Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Communications Psychology volume 4, Article number: 31 (2026)
1874
1
Metrics details
Deciding whether, when, and which information to sample is critical for making effective decisions, yet the cognitive mechanisms of this process are not well understood. Here, we propose that key aspects of human information demand are explained by non-linear subjective perceptions of probabilistic losses or gains. Using behavioral testing and quantitative model comparisons across three independent participant samples (N = 50, 50, and 150), we show that a model that incorporates non-linear probability and value perception outperforms a model based on a linear mixture of motives in explaining instrumental and non-instrumental information demand. Moreover, individual non-linearities that best explained information demand were correlated with personality traits and with non-linearities explaining risk seeking/aversion in standard choice tasks. The results suggest that a computational framework rooted in the subjective perception of probability furthers our understanding of information demand and its relationship with decision making under risk and uncertainty.
A recent and rapidly growing body of research has emphasized the fact that, beyond the passive consumption of information, individuals actively select which information to engage with depending on their current context and goals. The importance of this active information demand (or information-seeking behavior) is heightened by the proliferation of digital information, which highlights how choices to consume or ignore information profoundly influence our self-perception1,2, understanding of the world3,4,5, decision-making6,7 and even mental health outcomes8,9,10,11,12,13. Despite its clear importance, information demand remains a relatively nascent area of inquiry within cognitive science, and the mechanisms underlying it are not well understood6.
Recent studies investigated information demand using monetary lottery tasks in which participants choose whether or when to request advance information about probabilistic rewards14,15,16,17,18,19,20,21,22,23,24,25. The results of some of these studies have been interpreted as supporting a mixed-motive account, whereby the value of information reflects the value assigned to the cognitive, instrumental, and hedonic aspects of the information7,14,16,18,21,26,27. Cognitive value refers to the ability of the information to resolve uncertainty about a future action or state. Instrumental value refers to the potential of the information to improve the outcomes of utility-relevant actions7,19,28. Finally, hedonic value refers to the ability of the information to induce a positive emotional state of anticipating a desirable outcome and avoid a negative state of dreading an undesirable outcome7. The three motives are proposed to have different strengths in different individuals, explaining the diversity of information-demand strategies7,26,29. Thus, some individuals may adopt strategies that are deemed normative in decision-theoretic approaches – assigning higher weight to cognitive or instrumental value thereby minimizing uncertainty and maximizing external rewards. In contrast, other individuals pursue apparently non-normative strategies – assigning higher weight to the hedonic component, prioritizing information that induces positive affect (e.g., anticipating potential financial gains but not potential losses14 or potentially adverse medical test results30,31,32) even at the expense of reducing uncertainty and maximizing rewards14,18,20,33.
While the mixed-motive idea offers a good descriptive account of the factors modulating information demand, significant questions remain about the quantitative description of these factors. Studies so far have assumed that hedonic, cognitive and instrumental motives reflect veridical (objective) probability and reward estimates that sum linearly with individually-variable weights7,14,16,18,21,26,27. This parametrization implies that all instances of non-normative information demand (i.e., demand that cannot be explained by uncertainty reduction or instrumental rewards) are ascribed to an additive component corresponding to the hedonic pursuit of desirable information. However, this functional form has not been rigorously evaluated against alternative explanations. A strong contender for such an explanation is that non-normative patterns of information demand reflect non-linear subjective perceptions of probabilistic rewards that have been documented in multiple contexts.
The idea that people and animals have non-linear subjective estimates of both outcome value (magnitude) and probability is supported by decades of theoretical and empirical research34,35,36,37,38,39,40,41,42,43. These studies have shown that the nonlinearities depend on both the probability and valence (gain or loss) of an outcome38, may be traced back to biological substrates34,35,36, and may explain complex and apparently non-normative patterns of risk attitudes and probabilistic reasoning. Value weighting functions are typically concave for gains and convex for losses, with the function being steeper for losses than gains. This asymmetry can explain why participants show higher sensitivity to losses than gains, as well as diminishing sensitivity to incremental increases in the magnitudes of both outcomes37,38,39,40,41,42. Similarly, probability weighting functions (PWFs) are non-linear, typically showing that individuals overestimate small probabilities and underestimate large ones. This pattern accounts for behaviors such as purchasing lottery tickets (gambling on low-probability gains) and buying insurance (protecting against low-probability losses)38,43. Because the decision-theoretic definition of “information” is in terms of risk and uncertainty (as quantified, for example, by the misclassification rate or the Shannon entropy of a distribution of future states44,45), non-normative patterns of information demand may be closely related to subjective probabilistic perception and risk attitudes, but this possibility has yet to be empirically investigated.
Here, we tested and found strong support for the hypothesis that models explicitly incorporating classical distortions in probabilistic reasoning outperform models based on linear mixtures of motives. Using behavioral testing and computational modeling, we show that this finding holds for both non-instrumental and instrumental information demand. We further demonstrate that individual non-linearities in subjective perceptions are shared across information-demand and risky-choice tasks, and are associated with personality metrics of Stress Tolerance and Need for Cognition. Thus, non-normative patterns of information demand that have previously been attributed to hedonic motivations are better explained by subjective perceptions of probabilistic rewards, linking models of information demand with broader theories of risk and probabilistic reasoning.
Participants were recruited through the online platform Prolific46 between May and December 2024, following experimental procedures approved by the Institutional Review Board of Columbia University. All participants provided informed consent prior to their participation. Based on preliminary power analyses (see Supplementary Note 2), we aimed to collect 50 datasets in Exp. 1 and 2 and 150 datasets in Exp. 3.
We implemented several prospective and post-hoc measures to ensure data quality. Prospectively, we restricted enrollment to individuals who: (a) were 18–40 years old, (b) resided in the United States, (c) were fluent in English, and (d) had completed at least 20 prior Prolific submissions with a ≥ 90% approval rate. Further, we did not proceed to data collection for participants who met enrollment criteria but failed comprehension and attention checks during the pre-task instruction stage. In post-hoc analyses, we excluded participants whose variance across the trial-by-trial ratings differed from the average variance of the entire participant sample by more than three standard deviations in either Gain or Loss trials in any experiment. This criterion was independent of the strategies we examined, instead reflecting the fact that excessively low or high variance suggests, respectively, stereotyped or random responding that hinders hierarchical modeling. Post-hoc screening resulted in exclusion 9/59, 7/57, and 29/179 of participants in Experiments 1, 2, and 3, respectively. The final samples we analyzed comprised 50 participants in Exp. 1 (22 female, 27 male, 1 other, average age 31.96 years, SD = 5.09, range 21–40), 50 participants in Exp. 2 (20 female, 30 male, average age 30.78 years, SD = 5.95, range 20-40) and 150 participants in Exp. 3 (70 female, 77 male, 3 other, average age 30.99 years, SD = 5.43, range 18-40). Gender was self-reported and no data were gathered on race or ethnic identity.
Stimuli were generated and presented using the jsPsych 6.3.1 library47, and were presented during Instruction, Task and Questionnaire phases of the experiment. The study was not formally preregistered.
Before starting each task, participants viewed detailed instructions outlining the task structure and illustrating example trials. They then answered a series of comprehension questions to confirm they had understood the instructions. If they failed a comprehension question, they received feedback and repeated the question. A second incorrect response resulted in termination of the experiment. Participants who continued then completed five practice trials and, finally, received the option to continue to the Task stage or exit the experiment (self-exclude) if they remained uncertain about any aspect of the task.
In each task, participants completed two blocks of Gain trials and two blocks of Loss trials that were presented in interleaved order and preceded by a screen announcing the type of the block. Each block had 20 trials, including two repetitions of nine possible mixtures (e.g., fish:seaweed or plain:colored mushroom ratios of 9:1, 8:2, 7:3, 6:4, 5:5, 4:6, 3:7, 2:8, 1:9), presented in a pseudorandom sequence in which identical lotteries did not appear on consecutive trials. Two additional trials contained only one item (10 fish or 10 seaweed in the information demand tasks, and 10 plain mushrooms or 10 colored mushrooms in the risky choice task) and were used to encourage participants to use the full scale—but were not used in subsequent analyses or plots. Given evidence that probabilistic reasoning is sensitive to presentation format48,49, we depicted the lotteries in each task as both pictorial mixtures of fish and seaweed and pie charts listing the percentages and monetary values of each item (Figs. 1A, 2A and 3A). In the instrumental task of Exp. 2 and 3, an additional graphic in the top-right corner reminded participants that they could receive an additional $2 for correctly reporting the outcome (Fig. 2A). In the risky choice task of Exp. 3, a similar graphic reminded participants that they could enter the lottery or receive a guaranteed value in the range of the outcomes (($0to $5) in gain trials, (-$5to 0) in loss trials; Fig. 3A).
