared to Study 1 (OR = 0.81, z = −6.26, P < 0.001) (Fig. 6B). This suggests that when participants had greater agency over their information-gathering experience, their choices to continue were more sensitive to reaching resolution.

Finally, using a series of linear regressions with Study as our predictor, we compared the overall affective experience between the two studies to assess whether the greater agency impacted how the information-gathering process felt. We found that participants in Study 2 reported overall higher ratings of enjoyment (β = 4.93, 95% CI [4.33, 5.54], t = 15.89, P < 0.001) and lower ratings of frustration (β = −8.43, 95% CI [−9.19, −7.67], t = −21.71, P < 0.001), with no statistically significant difference in tension ratings (β = −0.65, 95% CI [−1.40, 0.09], t = −1.72, P = 0.09) (Fig. 6C). These findings suggest that allowing participants to freely decide when to make Mid-Video Decisions further heightened the enjoyment of watching these videos and reduced the experienced frustration, pointing to the additional reward value inherent in exerting agency (41).

Collectively, these results reveal that relative to Study 1, participants in Study 2 who were able to control their information gathering reported a more positive affective experience, showed increased sensitivity to reaching resolution in choices to continue, and were more likely to wait longer to resolve the identity of the drawing. With greater agency over when to initiate Mid-Video Decisions, we might expect that participants would only provide guesses upon reaching a certain level of confidence. As such, this may be expected to weaken the association between curiosity and choices to continue. However, we found that participants still maintained the relationship between curiosity and choices to continue a video even when they could control the timing of their Mid-Video decisions.

Using the experimental paradigm, the Evolving Line Drawing task (ELDT), designed to evoke and characterize the dynamics of curiosity, we showed how curiosity can evolve as a function of the information available to a participant. We saw that curiosity rose with stimulus entropy, suggesting that moments of high environmental variance can sustain curiosity. Conversely, curiosity declined as confidence in stimulus identity increased, suggesting that curiosity subsided (but did not dissipate) with the reduction in subjective experience of uncertainty. Crucially, we also demonstrated that within ongoing information gathering, higher curiosity predicted the choice to delay resolution not only when the alternative was to forgo resolution, but even when the alternative was to immediately resolve uncertainty and view the final drawing. The tendency for higher curiosity to prolong information gathering persisted even after participants were confident in the drawing’s identity. These findings support the notion that under continuously evolving information, higher curiosity may increase the value of perceiving the temporal progression to completion, prioritizing the process of information gathering over hastening the outcome.

While curiosity encouraged continuing to watch videos, confidence and guess accuracy did promote choices to stop videos early, indicating that the subjective experience of reaching resolution also changed the value of continued information gathering. In both experiments, curiosity shared a consistent positive relationship with enjoyment and tension; however, curiosity could relate both positively and negatively to frustration, depending on the concurrent level of confidence. Moreover, giving participants full agency over when they could guess and choose to continue or stop videos (Study 2) strengthened the relationship between resolution and choice, such that participants were more likely to continue watching when uncertain and more likely to stop once they believed a resolution was reached. Overall, these results demonstrate that curiosity changed over the course of information gathering, yet consistently discouraged truncating the information-gathering process in a context that afforded the opportunity for continued discovery.

Amplifying and extending prior work exploring curiosity using discrete-resolution tasks (e.g., trivia Q&A paradigms), we demonstrate that when information is continuously evolving, curiosity motivates seeking resolution rather than forgoing it (10, 13, 14, 36). Additionally, high levels of uncertainty, as measured through low levels of confidence and incorrect guessing, also increased choices to continue rather than forgo information seeking, supporting past research (11, 32).

Expanding on recent work showing that curiosity encourages a preference for receiving partial information (hints) compared to immediately receiving the answer (26); here, we find that this preference for prolonging uncertainty persists throughout information gathering. Thus, curiosity experienced during information gathering can encourage prolonging the state of uncertainty, similar to how people prefer to experience movies, sports, or narratives that unfold across time (17, 19–22). However, higher curiosity still predicted continuing a video even after participants made correct, highly confident guesses. These findings demonstrate that the process of information gathering also holds value beyond attaining resolution of an event and suggest that curiosity increases the value of experiencing information unfold.

We also provide evidence for how the choice to stop information seeking is made: Once resolution was reached there was a distinct decrease in the desire for more information. Moreover, when participants had full control of time spent watching (Study 2), they were more likely to continue when uncertain but also more likely to stop when they resolved the drawing’s identity than when timing was constrained (Study 1). These findings suggest that people use the subjective experience of resolution as a marker of when the process of information gathering can be concluded, providing key insight into how people self-organize their information seeking (7, 42).

