Curiosity evolves as information unfolds

Curiosity is a fundamental driver of human behavior, often guiding actions toward reducing uncertainty and acquiring new knowledge. Experimental paradigms frequently show that curiosity directs choices to obtain information rather than forgo it, supporting a theoretical view of curiosity as a state that hastens uncertainty resolution. Yet real-world experiences, such as narratives, sports, and games, often involve prolonged anticipation and a preference to avoid spoilers, suggesting that people sometimes value the process of uncertainty unfolding. Understanding the temporal dynamics of curiosity during evolving information could clarify its behavioral manifestations and subjective affective experience, including why curiosity can feel positive or negative. Most lab tasks present discrete, single-shot revelations (e.g., trivia answers) or delays without new information, limiting insights into changing curiosity. This study introduces a dynamic paradigm in which information progressively reveals stimulus identity, allowing tests of three questions: how curiosity develops with informational state (entropy, accuracy, confidence), whether curiosity drives preferences for continued information gathering versus immediate resolution, and how the affective experience of curiosity interacts with informational states over time. Two studies compare externally imposed decision timing versus participant-controlled timing.

Prior work demonstrates that curiosity motivates information seeking and is linked to reward circuitry and learning enhancements. Theoretical accounts characterize curiosity as driven by uncertainty reduction (e.g., information-gap hypothesis, deprivation-like curiosity) or by prediction error and information gain, with some models proposing peak curiosity at intermediate confidence or near the Region of Proximal Learning. Empirically, uncertainty can increase curiosity but sometimes decreases happiness; people often prefer suspense and partial information (hints) over immediate answers, and may avoid spoilers. Curiosity’s affective tone varies, relating to both positive (enjoyment, interest) and negative (frustration, tension) emotions, and agency over choices can be inherently rewarding. Standard tasks often elicit curiosity via withholding a single answer with instantaneous resolution or delays without new information, limiting examination of curiosity’s evolution. The present work extends discrete-resolution paradigms by using continuously evolving stimuli to measure dynamic mappings among curiosity, uncertainty (entropy, confidence, accuracy), affect, and choice behavior.

Design: The Evolving Line Drawing Task (ELDT) presents animated single-line drawings that slowly resolve into familiar objects over 20–30 s. Participants make Mid-Video Decisions consisting of: (a) a guess of the drawing’s identity, (b) ratings of curiosity, (c) ratings of confidence (two variants collapsed in analysis), and (d) one affective state rating (enjoyment, tension, or frustration) assigned between subjects. After each Mid-Video Decision, participants choose to continue watching or stop. Stop-Outcome manipulation (between subjects): Forgo Resolution (stopping ends information gathering; no final revealed drawing) versus Jump-to Resolution (stopping immediately reveals the completed drawing). Timing manipulation (between studies): Study 1 (Predetermined Timing; N = 1,033): pauses 1–3 per video at random times to elicit Mid-Video Decisions. Study 2 (Free Timing; N = 1,010): a persistent textbox allows participants to freely submit guesses and Mid-Video Decisions at self-chosen times. Participants: 2,153 US adults recruited via Prolific (Sept 2020); after exclusions, final N = 2,043 (Study 1: N = 1,033; Study 2: N = 1,010). Demographics: mean age 32 years (SD 11.4), 48% women, 48% men, 4% gender nonconforming/prefer not to say; racial distribution: 69% White, 13% Asian, 8% Black, 10% mixed/other. Procedure: Each participant watched a minimum of 25 videos; could optionally watch up to 10 additional videos (32% chose extras; M = 2.72). Ratings used 1–100 sliders (confidence alternative time-needed question used 0–30 s). Guesses required common single English words (>1 character). Compensation fixed ($5.75; two-part version $9.00). No feedback on correctness; final answers entered post-video. Replay option (1.6% trials) allowed passive rewatch without pauses or guessing. Measures and computations: - Guess accuracy: mid-video guesses compared to participant’s standardized final answer set; mid-video guess accuracy = 47.9%; final answer accuracy = 98%. - Stimulus entropy: across all participants’ guesses at each video second, compute probabilities of unique guessed identities and Shannon entropy H(X) = −Σ P(x) log P(x), capturing variance and convergence over time. - Confidence: two questions collapsed into a single subjective confidence marker for analysis. Analytic approach: Mixed-effects linear and logistic regression models (lme4, lmerTest in R v4.0.3/RStudio v1.3.1093), reporting standardized β (and 95% CI) and z/t statistics. Models included random intercepts for participant. Key models: (1) curiosity predicted by entropy, confidence, guess accuracy, and Stop-Outcome; (2) choice (continue vs stop) predicted by curiosity, confidence, guess accuracy, entropy, interactions with Stop-Outcome; (3) affect (enjoyment, tension, frustration) predicted by curiosity, confidence, guess accuracy, entropy, interactions with Stop-Outcome. Visualization produced with Python (Seaborn/Matplotlib).