A Trial Display. Each trial presented a pond containing 10 items (fish or seaweed), the proportion of which governed the probability of a monetary outcome, and was additionally indicated by a percentage/pie-chart. The outcome depended on a random draw from the pond, and participants used a slider to indicate their curiosity to receive advance information regarding the draw. B Gain and Loss trials were presented in randomized order and had various mixtures of seaweed and gold fish (respectively, $0 or +$5) or seaweed and red fish ($0 or −$5). C Payout Stage. One trial was selected from the 80 that participants viewed, and its outcome determined the final payout (a base pay of $10 plus the lottery outcome of $0, $5 or -$5). Participants were instructed that they had a chance of obtaining early information regarding the outcome (versus receiving it only at the end of the session after completing personality rating scales). To ensure questionnaire responses were not affected, no participant received early information. D Observed and predicted curiosity ratings. Ratings for Gain (orange) and Loss (red) lotteries versus the probability of the outcome. The left panel shows the empirical data, and the middle and left panels show the model predictions using the best fitting parameters of the mixed-motive (MM) and full probabilistic reasoning (PR) model. All values are mean and SEM over N = 50 participants after z-scoring within participants. E Model comparisons. Left: WAIC score differences relative to the full probabilistic reasoning model (mean and SEM over 50 participants); *p < .01 **p < .001 improved fit versus the best mixed-motive model. Right: Individual Akaike weights, ranging from 0.5 (center thin dotted line) to 1.0 toward the right (orange; favoring mixed-motive) or left (blue; favoring full probabilistic reasoning). Numbers show percentages, out of 50 participants, in a weight range.
A Participants viewed the same lotteries as in Exp. 1 but, different from Exp. 1, could also obtain a $2 instrumental reward for correctly reporting the outcome and reported their willingness to pay for information before reporting their guess. B Gain and Loss lotteries were identical to those in Exp. 1. C Payout Stage. A random trial was selected for payout, and participants did or did not see its outcome according to a BDM auction. Participants then reported their best guess about the lottery outcome, completed a ~ 15-min questionnaire battery and, finally, received the payout consisting of the non-instrumental and instrumental outcomes and the BDM price if information was given (see text for details). D Observed and Predicted Curiosity Ratings. Average z-scored ratings observed in participants (left) and predicted by the mixed-motive (MM; center) and full probabilistic reasoning (PR; right) model. All points show the mean and standard error over N = 50 participants. Dashed gray lines indicate the normative response based on the extent to which information would improve classification accuracy; all other conventions as in Fig. 1D. E Model comparisons. Left: WAIC score differences relative to the full probabilistic reasoning model (mean and SE over 50 participants); **p < .001 improved fit versus the mixed-motive model. Right: Individual Akaike weights. All other conventions as in Fig. 1E.
The Risky Choice task cartoon shows (A) the task lotteries/threshold elicitation, (B) Gain and Loss lotteries, and (C) Payout procedure in the same format as in Fig. 2 A-C. See text for further details. D Willingness to pay for information for the 150 participants in Exp. 3, in the same format as in Fig. 2D (left). E Thresholds in the Risky Choice task as a function of probability for Gain and Loss lottery. The dotted gray line indicates the normative response. Data show the mean and SEM for the same participants as in D. F Average PWFs for Gain (left) and Loss (center) lotteries, and average Value Function (right). Each plot shows the normative identity line (thin dashed traces; defined as (w(p)=p) for probability-weighting and (v(x)=x) for the value function), and the function derived from the information demand task (solid lines) and from the risky-choice task (dotted lines). G Scatterplots for the free parameters in the two tasks. Each point indicates one participant. All parameters are shown in log-space as they were estimated using log-normal hyperparameters. Each plot shows the Pearson correlation coefficient (r) and its p-values, and solid lines denoting the best-fit linear regressions.
After viewing a lottery, participants were required to indicate their desire for advance information by moving a slider at least once before clicking a “Continue” button to progress to the next trial. In all experiments, the initial slider position was randomized across trials and no time limit imposed on responses. In the non-instrumental task of Exp. 1, the slider indicated the desire for advance information (“curiosity”) on an arbitrary scale. To incentivize participants to use the full scale, they were instructed that the probability that they would receive information depended on the trial’s curiosity rating relative to the other ratings they gave. Specifically, we used verbal and graphical instructions to illustrate how responding with high curiosity to all trials would result in an equal probability of receiving information for any selected trial, while spreading out their curiosity ratings would result in a higher likelihood of receiving information about high-curiosity trials and lower likelihood for low-probability trials.
In the Exp. 2 and 3, the slider indicated monetary values corresponding to willingness to pay for information (in the instrumental information demand task) or minimum acceptable guaranteed value (in the risky choice task), each with labeled boundaries at the minimum and maximum response options. An additional numerical label updated as participants moved the slider, indicating its current position in 1-cent increments.
After completing the lottery trials, participants proceeded to the payout step described in Figs. 1C, 2C and 3C, in which a single trial determined their bonus. Although participants were instructed that the bonus trial would be selected at random from the 80 they played, this trial was in fact constant across all participants in a task. This manipulation was not announced and could not affect task behavior. However, it was necessary to ensure that our participants did not receive differentially valenced information (which may affect their mood for the subsequent Questionnaires; see below), and to guarantee that the average payment was at the higher end of Prolific guidelines, with no participant receiving less than the minimum required amount. Specifically, in Exp. 1, all participants were assigned a bonus trial with a 30% chance of a positive outcome and 70% chance of a neutral one. Each participant was awarded the neutral outcome and did not receive early information about the result. In Exp. 2 and 3, the bonus trial selection and outcomes were fixed (gain trial with 30% chance of winning, fixed to deliver a neutral outcome) based on the same rationale as above. To minimize the likelihood of early information affecting subsequent questionnaire responses, we fixed the price of information at $1.75 in the instrumental task (ensuring only a minority of participants – 8 out of 150 – bid enough to view it), and ensured that the outcome of the risky choice task was not delivered until the end of the experiment. This manipulation was not disclosed to participants, to avoid influencing their behavior.
In the instrumental information demand task, the probability of information delivery was governed by Becker–DeGroot–Marschak (BDM) auctions eliciting incentive-compatible bids50. In this auction procedure, participants were instructed their bid would be compared against a random price drawn from a uniform distribution ranging between $0 and $2 in 1-cent increments. If the drawn price exceeded the bid, that participant paid nothing and had to guess the identity of the item without advance information. If the drawn price was less than the bid, participants were shown the outcome before making the instrumental report and the drawn price was deducted from their payout. In the risky-choice task (Exp. 3), a BDM auction decided whether participants received a guaranteed payout or the lottery outcome. After reporting their threshold value (minimum acceptable guaranteed amount), a random value would be drawn from a uniform distribution between $0 and $5 (in 1-cent increments). If the drawn value was below the participant’s threshold, they received that fixed value; if it was equal to or above their threshold, they instead played the lottery and received its outcome.
In Exp. 1, all participants received a payment of $10. In Exp. 2, participants received between $9.25 and $12.00 (mean $11.55), depending on accuracy of their guess and the price they paid for information. In Exp. 3, participants received between $11.25 and $14.00 (mean, $13.53), depending on their thresholds and BDM outcomes on the risky choice task.
After completing the tasks of non-instrumental and instrumental information-demand, participants completed 5 questionnaires presented in a randomized order: Need for Cognition (NFC)51, the extraversion subscale of the Big-5 Personality Inventory52, the Thrill Seeking and Stress Tolerance Subscales of the Five-Dimensional Curiosity (5DC) scale53, and the Patient Health Questionnaire-4 (PHQ-4)54. For each questionnaire, instructions appeared in large font at the top of the screen (Supplementary Fig. 13). Participants selected their response by clicking the corresponding bubble; no time limit was imposed. The questionnaires were administered in Exp. 1 and 2 for procedural consistency and to induce a delay before the outcome of the lottery was known, but only responses from Exp. 3 – where we achieved a sample size sufficient for questionnaire analyses – were included in the primary analyses.
Questionnaire scores were computed following published scoring rules – reverse-coding items where specified – and were linearly rescaled to a 0–10 range for comparability across measures. To examine the relationship between questionnaire scores and performance, we computed three metrics based on the mean-absolute deviations (MAD) from normative behavior as explained in the Results section. We then log-transformed the MAD scores to approximate normal distributions and minimize the impact of extreme outliers, z-scored the MAD metrics and questionnaire scores across participants and, finally, fit multiple linear regression models predicting each z-scored MAD measure from the five standardized questionnaire scores.