Our study contributes to current theoretical accounts aimed at understanding the relationship between curiosity and uncertainty (1, 2, 5, 8, 15, 38, 43). Varying accounts have posited that curiosity is piqued 1) under intermediate levels of confidence as in the “information-gap” hypothesis (13, 15), 2) when uncertainty is maximal, as in prediction error accounts, (10, 33), or 3) when uncertainty is minimal and a person feels they are on the precipice of knowing, as in the Region of Proximal Learning (38, 44). Moreover, recent work has documented that the relationship between curiosity and confidence is contextual; depending on the future utility of information (5). Collectively, these findings suggest that the relationship between curiosity and uncertainty cannot be described in a single functional form, suggesting instead that form is determined both by how curiosity is measured and the structure of the task. Expanding on this notion, we demonstrate that the relationship between uncertainty and curiosity can also change across an evolving informational state. Specifically, curiosity can be piqued by high states of variance or entropy but also when people report correct guesses but are not very confident (33, 45, 46). This suggests that curiosity can flexibly transition from a more diversive curiosity (i.e., what could this drawing be?) to a more specific curiosity (i.e., is it going to be a dog?) (47), providing support to the notion that curiosity is multifaceted even within a single episode (2). Future work is needed to understand when these different mappings between curiosity and uncertainty arise and how both subjective states and stimulus features can direct ongoing information-seeking behavior.

In both of our experiments, information gathering elicited moments of enjoyment, tension, and frustration, depending on the informational state of a participant. Our findings reveal an experience of curiosity that was mainly positive rather than aversive, but complex and dynamic, further suggesting that curiosity can take on a variety of affective flavors even within one episode. The positive relationship between curiosity and enjoyment stands in contrast with previous work characterizing a state of curiosity or prolonging uncertainty as aversive (15, 32, 34) and linked to unhappiness (16). Although the link between curiosity and enjoyment has been explored in personality trait differences in the desire for information (12, 27, 28), understanding within-person changes in the affective experience of curiosity is underexplored. Overall, we found that curiosity and confidence were both related to enjoyment, suggesting that both the state of wanting to know and the state of moving toward knowing can elicit enjoyment.

We found across both studies that curiosity was related to feelings of tension, possibly due to the experience of uncertainty. Interestingly, we found that curiosity shared a complex relationship with frustration, and this relationship also differed across Study 1 (Predetermined Timing) and Study 2 (Free-Timing). In Study 1, curiosity was negatively related to frustration under low confidence but positively related to frustration under high confidence. In Study 2, we found the opposite: Curiosity was positively related to frustration under low confidence and not related under high confidence. We speculate that in Study 1, the positive mapping of curiosity and frustration under high confidence could reflect a more deprivation-like curiosity that is associated with wanting confirmation about the correctness of an answer (28, 48). In Study 2, our findings aligned with prior work reporting feelings of annoyance related to long delays (35). This difference between Study 1 and 2 suggests that when participants have more control over the timing of their guesses and videos, being far away from resolution could feel more agitating, while coming to and validating an answer could coincide with less conflict. However, we found that overall, increasing a person’s agency over the timing of choices in Study 2 increased the positive affect experienced and decreased the negative affect, similar to previous work emphasizing the value of exerting control (41, 45). Although these affective relationships with curiosity provide insights into the range of emotional experiences that can be elicited by curiosity, we probed one emotion for each participant, making it challenging to compare the affective experience holistically. Future work will need to further characterize the affective experience of curiosity to unpack how the emotional experience of curiosity can impact behavior.

The divergent effects of curiosity on information seeking generate new questions about boundary conditions: When does curiosity prioritize the process of information gathering over a rapid resolution of uncertainty? Prior work has demonstrated that expectations around the structure of the environment can influence information-seeking choices (36, 46). An individual might deem prolonging uncertainty to be worthwhile when its resolution is definite, but not when resolution may not occur. Another important factor to consider, and a main motivation for this paradigm, is what process is happening while the resolution of uncertainty is being delayed. Within the ELDT, information continued to unfold across time, allowing for active hypothesis testing and updating. This stands in contrast with delay periods in which the participant passively waits for resolution, without the opportunity to gather more information (11, 36, 37). Under conditions where no further information is obtained, receiving immediate information, indeed, seems favored and is linked to anticipatory utility (37). Future studies will need to elucidate under what contexts curiosity promotes immediate uncertainty reduction or engaging in a prolonged gathering process.

What does it mean if curiosity sometimes directs choices toward and other times away from immediate resolution of uncertainty? One implication is that curiosity can prioritize understanding how information fits together. Curiosity has been linked to the process of sense-making (49), explanation-seeking (50–52), and in the creations of knowledge networks (53). All of these conceptualizations emphasize the process of experiencing an evolving informational state in building representations of the world. The choice to suspend resolution in this study, consequently, could reflect curiosity’s preference for and valuation of understanding the process of how each drawing comes to fruition. Delaying resolution to experience the process of information gathering can be a rational solution to maximize information gain (5). Thus, curiosity as prioritizing the process of information gathering offers an avenue to extend current theoretical and computational models of noninstrumental information seeking.

To conclude, by using the Evolving Line Drawing Task, designed to elicit curiosity over continuous information gathering, we reveal the dynamic evolution of curiosity along with information state. This experience of curiosity could reflect a state where prolonging uncertainty presents opportunities for hypothesis generation, testing, exploration, and wonder. Such curious states may promote information seeking beyond simple resolution to prioritize an understanding of how information fits together to create meaning.