Informational dynamics: - Stimulus entropy decreased across time: β = −0.59, 95% CI [−0.59, −0.58], z = −307.65, P < 0.001. - Guess accuracy increased across time: β = 2.31, 95% CI [2.28, 2.35], z = 127.59, P < 0.001. - Confidence increased across time: β = 0.50, 95% CI [0.49, 0.51], z = 149.58, P < 0.001. Curiosity mappings: - Curiosity was positively related to stimulus entropy (Study 1: β = 1.15, z = 6.19, P < 0.001; Study 2: β = 1.84, z = 8.74, P < 0.001) and negatively related to confidence (Study 1: β = −3.63, z = −17.82, P < 0.001; Study 2: β = −3.21, z = −13.64, P < 0.001); no main effect of guess accuracy. - Interaction entropy × confidence: high stimulus entropy weakened the negative relationship between confidence and curiosity (Study 1: β = 1.67, z = 11.00, P < 0.001; Study 2: β = 1.28, z = 8.21, P < 0.001). - Interaction confidence × guess accuracy: when confidence was low, curiosity was higher for correct vs. incorrect guesses; when confidence was high, curiosity was higher for incorrect vs. correct guesses (Study 1: β = −2.89, z = −9.30, P < 0.001; Study 2: β = −1.48, z = −4.86, P < 0.001). - Within trials, updating from incorrect to correct guesses decreased curiosity (Study 1: β = −3.60, z = −20.04, P < 0.001; Study 2: β = −2.41, z = −13.9, P < 0.001); updating incorrect to incorrect increased curiosity in Study 1 (β = 1.12, z = 9.86, P < 0.001) and was unchanged in Study 2. Choice behavior: - Curiosity promoted continuing in Forgo Resolution (Study 1: B = 0.81, z = 27.02, P < 0.001; Study 2: B = 0.87, z = 21.05, P < 0.001) and also in Jump-to Resolution (Study 1: β = 0.64, z = 22.59, P < 0.001; Study 2: β = 0.50, z = 13.55, P < 0.001). - Controlling for informational state, curiosity remained a significant predictor of continuation (Study 1: β = 0.50, z = 10.54, P < 0.001; Study 2: β = 0.50, z = 7.41, P < 0.001). - Higher stimulus entropy increased continuation (Study 1: β = 0.11, z = 2.75, P = 0.006; Study 2: β = 0.33, z = 10.31, P < 0.001). - Correct guesses and higher confidence decreased continuation (Correct: Study 1: β = −2.08, z = −25.14, P < 0.001; Study 2: β = −1.94, z = −17.66, P < 0.001. Confidence: Study 1: β = −0.79, z = −17.53, P < 0.001; Study 2: β = −0.78, z = −13.59, P < 0.001). - Among highly confident correct guesses, curiosity still predicted continuation (Forgo: Study 1: β = 0.96, z = 9.92, P < 0.001; Study 2: β = 0.71, z = 7.42, P < 0.001. Jump-to: Study 1: β = 0.62, z = 7.23, P < 0.001; Study 2: β = 0.27, z = 3.25, P = 0.001). - Jump-to vs. Forgo: overall, Jump-to reduced continuation relative to Forgo (Study 1: β = −1.68, z = −8.48, P < 0.001; Study 2: β = −1.44, z = −5.76, P < 0.001), especially after incorrect guesses (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 correct guesses. - Curiosity’s predictive slope was steeper in Forgo than Jump-to (Study 1: β = 0.17, z = 3.35, P < 0.001; Study 2: β = 0.32, z = 4.98, P < 0.001). - Study differences: In Study 1, curiosity dampened effects of guess accuracy and confidence on choice; in Study 2, curiosity’s influence was more strongly modulated by stimulus entropy. Affective dynamics: - Across time, enjoyment increased (β = 3.64, z = 15.62, P < 0.001), frustration decreased (β = −3.94, z = −13.61, P < 0.001), tension decreased (β = −1.51, z = −5.26, P < 0.001). - Enjoyment was positively associated with confidence and guess accuracy (Study 1: β = 3.83 and β = 3.39, P < 0.001; Study 2: β = 2.80 and β = 1.66, P < 0.001); curiosity also positively related to enjoyment, with stronger association under low confidence (interaction curiosity × confidence significant in both studies). - Tension positively related to curiosity (Study 1: β = 5.07, z = 13.47, P < 0.001; Study 2: β = 3.26, z = 6.36, P < 0.001) and negatively to confidence (Study 1: β = −3.37; Study 2: β = −3.41; both P < 0.001). - Frustration: Study 1 showed negative associations with curiosity, confidence, and accuracy; Study 2 showed a negative association only with confidence. Confidence modulated the curiosity–frustration relationship in opposite directions across studies: in Study 1, under low confidence, curiosity negatively related to frustration, but under high confidence, positively related; in Study 2, under low confidence, curiosity positively related to frustration; under high confidence, no relation. Agency (Study 1 vs. Study 2) comparisons: - Higher proportion of correct last Mid-Video guesses in Study 2 vs. Study 1: 70.6% vs. 49.2%; χ²(1, 40,754) = 2,065.1, P < 0.001. - When stopping early, participants watched more of the video in Study 2: 65.1% vs. 48.9%; F(1, 21,673) = 1841, P < 0.001. - Continuation was more sensitive to reaching resolution in Study 2: lower continuation after correct guesses; higher continuation under incorrect guesses (ORs reported). - Study 2 reported higher enjoyment (β = 4.93, t = 15.89, P < 0.001) and lower frustration (β = −8.43, t = −21.71, P < 0.001); no difference in tension (β = −0.65, t = −1.72, P = 0.09).