For each task, we fit information-seeking behavior with models belonging to a mixed-motive or probabilistic reasoning family, and hybrid models integrating features of both. Here we provide an overview and summary, and in Supplementary Notes 1 and 2 we give full details of our modeling strategy.
In the mixed-motive models, information value was fit as a weighted sum of two factors, which differed for non-instrumental and instrumental conditions. In the non-instrumental condition (Exp. 1), the factors were the hedonic value and the cognitive value (uncertainty) of the lottery. Hedonic value was the expected value of the lottery (i.e., the product of outcome magnitude; +$5 or −$5) and outcome probability (the fraction of fish in the pond). Uncertainty was parametrized in two ways – as the misclassification rate (the probability of incorrectly guessing the outcome without information), and the Shannon entropy of the lottery.
In the instrumental condition, the factors in the mixed-motive model were hedonic value and instrumental value. Hedonic value was defined as in Exp. 1 based on the non-instrumental value of the lottery (i.e., +$5 or −$5) multiplied by the probability of drawing that outcome. Instrumental value was defined as the change in the expected instrumental reward with versus without the information (i.e., the reward for a correct guess; $2, multiplied by the probability of guessing correctly). The probability of guessing correctly was in turn equal to 1 minus the misclassification rate (i.e., 1 minus the uncertainty of the lottery). Importantly, this meant that instrumental and cognitive value were perfectly positively correlated and could not be separately estimated in the mixed-motive approach. In Exp. 2 and 3, therefore, the mixed-motive model captured the potential effects of cognitive and instrumental value with a single parameter measuring the instrumental benefits – the change in the misclassification rate expected from obtaining advance information.
In contrast with the mixed-motive model, the probabilistic-reasoning models predicted information demand from only the uncertainty or instrumental value of the information. However, rather than being parametrized based on their objective (true) values, each value was calculated from subjective perceptions of reward probability and magnitude. Across both tasks, we estimated subjective reward probabilities using the Prelec probability weighting function (PWF)55, with parameters (alpha) (curvature) and (beta) (elevation). To capture valence-dependent effects, we first employed distinct PWFs for gain and loss trials, estimating separate (alpha) and (beta) parameters for each. Second, we incorporated (lambda), a parameter capturing Prospect-theoretic loss aversion, which modified the relative sensitivity to gains and losses. For the instrumental information demand task, we modeled subjective reward magnitudes using the full Prospect-theoretic value function38. To achieve this, we combined (lambda) with (gamma), a parameter that captured decreasing sensitivity to larger rewards.
In addition to the pure mixed-motive/probabilistic reasoning models, we evaluated hybrid models that contained the weighted sums of the mixed-motive models and the PWF/value functions of the probabilistic reasoning models as described in the text.
In the risk task of Exp. 3, we modeled participants’ threshold responses using the same non-linearities as in the probabilistic reasoning models for instrumental demand.
Finally, prior to conducting linear regression analyses on the questionnaire results, all necessary assumptions were formally assessed. The normality of standardized residuals was evaluated through visual inspection of Normal Q-Q plots and confirmed using the Shapiro-Wilk test. Homoscedasticity was verified via visual examination of scatter plots of standardized residuals versus fitted values.
Participants were recruited on the online platform Prolific and completed a task in which they rated their curiosity about non-instrumental losses or gains. In each trial, participants were presented with a lottery depicted as a pond containing a mixture of fish and seaweed and a pie-chart reinforcing the probabilities and rewards of each item (Fig. 1A). Participants were instructed that their bonus at the end of the session would depend on one item that was randomly drawn from the pond, such that seaweed did not change the base payout of $10, a gold fish increased the payout by $5, and a red fish decreased the payout by $5. Participants viewed 80 lotteries of this kind that differed in the ratios of fish to seaweed and in whether they contained gold fish (signaling a possible Gain) or red fish (signaling a possible Loss; Fig. 1B). Each lottery outcome was determined purely by chance, but participants rated their curiosity, or desire to obtain advance information regarding the outcome (Fig. 1A). Participants were instructed that, at the end of the session, their bonus will be based on a single trial that was randomly selected from the 80 they had completed (Fig. 1C). This random-lottery system has encourages participants to treat each trial as meaningful rather than regarding the entire experiment as a meta-lottery56,57,58,59 – which we further reinforced with specific instructions to report curiosity as if the trial were selected for payment and using the full scale (Methods).
Consistent with previous findings, curiosity ratings varied with both the uncertainty and valence of the lottery14,18 (Fig. 1D, left). The average ratings (N = 50 participants who passed attention checks; Methods) peaked for lotteries that had maximal uncertainty (outcome probability of 0.5). In addition, the ratings showed an interaction between valence and probability, such that curiosity was higher for gains versus losses at probabilities above 0.5 but reversed to be lower for gains versus losses at probabilities below 0.5—a probability-dependent reversal that was also noted in prior reports7,14.
To understand the mechanisms producing this pattern, we compared two computational models describing behavior based either on a mixture of motives or a non-linear weigthing of probabilities. The mixed-motive model predicted the ratings based on the weighted sum of an uncertainty (cognitive) term and an expected value (hedonic) term, with the weights of each term specified by participant-specific free-parameters. Conversely, the probabilistic reasoning model used only an uncertainty term, but allowed for non-linear distortion in the probabilities that determined this value. For the full probabilistic reasoning model, we used independent Prelec probability-weighting functions (PWFs) for losses and gains, each characterized by a curvature ((alpha )) and elevation ((beta )) parameter, with an additional multiplicative parameter ((lambda )) capturing loss aversion (see Methods and Supplementary Note 1 for full model specifications). Extensive simulations determined that the models were identifiable with high accuracy and all parameters could be reliably recovered given a 50-participant sample (see Supplementary Note 2, Supplementary Tables 3 and 4).
Hierarchical Bayesian modeling (Methods) showed that the probabilistic reasoning model outperformed the mixed-motive model in fitting the participants’ data. Importantly, while both models reproduced the peak in curiosity ratings at probability of 0.5, the mixed-motive model failed to capture the interaction between valence and probability shown by the data. Because the mixed-motive model captures valence effects using an additive linear term, it predicts that the strength of the effects may increase with reward probability (expected reward), but their sign (whether curiosity is higher for gains or for losses) is constant across all probabilities (Fig. 1D, center).
In contrast, the full probabilistic reasoning model successfully captured the interaction between valence and probability using nonlinear PWFs that differed for losses and gains (Fig. 1D, right). Consistent with prior literature37,38,39,40,41,42,55,60, the best-fit PWF curves were S-shaped, overweighting low probabilities and underweighting high probabilities, and had higher elevation for losses than gains (Supplementary Fig. 5A). These non-linearities imply that, for low outcome probabilities, a loss is perceived as being more uncertain (its probability is perceived as closer to 0.5) relative a gain, resulting in higher curiosity for losses than gains. In contrast, for higher probabilities the relationship is reversed, with a possible gain being subjectively more uncertain and evoking higher curiosity relative to a possible loss.
Quantitative comparisons confirmed the advantage of the full probabilistic reasoning model in reproducing the data. The average difference in the Widely Applicable Information Criterion (ΔWAIC)61 was positive for the mixed-motive versus the full probabilistic reasoning model, confirming that the probabilistic reasoning model provided a superior fit even after accounting for its higher complexity (Fig. 1E, left: ({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}}=857pm 76.3) [SE]; p < .001). Moreover, analysis of individual Akaike weights62 showed that 86% of participants were better fit by the full probabilistic reasoning than the mixed-motive model, with 80% strongly favoring the former model (with a likelihood above 0.75; Fig. 1E, right).
A possible explanation of the low mixed-motive model performance was that, in contrast with the rounded peak in curiosity in the data, the model predicted a sharp peak at 0.5 due to our use of misclassification rate as the measure of uncertainty (Fig. 1D). However, re-estimating the mixed-motive model using Shannon entropy44 resulted in significantly worse mixed-motive model fit (Fig. 1E; ({{{rm{WAIC}}}}_{{{rm{MM}}}[{{rm{Shannon}}}]}-{{{rm{WAIC}}}}_{{{rm{MM}}}}=173pm 30.6) [SE]; p = .018), despite producing a rounded peak that was more similar to the data (Supplementary Fig. 6A). This result is consistent with a prior report showing that the misclassification rate in a two-alternative task explains human information demand better than several alternative metrics45, and suggests that this conclusion extends to non-instrumental information demand. Importantly, although using Shannon entropy also slightly reduced the fit quality for the probabilistic reasoning model, it did not alter the model’s clear superiority over the best-fitting mixed-motive model (Fig. 1E; ({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Shannon}}}]}-{{{rm{WAIC}}}}_{{{rm{MM}}}}=844pm 73.7) [SE]; p < .001). Thus, the decisive advantage of the probabilistic reasoning over the mixed-motive model did not depend on the precise uncertainty metric we used (Fig. 1E; Supplementary Fig. 6).
Additional analyses demonstrated that the probabilistic reasoning model’s advantage persisted even when its complexity was reduced. Specifically, this advantage was maintained when the model used only one mechanism to generate differential behavior on gain and loss trials – either a loss aversion parameter ((lambda)) or distinct (alpha) and (beta) parameters for gains and losses. Simplified probabilistic reasoning models that omitted loss aversion (PR[No Loss Aversion]) or included it but used a single PWF function for losses and gains (PR[Single PWF]) decisively outperformed the mixed-motive model (Fig. 1E; ({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{NoLossAversion}}}]}=654pm 74.4) [SE]; p < .001; ({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{SinglePWF}}}]}=355pm 62.0) [SE]; p = .008). Importantly, both reduced versions significantly underperformed the full probabilistic reasoning model (Fig. 1E; ({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{NoLossAversion}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=191pm 28.8) [SE]; p = .002; ({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{SinglePWF}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=490pm 58.8) [SE]; p < .001), showing that the data justified the inclusion of all parameters. Examination of the models’ predictions showed that the simplified versions reproduced the interaction between valence and probability but did not capture it as completely as the full model (Supplementary Fig. 6).
Finally, we tested two hybrid models that combined the mixed-motive and probabilistic reasoning strategies, by modeling information demand as a linear combination of hedonic value and cognitive value, while also transforming the probabilities using non-linear probability weighting as in the probabilistic reasoning model. When fit to empirical data, these hybrid models showed severe convergence failure (high (hat{R}) and low effective sample size across key individual- and group-level parameters). These issues persisted irrespective of whether the model included a single PWF or separate PWFs for gains and losses (respectively, MM[Single PWF] and MM[Dual PWF]; see Supplementary Table 6). Posterior pairs plots of the group-level hyperparameters (Supplementary Figs. 7 and 8) indicated that the underlying cause of this failure was weak identifiability due to trade-off degeneracy. That is, the hybrid models were able to produce similar patterns of results with different combinations of parameter values, suggesting they are ill-specified to the data.
In sum, the full probabilistic reasoning model provides the best fit to the data and has a decisive advantage over the best-performing mixed-motive model in fitting non-instrumental information demand in our task.
To assess whether the probabilistic reasoning model also explained instrumental information demand, we recruited a new cohort of participants to perform Exp. 2 (Fig. 2). Participants in Exp. 2 faced the same non-instrumental gain and loss lotteries as in Exp. 1 (Fig. 2A, B) but were informed that, in addition to the random non-instrumental reward they obtained from the lottery, they could earn an instrumental reward ($2) if they correctly reported the non-instrumental reward (Fig. 2A). Instead of rating their curiosity, participants placed monetary bids indicating their willingness to pay (WTP) for advance information about the lottery outcome (Fig. 2A). A randomly selected trial determined the payout as in Exp. 1, but in this case, participants received (or did not receive) the information based on a Becker–DeGroot–Marschak (BDM) auction, which incentivizes truthful WTP bids as if each trial had been selected for payout50 (Methods). Participants then reported the lottery outcome, completed a questionnaire battery, and received their payout, which comprised their base compensation ($10), non-instrumental lottery outcome ($0, +$5, or −$5) and instrumental reward ($2 if the outcome report was correct; $0 otherwise), from which the BDM price was subtracted if the participant received information (Fig. 2C, Methods).
Because only the instrumental component of participants’ bonus could be affected by advance information, the optimal strategy was to bid based on the extent to which information could improve the accuracy of the instrumental report. Instrumental value, in turn, was equal to the instrumental reward ($2) multiplied by the misclassification rate (the probability of making an error without information; see Methods, Data analysis and Computational modeling) and was an inverted-V shaped function of probability for both losses and gains, as shown by the thin dotted lines in Fig. 2D. Participants’ bids generally mirrored this normative pattern but deviated systematically, undershooting optimal bids at maximal uncertainty and overshooting them at lower uncertainty (Fig. 2D, left; colored traces).
To formally assess behavior in Exp. 2, we fit the bids with mixed-motive and probabilistic reasoning models adapted to reflect the instrumental contingencies (Methods, Data analysis and Computational modeling). The mixed-motive model included an instrumental component alongside a hedonic component related to the expected value of the lottery. The probabilistic reasoning model had an instrumental component with non-linear PWFs, which were specified as in Exp. 1 except for an additional free parameter representing value weighting ((gamma)) capturing the larger variability in the possible bonus amounts (see Supplementary Note 1 for full description of modeling approach).
Although both models reproduced the peak in bids at outcome probability of 0.5, only the probabilistic reasoning model captured the systematic underbidding and overbidding observed in the data (Fig. 2D). Quantitative comparisons confirmed that the probabilistic reasoning model was superior to the mixed-motive model after accounting for its higher complexity (Fig. 2E, left; ({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=1004pm 79.1) [SE]; p < .001). Analysis of individual Akaike weights showed that 72% of participants were better fit by the full probabilistic reasoning model compared to the mixed-motive model, with 68% showing strong support for the full probabilistic reasoning model (weight >.75; Fig. 2E, right). The best-fit PWFs suggested that, for both losses and gains, participants over-estimated uncertainty at both the high and low ends of the probability scales, motivating them to over-bid in these ranges (Supplementary Fig. 5B).
As in Exp. 1, we found that several simplified variants of the probabilistic reasoning model also far outperformed the mixed-motive model but underperformed the full model (Fig. 2E). The mixed-motive model was overperformed by a simpler probabilistic reasoning model that was constrained to use a single Prelec function for losses and gains (({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{SinglePWF}}}]}=815pm 78.0) [SE]; p < .001), and by a model that used separate PWFs for losses and gains, but omitted value weighting (({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{NoVW}}}]}=694pm 60.9) [SE]; p < .001). However, both models were overperformed by the full probabilistic reasoning model (({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{SinglePWF}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=189pm 38.0) [SE]; p < .001; ({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{PWonly}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=310pm 57.3) [SE]; p < .001), showing that all free parameters improved goodness of fit. As expected, a model that excluded PWF and used value weighting alone strongly underperformed the worst probabilistic reasoning model, confirming that non-linear PWF were essential to this modeling framework (({{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{NoPWF}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{NoVW}}}]}=978pm 74.9) [SE]; p < .001).
Finally, we attempted to fit hybrid models in which bids were predicted by a mixed-motive model that also included non-linearities in subjective probability and reward estimates. As in Exp. 1, we found that even the simplest such model that used a single PWF displayed severe convergence failures across a range of parameters (high (hat{R}) and low effective sample size; Supplementary Table 6) and trade-off degeneracy (Supplementary Fig. 9). An mixed-motive model that included value weighting but no PWF produced a marginal improvement over the standard mixed-motive model (Fig. 2E; ({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{MM}}}[{{rm{VW}}}]}=58.7pm 24.2) [SE]; p = .067) but remained significantly worse than all probabilistic reasoning models (({{{rm{WAIC}}}}_{{{rm{MM}}}[{{rm{VW}}}]}-{{{rm{WAIC}}}}_{{{rm{PR}}}[{{rm{Full}}}]}=945pm 77.3) [SE]; p < .001).
Intriguingly, the best-fit PWFs in Exp. 2 appeared closer to the veridical, linear values relative to those in Exp. 1 (see Supplementary Figs. 5A and 5B) suggesting that probability estimates may be less distorted when instrumental rewards are at stake. Indeed, we found that PWF parameters were slightly but significantly closer to their normative values in instrumental versus non-instrumental conditions for both losses and gains. For gains, PWF curvature was significantly smaller (({alpha }_{{{rm{Gain}}}}) values were closer to 1) for the instrumental versus non-instrumental condition (t-tests across tasks: (log {alpha }_{{{rm{Gain}}}}) t(96.14) = 2.13, p = .035, 95% CI = [0.02, 0.44], Cohen’s d = 0.43; ({{rm{abs}}}(log {alpha }_{{{rm{Gain}}}})), t(95.58) = 2.38, p = .019, 95% CI = [−0.35, −0.03], Cohen’s d = 0.48), with no significant change in the elevation parameters (t-test of (log {beta }_{{{rm{Gain}}}}): t(89.83) = 0.65, p = .515, 95% CI = [−0.13, 0.07], Cohen’s d = 0.13). For losses, PWF elevation was significantly lower (({beta }_{{{rm{Loss}}}}) values were closer to 1) in Exp. 2 versus Exp. 1 (t-tests of (log {beta }_{{{rm{Loss}}}}): t(81.92) = 3.48, p < .001, 95% CI = [0.08, 0.30], Cohen’s d = 0.70; ({{rm{abs}}}(log {beta }_{{{rm{Loss}}}})): t(75.62) = 3.32, p = .001, 95% CI = [-.021, -0.05], Cohen’s d = 0.66), with no significant differences in curvature (t-test of (log {alpha }_{{{rm{Loss}}}}): t(96.64) = 0.22, p = .792, 95% CI = [-0.22, 0.17], Cohen’s d = 0.05).