The ELDT demonstrates that curiosity evolves dynamically with ongoing information accumulation. Curiosity increased with stimulus entropy and decreased with subjective confidence, yet remained elevated in contexts of high variance or when participants held low-confidence correct guesses, indicating that curiosity is multifaceted rather than a one-dimensional reflection of uncertainty. Importantly, curiosity consistently promoted continuing to gather information—even when immediate resolution (a spoiler) was available—supporting the idea that curiosity can prioritize valuing the process of discovery over hastening outcomes. Confidence and accurate guessing, by contrast, increased stopping behavior, suggesting that subjective resolution marks the natural conclusion of information gathering. Affectively, curiosity was linked to enjoyment and tension and had complex relations with frustration that depended on confidence and agency over decision timing. Granting participants control over timing (Study 2) enhanced positive affect, reduced frustration, and increased sensitivity to reaching resolution in choices, highlighting the inherent reward of agency and its role in structuring information-seeking behavior. These findings extend discrete-resolution models by showing that under continuously evolving information, curiosity motivates both resolution-seeking and the experiential value of perceiving information unfold, aligning with accounts of sense-making, explanation-seeking, and building knowledge networks.

Curiosity changes as information unfolds, tracking both stimulus-driven uncertainty (entropy) and subjective informational state (confidence, accuracy). Across two experiments, higher curiosity reliably encouraged participants to continue watching unfolding stimuli—even when immediate resolution was available—and often beyond achieving high-confidence correct guesses, indicating value in the process of discovery itself. Confidence and correct identification increased stopping, illustrating how subjective resolution organizes information seeking. Affect during information gathering is dynamic: curiosity is associated with enjoyment and tension, and its relation to frustration depends on confidence and agency. Future research should delineate boundary conditions under which curiosity favors immediate uncertainty reduction versus prolonged information gathering, examine contexts where delays do or do not provide new information, integrate computational models that capture process valuation and information gain, and more comprehensively probe multiple affective states within participants to characterize the emotional phenomenology of curiosity.

Affective measures were limited to probing one emotion per participant (enjoyment, tension, or frustration), making it challenging to compare the affective experience holistically across multiple emotions within individuals. The study did not provide feedback about correctness, so conclusions rely on participants’ subjective resolution; future work is needed to elucidate contexts where curiosity promotes immediate uncertainty reduction versus prolonged gathering, and to further characterize the affective experience of curiosity and its impact on behavior.