In sum, despite slight improvements in instrumental versus non-instrumental conditions, participants’ bids deviated from optimal values even in instrumental conditions. As we found for non-instrumental demand, their behavior was better explained by the probabilistic reasoning than the mixed-motive model.
In the final experiment, we sought to replicate the findings of Exp. 2 in a larger participant cohort and test the hypothesis that information demand and risky decisions can be captured in a common framework. To this end, we recruited a new cohort of participants to perform the same instrumental information demand task as in Exp. 2 alongside an equivalent risky-choice task. The two tasks were presented in blocks in an order counterbalanced across participants, and we analyzed data from 150 participants who passed quality checks for both tasks (Methods). Replicating our findings in Exp. 2, participants’ bids on the information demand task showed modest departures from normative values (Fig. 3D) and were substantially better fit by the full probabilistic reasoning model compared to the mixed-motive model (({{{rm{WAIC}}}}_{{{rm{MM}}}}-{{{rm{WAIC}}}}_{{{rm{PR}}}}=1928pm 136.2) [SE]; N = 150, p < .001). Individual Akaike weights favored the probabilistic reasoning model in 77.3% of participants, with 72.7% showing likelihoods above 0.75 independent of task order (Supplementary Fig. 10A, B).
The critical question was whether similar subjective perceptions of probabilistic rewards that accounted for information demand also accounted for risk preferences in a risky-choice task. To test this hypothesis, the same participants completed a risky-choice task in which they made choices based on lotteries that were numerically identical to those testing information demand, but were recontextualized as fields containing colored mushrooms rather than ponds to prevent confusion across the two tasks (Fig. 3A, B). Rather than bidding for information, participants were asked to report the minimum guaranteed amount they would accept in place of a randomly drawn item (their “threshold” value for accepting the lottery gamble; Fig. 3A). At the end of the session, one field was randomly selected, and participants received a payout equal to their base pay ($12) plus either the value of a randomly drawn item ($5, $0, or –$5) or a fixed amount determined through a BDM auction based on their threshold report (Fig. 3C).
The BDM auction incentivized participants to report the lowest guaranteed value they would accept in lieu of a random draw from the lottery50 which, for a normative, risk-neutral agent, is the expected value of the lottery (Fig. 3E, thin dotted lines). Participants’ thresholds showed deviations from this benchmark that depended on both the probability and valence of the lottery (Fig. 3E) and were fit by S-shaped PWFs (Fig. 3F; see Supplementary Table 5 for full statistics of estimated parameters). Consistent with previous work38,43 the best-fit PWFs suggested that, when outcome probabilities were low, participants overestimated the probability of a gain or a loss. Consequently, participants tended to place their thresholds higher than the normative response for low-probability gains, and lower for low-probability losses (Fig. 3E). Conversely, at high probabilities of an outcome, participants underestimated the probabilities of a gain or a loss, resulting in a reversal in which their thresholds were lower than normative for gains and higher for losses (Fig. 3E). We also note that, while bids in the instrumental information task were centered in the lower half of the scale (see Fig. 3D and Fig. 2A), thresholds in the risky-choice task were centered at the midpoint of the scale, ruling out sensorimotor biases toward a particular region on the response scale.
Importantly, we hypothesized that, if the same individual-level perceptions of value and probability explained inter-individual variability in both tasks, the best-fitting probabilistic reasoning model parameters should be comparable and correlated in the risky choice and information demand tasks. Consistent with this hypothesis, the average PWF for both gains and losses (Fig. 3F, left and center, respectively) and the value weighting functions (Fig. 3F, right) were very similar across the two tasks. Crucially, individual-level analyses showed that all the parameters were significantly correlated between the two tasks, including ({alpha }_{{gain}},{beta }_{{gain}},{alpha }_{{loss}},{beta }_{{loss}}gamma ,{{rm{ and}}} lambda) In Fig. 3G we show the results in log-space, consistent with the log-normal hyperpriors used in model fitting; though the results also held for raw parameter estimates (Supplementary Fig. 11) and were robust to task order (Supplementary Fig. 10C). As additional verification, we conducted across-task analyses in which we simulated behavior on one task using the best-fit parameters from the other. The pattern of bids on the information demand task was reproduced when using the best-fit parameters from the risky-choice task and conversely, the pattern of thresholds on the risky choice task was reproduced when using best-fit parameters from the information demand task (Supplementary Fig. 12). Together, these findings support the hypothesis that shared patterns of subjective probability perception underlie behavior in both information-demand and risky-choice contexts.
Given the consistency of individual parameters across the information demand and risky choice tasks, we next asked whether these parameters were associated with personality traits. We focused on traits previously linked to information demand13,63,64,65—namely, Need for Cognition (NFC; the tendency to engage in intellectually challenging activities), the Extraversion subscale of the Big-5 Personality Inventory52, the Thrill Seeking and Stress Tolerance Subscales of the Five-Dimensional Curiosity inventory53, and the Patient Health Questionnaire-454 (PHQ-4; a brief measure of symptomology associated with anxiety and depression). For each participant, we extracted three mean absolute deviation (MAD) metrics summarizing their tendency to deviate from normative behavior, regardless of the direction of the deviation: (1) the behavioral MAD was the magnitude of the deviation between each participant’s bid or threshold response and the response that maximized expected payoff, (2) the probability-weighting MAD was the average absolute deviation of the fitted Prelec probability-weighting function from a linear (identity) mapping over the full (0to 1) probability range and (3) the value-weighting MAD was the average absolute deviation of the fitted power utility function from a linear value function across the range of possible outcomes ((-$5to 5)). After confirming that questionnaire scores were not excessively intercorrelated (Supplementary Fig. 14), we fit 3 linear models predicting each MAD metric from all questionnaire scores across the 150 participants (using z-scored MAD values and questionnaire scores rescaled to a 0-10 range to enable comparison of coefficient magnitudes).
Two scales emerged as consistent predictors across the two tasks (Fig. 4). NFC had significant negative associations with the MAD metrics, suggesting that higher NFC scores were associated with more normative estimates (smaller MAD values) in both the information-demand and risky-choice tasks. In addition, Stress Tolerance scores had positive associations for all metrics suggesting that higher stress tolerance was associated with less adherence to normative patterns in both tasks. With the sole exception for the MAD from the normative bid for the information-demand task, which fell just short of reaching significance, all the coefficients for the NFC and Stress tolerance scores were significant for all metrics, highlighting the robustness of the associations across the risky choice and information-demand tasks (Fig. 4). We found no consistent associations between personality scores and signed (rather than absolute) deviations from the normative PWF, confirming that the personality scores predicted the deviation from normative strategies rather than specific biases toward placing high or low bids/thresholds.
Each personality score was entered as regressor explaining three performance metrics based, respectively, on behavior, linear probability-weighting, and linear value-weighting. The resulting coefficients are color-coded as indicated in the title. Error bars show 95% confidence intervals. Filled points/thick lines indicate p < .05 (the CI does not include 0).
Using behavioral testing and computational modeling, we provide evidence that certain features of information demand that have been attributed to mixtures of motives7,14,16,18,26,27 are better explained by subjective perceptions of value and probability. We show that an established mixed-motive model predicated on linear mixtures of expected value, uncertainty and instrumental rewards, was outperformed by a model that accounts for non-linear subjective perceptions of probabilistic rewards. Inter-individual variations in subjective perceptions were correlated with the same non-linearities estimated from risky choice preferences and with the personality traits of Need for Cognition and Stress Tolerance. The results highlight the importance of adequately parametrizing models and their functional forms, and help explain the mechanisms of information demand in the framework of decisions under uncertainty.
Our study extends decades of decision research that has held that non-linear weighting of probabilities and rewards explains non-normative risky-choice behavior43 by showing that this framework can also explain information demand. Specifically, we show that by using a non-linear parametrization of subjective probability, a probabilistic reasoning model captures complex interactions between valence and probability that are not explained by the mixed-motive approach. In the mixed-motive approach, valence effects are parametrized as a linear-additive term. This enables the framework to account for differential information demand for gains versus losses, but restricts it to predict that the direction of any such preference would be constant across all outcome probabilities. In the prospect-theory framework, in contrast, valence modifies non-linear perceptions of probability, producing complex interactions between valence and probability. This gives the probabilistic reasoning model unique power to reproduce information demand that is stronger for gains versus losses at high probabilities but reverses to be higher for losses than gains at low probabilities of an outcome, as seen in the data. Thus, while both the mixed-motive and probabilistic reasoning models account for the effects of valence, uncertainty and instrumentality on information demand, the non-linear interactions formalized by the latter are far superior in predicting the complexity of the empirical data.
A salient feature of our results is that the interaction effects in information demand and the non-linearities in subjective probability estimates were smaller than those typically reported in the literature38,43 (i.e., parameters (gamma ,lambda), (alpha) and (beta) were close to their normative value of 1). We believe this reflects design and analytic choices we made and is a strength rather than weakness of our data. Specifically, the fact that the probabilistic reasoning model was clearly superior in our conditions adds confidence it would continue to be so in other situations that show larger non-linearities of the type that the probabilistic reasoning model is best equipped to predict.
One likely explanation for the relatively normative responses we found is that we conveyed probabilities using not only pie charts but also discrete frequencies, which was shown to attenuate heuristic biases by improving comprehension49, and fostering Bayesian-like updating48. This may at least partly explain why the valence effects we found for non-instrumental information demand (Fig. 1D) appear similar but weaker than those reported in an earlier study that used only analog descriptions of probabilities14. In comparison with these previous findings, our weakened valence effect produced only a modest crossover in preference for information about losses to information about gains as the likelihood of each increased (Fig. 1D, left). Given this is a key trend in the data that the mixed-motive model cannot explain (see posterior predictive checks; Fig. 1D, center), the success of the probabilistic reasoning model over the mixed-motive model despite this attenuated effect suggests the present results are a robust check of the relative strength of the model.
A second factor that may contribute to our relatively normative parameter estimates is our granular modeling approach, in which we used separate gain- and loss-domain Prelec functions and a separate value function dictated by a loss aversion parameter. Our model comparisons showed that this model complexity was fully justified by the data because each additional parameter significantly improved the probabilistic reasoning model fit (even as simplified models still outperformed the mixed-motive model). This raises the possibility that some studies using simpler models may report inflated parameter estimates (e.g., if the subjective probability of a loss is higher than that of an equivalent gain, as we and others37,39,40,41,42 propose, models that use a single PWF may misattribute this difference to the loss-aversion parameter (lambda), artificially inflating its estimates). In a similar vein, a recent study failed to detect a relationship between information demand and risky decisions, potentially because it parametrized uncertainty effects using linear-regression analyses without modeling non-linearities in value and probability functions66. Together, these considerations highlight the importance of accurate parametrization and quantitative model comparisons in describing risky decisions and information demand.
Careful parametrization is also important for understanding the neural mechanisms of this process, and may explain variable success in identifying the neural correlates of the hedonic/valence effects. Inspired by a mixed-motive approach, Charpentier and colleagues reported that the mesolimbic reward systems and the orbitofrontal cortex encode, respectively, the hedonic and cognitive components of information demand (based on their respective valence-dependent versus valence-independent responses to the opportunity to resolve uncertainty14). However, Goh and colleagues20 and van Lieshout and colleagues15 found minimal responses associated with a putative hedonic component, despite identifying robust clusters associated with uncertainty resolution in the anterior cingulate, parietal, insular and orbitofrontal cortices.
Combined with these findings, our results suggest that the ability to identify the neural correlates of valence effects may be improved by using a prospect-theory based parametrization and focusing on brain areas implicated in decision making under risk34,35,36,67,68,69. Electrophysiological recordings in non-human primates have shown that neurons in the dorsal and ventral striatum integrate reward magnitude and probability multiplicatively, exhibiting distortions that overweight small probabilities and underweight large ones36, and neurons in the anterior insular cortex encode key characteristics of the prospect theory value function34. Complementary studies using fMRI further show that regions of the human brain such as the orbitofrontal cortex – particularly its central subdivision – and the dorsolateral prefrontal cortex display BOLD responses that vary nonlinearly with probability and reward magnitude35. Thus, the non-linearities in probabilistic reasoning may have natural biological instantiations and enhance our understanding of the neural mechanisms of information demand.
Our finding that normative task performance was correlated with questionnaire scores adds to the growing body of evidence that information demand strategies can be predicted by some personality traits13,63,64,65,70. Despite the relatively small deviations from normative responses in our task, the correlations we found were robust across multiple alternative metrics. Moreover, we found similar correlations in the information demand and risk tasks, supporting our conclusion of a mechanistic relationship between the two tasks.
Our findings support previous studies65 suggesting that Need for Cognition (NFC) and Stress Tolerance are particularly consistent predictors of information demand across different task contexts. NFC scores were negatively correlated with MAD metrics, indicating that participants who were more drawn to cognitively complex tasks had more normative (veridical) probability estimates. This extends previous findings that high-NFC individuals exhibit better performance on decision-making measures like the Iowa Gambling Task71 and smaller subjective distortions of probability72. Interestingly, higher Stress Tolerance predicted behavior that was less normative – scores were positively correlated with MAD metrics. At first glance, this seems to conflict with reports that Stress Tolerance predicts more uncertainty-driven information demand strategies65 and lower proneness to anxiety73 (confirmed by the negative correlation between Stress Tolerance and PHQ scores in our task; Supplementary Fig. 14). This suggests the intriguing idea that deviations from normative estimates may not indicate “deficits” but, rather, adaptive responses that allow people to better tolerate stress. Thus, individuals who have lower ability to tolerate stress and are more prone to anxiety may lean on rigid, formula-driven choices. In contrast, people who are better able to tolerate stress and uncertainty-minimizing exploration may also be more comfortable using heuristics over exact calculations. Together, the findings underscore the need to interpret individual differences in probabilistic perception in the broader context of cognitive, emotional and adaptive responses to stress.
The main limitations of our results stem from the relative simplicity of the behavioral tasks we have used, which limited the complexity of the models we could test and interpret. Since our tasks are similar to others currently used in the field, this highlights the need for expanding the repertoire of empirical and computational tests to capture information demand in more complex settings.
An important limitation of our probabilistic reasoning model is explaining information avoidance – whereby individuals deliberately avoid highly-charged emotional information about medical diagnoses or identity-related matters31,32,74,75. Although the probabilistic reasoning model can reproduce an affirmative or null desire to obtain information (e.g., a positive or zero willingness-to-pay), it cannot predict negative WTP; the desire to avoid information. Information avoidance may instead be explained by a hybrid mixed-motive/probabilistic reasoning model that integrates non-linear subjective perceptions with a hedonic component that can assign negative value to advance information. However, as our analyses demonstrate, such a model could not be reliably estimated based on our data – and thus the feasibility of a hybrid mixed-motive/probabilistic reasoning model in explaining information avoidance remains an important topic for future investigation.
An additional difficulty we encountered is that the mixed-motive framework could not separately formalize the contributions of cognitive and instrumental value, as the two were perfectly correlated in our instrumental conditions. While a correlation between instrumental and cognitive values characterizes many natural settings – as decision-relevant information typically increases expected rewards because it reduces the uncertainty about the correct action – the two quantities can be decoupled in more complex tasks76. We have shown that the probabilistic reasoning model obviates this difficulty by using a single term related to instrumental rewards, and whether it can separately identify instrumental utility and uncertainty in more complex tasks is an important question for future investigations.
A final important result we report is that information demand was significantly more normative in the instrumental versus the non-instrumental conditions (Exp. 2 versus Exp. 1). This reinforces our findings in an earlier task63 and, together with the correlations with NFC scores reported above, suggests an interpretation in terms of computational rationality76. According to this hypothesis, information demand is cognitively effortful77 and instrumental incentives encourage participants to exert the effort required to make more veridical (normative) estimates – an idea that remains to be verified by more detailed comparisons across different instrumental and non-instrumental tasks.
Our results underscore the importance of non-linear probability and value perceptions in shaping information demand and its links with risky choices and personality traits. The results support future studies of the shared cognitive mechanisms of these processes, and may inform practical applications such as designing interventions to improve decision-making in high-stakes domains like healthcare and finance. Bridging the gap between information demand and risk-taking research moves us closer to a comprehensive understanding of the cognitive, motivational and emotional mechanisms by which humans navigate uncertainty in an information-rich world.
The data analyzed reported in this manuscript are available from the Open Science Framework: https://osf.io/rsydx/. Data and analysis code will be made available upon publication in a repository hosted on the Open Science Framework.
The code used to conduct the analyses reported in this manuscript are available from the Open Science Framework: https://osf.io/rsydx/.
Wilcox, K. & Laird, J. D. The impact of media images of super-slender women on women’s self-esteem: identification, social comparison, and self-perception. J. Res. Pers. 34, 278–286 (2000).
Google Scholar
Fathurohman, F., Marzuki, M. & Baharta, R. The influence of social media use on the self-perception and social relations of teenagers in the digital era. J. Kaji. Pendidik. dan. Psikol. 1, 111–119 (2023).
Google Scholar
Morgan, M. & Shanahan, J. Two decades of cultivation research: an asppraisal and meta-analysis. Ann. Int. Commun. Assoc. 20, 1–45 (1997).
Google Scholar
Stroud, N. J. Polarization and partisan selective exposure. J. Commun. 60, 556–576 (2010).
Google Scholar
Stroud, N. J. Selective Exposure Theories. in The Oxford Handbook of Political Communication (eds. Kenski, K. & Jamieson, K. H.) Vol. 1 (Oxford University Press, Oxford, UK, 2017).
Gottlieb, J. & Oudeyer, P.-Y. Towards a neuroscience of active sampling and curiosity. Nat. Rev. Neurosci. 19, 758–770 (2018).
PubMed Google Scholar
Sharot, T. & Sunstein, C. R. How people decide what they want to know. Nat. Hum. Behav. 4, 14–19 (2020).
PubMed Google Scholar
Gray, K. & Tonge, B. Are there early features of autism in infants and preschool children? J. Paediatr. Child Health 37, 221–226 (2001).
PubMed Google Scholar
Locander, W. B. & Hermann, P. W. The effect of self-confidence and anxiety on information seeking in consumer risk reduction. J. Mark. Res. 16, 268–274 (1979).
Google Scholar
Aderka, I. M., Haker, A., Marom, S., Hermesh, H. & Gilboa-Schechtman, E. Information-seeking bias in social anxiety disorder. J. Abnorm. Psychol. 122, 7–12 (2013).
PubMed Google Scholar
Hildebrand-Saints, L. & Weary, G. Depression and social information gathering. Pers. Soc. Psychol. Bull. 15, 150–160 (1989).
Google Scholar
Camp, C. J. I am curious-grey: Information seeking and depression across the adult lifespan. Educ. Gerontol. 12, 375–384 (1986).
Google Scholar
Bennett, D., Sutcliffe, K., Tan, N. P., Smillie, L. D. & Bode, S. Anxious and obsessive-compulsive traits are independently associated with valuation of noninstrumental information. J. Exp. Psychol. Gen. 150, 739–755 (2021).
PubMed Google Scholar
Charpentier, C. J., Bromberg-Martin, E. S. & Sharot, T. Valuation of knowledge and ignorance in mesolimbic reward circuitry. Proc. Natl. Acad. Sci. USA 115, E7255–E7264 (2018).
PubMed PubMed Central Google Scholar
van Lieshout, L. L. F. et al. Induction and relief of curiosity elicit parietal and frontal activity. J. Neurosci. 38, 2579–2588 (2018).
PubMed PubMed Central Google Scholar
van Lieshout, L. L. F., Traast, I. J., de Lange, F. P. & Cools, R. Curiosity or savouring? Information seeking is modulated by both uncertainty and valence. PLoS One 16, e0257011 (2021).
PubMed PubMed Central Google Scholar
Jiwa, M., Cooper, P. S., Chong, T. T.-J. & Bode, S. Choosing increases the value of non-instrumental information. Sci. Rep. 11, 1–11 (2021).
Embrey, J. R., Liew, S. X., Ghai, I. & Newell, B. The effect of uncertainty and reward probability on information seeking behaviour. Proceedings of the Annual Meeting of the Cognitive Science Society 43, (2021).
Kobayashi, K. & Hsu, M. Common neural code for reward and information value. Proc. Natl. Acad. Sci. USA 201820145 https://doi.org/10.1073/pnas.1820145116 (2019).
Goh, A. X. A., Bennett, D., Bode, S. & Chong, T. T.-J. Neurocomputational mechanisms underlying the subjective value of information. Commun. Biol. 4, 1–14 (2021).
Google Scholar
Bode, S. et al. Non-instrumental information seeking is resistant to acute stress. Sci. Rep. 13, 19505 (2023).
PubMed PubMed Central Google Scholar
Bode, S. et al. When knowledge hurts: humans are willing to receive pain for obtaining non-instrumental information. Proceedings of the Royal Society B: Biological Sciences 290, (2023).
Dan, O. et al. Moderate stability of risk and ambiguity attitudes across quantitative and qualitative decisions. Sci. Rep. 15, 3119 (2025).
PubMed PubMed Central Google Scholar
Brydevall, M., Bennett, D., Murawski, C. & Bode, S. The neural encoding of information prediction errors during non-instrumental information seeking. Sci. Rep. 8, 6134 (2018).
PubMed PubMed Central Google Scholar
Bennett, D., Bode, S., Brydevall, M., Warren, H. & Murawski, C. Intrinsic Valuation of Information in Decision Making under Uncertainty. PLoS Comput Biol. 12, e1005020 (2016).
PubMed PubMed Central Google Scholar
van Lieshout, L. L. F., de Lange, F. P. & Cools, R. Motives underlying human curiosity. Nat. Hum. Behav. 1 https://doi.org/10.1038/s41562-019-0565-y (2019).
Kobayashi, K., Ravaioli, S., Baranès, A., Woodford, M. & Gottlieb, J. Diverse motives for human curiosity. Nat. Hum. Behav. 3, 587–595 (2019).
PubMed Google Scholar
Foley, N. C., Kelly, S. P., Mhatre, H., Lopes, M. & Gottlieb, J. Parietal neurons encode expected gains in instrumental information. Proc. Natl. Acad. Sci. USA 114, E3315–E3323 (2017).
PubMed PubMed Central Google Scholar
van Lieshout, L. L., de Lange, F. P. & Cools, R. Why so curious? Quantifying mechanisms of information seeking. Curr. Opin. Behav. Sci. 35, 112–117 (2020).
Google Scholar
Lerman, C. et al. What you don’t know can hurt you: adverse psychologic effects in members of BRCA1-linked and BRCA2-linked families who decline genetic testing. J. Clin. Oncol. 16, 1650–1654 (1998).
PubMed Google Scholar
Persoskie, A., Ferrer, R. A. & Klein, W. M. P. Association of cancer worry and perceived risk with doctor avoidance: an analysis of information avoidance in a nationally representative US sample. J. Behav. Med 37, 977–987 (2014).
PubMed Google Scholar
Dwyer, L. A., Shepperd, J. A. & Stock, M. L. Predicting avoidance of skin damage feedback among college students. Ann. Behav. Med. 49, 685–695 (2015).
PubMed PubMed Central Google Scholar
Jiwa, M., Cooper, P. S., Chong, T. T.-J. & Bode, S. Hedonism as a motive for information search: biased information-seeking leads to biased beliefs. Sci. Rep. 13, 2086 (2023).
PubMed PubMed Central Google Scholar
Yang, Y.-P., Li, X. & Stuphorn, V. Primate anterior insular cortex represents economic decision variables proposed by prospect theory. Nat. Commun. 13, 717 (2022).
PubMed PubMed Central Google Scholar
Hsu, M., Krajbich, I., Zhao, C. & Camerer, C. F. Neural response to reward anticipation under risk is nonlinear in probabilities. J. Neurosci. 29, 2231–2237 (2009).
PubMed PubMed Central Google Scholar
Imaizumi, Y., Tymula, A., Tsubo, Y., Matsumoto, M. & Yamada, H. A neuronal prospect theory model in the brain reward circuitry. Nat. Commun. 13, 5855 (2022).
PubMed PubMed Central Google Scholar
Brooks, P., Peters, S. & Zank, H. Risk behavior for gain, loss, and mixed prospects. Theory Decis. 77, 153–182 (2014).
Google Scholar
Tversky, A. & Kahneman, D. Advances in prospect theory: cumulative representation of uncertainty. J. Risk Uncertain. 5, 297–323 (1992).
Google Scholar
Fehr-Duda, H., de Gennaro, M. & Schubert, R. Gender, financial risk, and probability weights. Theory Decis. 60, 283–313 (2006).
Google Scholar
Abdellaoui, M. Parameter-free elicitation of utility and probability weighting functions. Manag. Sci. 46, 1497–1512 (2000).
Google Scholar
Abdellaoui, M., Vossmann, F. & Weber, M. Choice-based elicitation and decomposition of decision weights for gains and losses under uncertainty. Manag. Sci. 51, 1384–1399 (2005).
Google Scholar
Pachur, T. & Kellen, D. Modeling gain-loss asymmetries in risky choice: the critical role of probability weighting. Proceedings of the Annual Meeting of the Cognitive Science Society 35, (2013).
Thaler, R. H. Behavioral economics: past, present, and future. Am. Econ. Rev. 106, 1577–1600 (2016).
Google Scholar
Shannon, C. E. A mathematical theory of communication. Bell Syst. Tech. J. 27, 379–423 (1948).
Google Scholar
Nelson, J. D., McKenzie, C. R. M., Cottrell, G. W. & Sejnowski, T. J. Experience matters. Psychol. Sci. 21, 960–969 (2010).
PubMed PubMed Central Google Scholar
Palan, S. & Schitter, C. Prolific.ac—A subject pool for online experiments. J. Behav. Exp. Financ. 17, 22–27 (2018).
Google Scholar
de Leeuw, J. R. jsPsych: a JavaScript library for creating behavioral experiments in a Web browser. Behav. Res. Methods 2014 47:1 47, 1–12 (2014).
Gigerenzer, G. & Hoffrage, U. How to improve Bayesian reasoning without instruction: frequency formats. Psychol. Rev. 102, 684–704 (1995).
Google Scholar
Garcia-Retamero, R. & Cokely, E. T. Designing visual aids that promote risk literacy: a systematic review of health research and evidence-based design heuristics. Hum. Factors: J. Hum. Factors Ergonom. Soc. 59, 582–627 (2017).
Google Scholar
Becker, G. M., DeGroot, M. H. & Marschak, J. Measuring utility by a single-response sequential method. Behav. Sci. 9, 226–232 (1964).
PubMed Google Scholar
Cacioppo, J. T. & Petty, R. E. The need for cognition. J. Pers. Soc. Psychol. 42, 116–131 (1982).
Google Scholar
Costa, P. T. & McCrae, R. R. The five-factor theory of personality. in The five-factor model of personality: Theoretical perspectives Vol. 2 51–87 (The Guilford Press, New York, 1999).
Kashdan, T. B. et al. The five-dimensional curiosity scale: Capturing the bandwidth of curiosity and identifying four unique subgroups of curious people. J. Res. Pers. 73, 130–149 (2018).
Google Scholar
Kroenke, K., Spitzer, R. L., Williams, J. B. W. & Löwe, B. An ultra-brief screening scale for anxiety and depression: the PHQ–4. Psychosomatics 50, 613–621 (2009).
PubMed Google Scholar
Prelec, D. The probability weighting function. Econometrica 66, 497 (1998).
Google Scholar
Holt, C. A. & Laury, S. K. Risk aversion and incentive effects. Am. Econ. Rev. 92, 1644–1655 (2002).
Google Scholar
Myagkov, M. & Plott, C. R. Exchange economies and loss exposure: experiments exploring prospect theory and competitive equilibria in market environments. Am. Econ. Rev. 87, 801–828 (1997).
Google Scholar
Harrison, G. W., Lau, M. I. & Williams, M. B. Estimating individual discount rates in denmark: a field experiment. Am. Econ. Rev. 92, 1606–1617 (2002).
Google Scholar
Starmer, C. & Sugden, R. Does the random-lottery incentive system elicit true preferences? An experimental investigation. Am. Econ. Rev. 81, 971–978 (1991).
Google Scholar
Cavagnaro, D. R., Pitt, M. A., Gonzalez, R. & Myung, J. I. Discriminating among probability weighting functions using adaptive design optimization. J. Risk Uncertain. 47, 255–289 (2013).
PubMed PubMed Central Google Scholar
Watanabe, S. WAIC and WBIC for mixture models. Behaviormetrika 48, 5–21 (2021).
Google Scholar
Wagenmakers, E.-J. & Farrell, S. AIC model selection using Akaike weights. Psychon. Bull. Rev. 11, 192–196 (2004).
PubMed Google Scholar
Rischall, I., Hunter, L., Jensen, G. & Gottlieb, J. Inefficient prioritization of task-relevant attributes during instrumental information demand. Nat. Commun. 14, 3174 (2023).
PubMed PubMed Central Google Scholar
Kelly, Christopher. A. & Sharot, T. Individual differences in information-seeking. Nat. Commun. 2021 12:1 12, 1–13 (2021).
Jach, H. K. et al. Individual differences in information demand have a low dimensional structure predicted by some curiosity traits. Proc. Natl. Acad. Sci. USA 121, e2415236121 (2024).
Bromberg-Martin, E. S. et al. A neural mechanism for conserved value computations integrating information and rewards. Nat. Neurosci. 27, 159–175 (2024).
PubMed PubMed Central Google Scholar
Jung, W. H., Lee, S., Lerman, C. & Kable, J. W. Amygdala functional and structural connectivity predicts individual risk tolerance. Neuron 98, 394–404.e4 (2018).
PubMed PubMed Central Google Scholar
Hsu, M., Bhatt, M., Adolphs, R., Tranel, D. & Camerer, C. F. Neural systems responding to degrees of uncertainty in human decision-making. Science (1979) 310, 1680–1683 (2005).
Breiter, H. C., Aharon, I., Kahneman, D., Dale, A. & Shizgal, P. Functional imaging of neural responses to expectancy and experience of monetary gains and losses. Neuron 30, 619–639 (2001).
PubMed Google Scholar
Jach, H. K. & Smillie, L. D. Testing the information-seeking theory of openness/intellect. Eur. J. Pers. 35, 103–119 (2021).
Google Scholar
Harman, J. L. Individual differences in need for cognition and decision making in the Iowa Gambling Task. Pers. Individ. Dif. 51, 112–116 (2011).
Google Scholar
Choi, S., Kim, J., Lee, E. & Lee, J. Probability weighting and cognitive ability. Manag. Sci. 68, 5201–5215 (2022).
Google Scholar
Charpentier, C. J., Aylward, J., Roiser, J. P. & Robinson, O. J. Enhanced risk aversion, but not loss aversion, in unmedicated pathological anxiety. Biol. Psychiatry 81, 1014–1022 (2017).
PubMed Google Scholar
Golman, R., Hagmann, D. & Loewenstein, G. Information avoidance. J. Econ. Lit. 55, 96–135 (2017).
Google Scholar
Sweeny, K., Melnyk, D., Miller, W. & Shepperd, J. A. Information avoidance: who, what, when, and why. Rev. Gen. Psychol. https://doi.org/10.1037/a0021288 (2010).
Article Google Scholar
Silvetti, M., Lasaponara, S., Daddaoua, N., Horan, M. & Gottlieb, J. A reinforcement meta-learning framework of executive function and information demand. Neural Netw. 157, 103–113 (2023).
PubMed Google Scholar
Petitet, P., Attaallah, B., Manohar, S. G. & Husain, M. The computational cost of active information sampling before decision-making under uncertainty. Nat. Hum. Behav. 2021 5:7 5, 935–946 (2021).
Download references
The authors are grateful for generous support for this work from the Mortimer B. Zuckerman Institute at Columbia University. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, USA
Matthew W. Jiwa & Jacqueline Gottlieb
Department of Neuroscience, Columbia University, New York, NY, USA
Matthew W. Jiwa & Jacqueline Gottlieb
Kavli Institute for Brain Science, Columbia University, New York, NY, USA
Jacqueline Gottlieb
PubMed Google Scholar
PubMed Google Scholar
M.J. and J.G. conceived of the study; M.J. collected and analyzed the data; M.J. and J.G. wrote the manuscript.
Correspondence to Matthew W. Jiwa or Jacqueline Gottlieb.
The authors declare no competing interests.
Communications Psychology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Mael Lebreton and Troby Ka-Yan Lui. A peer review file is available.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Jiwa, M.W., Gottlieb, J. Modeling information demand in the framework of probabilistic reasoning. Commun Psychol 4, 31 (2026). https://doi.org/10.1038/s44271-026-00398-8
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s44271-026-00398-8
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Communications Psychology (Commun Psychol)
ISSN 2731-9121 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.