Rational
Differences in Learning Across the Lifespan Emerge via Resource-Rational Computations
Rasmus Bruckner et al.
Psychological Review, forthcoming
Abstract:
Learning accurate beliefs about the world is computationally demanding but critical for adaptive behavior across the lifespan. Here, we build on an established framework formalizing learning as predictive inference and examine the possibility that age differences in learning emerge from efficient computations that consider available cognitive resources differing across the lifespan. In our resource-rational model, beliefs are updated through a sampling process that stops after reaching a criterion level of accuracy. The sampling process navigates a trade-off between belief accuracy and computational cost, with more samples favoring belief accuracy and fewer samples minimizing costs. When cognitive resources are limited or costly, a maximization of the accuracy-cost ratio requires a more frugal sampling policy, which leads to systematically biased beliefs. Data from two lifespan studies (N = 129 and N = 90) and one study in younger adults (N = 94) show that children and older adults display biases characteristic of a more frugal sampling policy. This is reflected in (a) more frequent perseveration when participants are required to update from previous beliefs and (b) a stronger anchoring bias when updating beliefs from an externally generated value. These results are qualitatively consistent with simulated predictions of our resource-rational model, corroborating the assumption that the identified biases originate from sampling. Our model and results provide a unifying perspective on perseverative and anchoring biases, show that they can jointly emerge from efficient belief-updating computations, and suggest that resource-rational adjustments of sampling computations can explain age-related changes in adaptive learning.
A watched pot seems slow to boil: Why frequent monitoring decreases perceptions of progress
André Vaz, André Mata & Clayton Critcher
Journal of Experimental Psychology: General, April 2025, Pages 895-918
Abstract:
In evaluating changing attributes (e.g., work output, pollution levels), perceivers care not only about an attribute’s level but its rate of change. Two employees likely have different value in the eyes of a supervisor if they take different amounts of time to complete the same work. Ten studies in the main article (and five in the Supplemental Materials) document and explore a monitoring frequency effect (MFE): Progress is seen to slow to the extent it is monitored more frequently. This effect was observed across various domains (workplace, public health, environmental, investment, physical growth) and was robust to financial incentives that encouraged accuracy. Several factors are identified that affect preferences for monitoring targets more or less frequently. Participants also displayed preferences for how frequently they themselves would be monitored; this investigation directly revealed the counterintuitive nature of the MFE. Although the MFE was robust to all tested variants, the size of the MFE did depend on how information about attribute changes was presented. Two mechanistic accounts -- one rooted in memory biases for tracked information and the other in a failure to synthesize the tracked information in a normative way -- were tested. Only the latter was supported. Discussion focuses on how the MFE complements or only superficially contradicts previous work on myopic loss aversion, the ratio bias, partition dependence, and tracking goal progress. The MFE identifies a qualitatively distinct way by which prior evaluations and beliefs can color evaluations of targets, thereby reinforcing even misguided priors.
Bridging the human–AI knowledge gap through concept discovery and transfer in AlphaZero
Lisa Schut et al.
Proceedings of the National Academy of Sciences, 1 April 2025
Abstract:
AI systems have attained superhuman performance across various domains. If the hidden knowledge encoded in these highly capable systems can be leveraged, human knowledge and performance can be advanced. Yet, this internal knowledge is difficult to extract. Due to the vast space of possible internal representations, searching for meaningful new conceptual knowledge can be like finding a needle in a haystack. Here, we introduce a method that extracts new chess concepts from AlphaZero, an AI system that mastered chess via self-play without human supervision. Our method excavates vectors that represent concepts from AlphaZero’s internal representations using convex optimization, and filters the concepts based on teachability (whether the concept is transferable to another AI agent) and novelty (whether the concept contains information not present in human chess games). These steps ensure that the discovered concepts are useful and meaningful. For the resulting set of concepts, prototypes (chess puzzle–solution pairs) are presented to experts for final validation. In a preliminary human study, four top chess grandmasters (all former or current world chess champions) were evaluated on their ability to solve concept prototype positions. All grandmasters showed improvement after the learning phase, suggesting that the concepts are at the frontier of human understanding. Despite the small scale, our result is a proof of concept demonstrating the possibility of leveraging knowledge from a highly capable AI system to advance the frontier of human knowledge; a development that could bear profound implications and shape how we interact with AI systems across many applications.
A Quest for AI Knowledge
Joshua Gans
NBER Working Paper, March 2025
Abstract:
This paper examines how the introduction of artificial intelligence (AI), particularly generative and large language models capable of interpolating precisely between known data points, reshapes scientists' incentives for pursuing novel versus incremental research. Extending the theoretical framework of Carnehl and Schneider (2025), we analyse how decision-makers leverage AI to improve precision within well-defined knowledge domains. We identify conditions under which the availability of AI tools encourages scientists to choose more socially valuable, highly novel research projects, contrasting sharply with traditional patterns of incremental knowledge growth. Our model demonstrates a critical complementarity: scientists strategically align their research novelty choices to maximise the domain where AI can reliably inform decision-making. This dynamic fundamentally transforms the evolution of scientific knowledge, leading either to systematic “stepping stone” expansions or endogenous research cycles of strategic knowledge deepening. We discuss the broader implications for science policy, highlighting how sufficiently capable AI tools could mitigate traditional inefficiencies in scientific innovation, aligning private research incentives closely with the social optimum.
Typical Ranges as Scale-Specific Benchmarks: When and Why Percentages Amplify Relative Magnitudes and Their Differences
Joowon Klusowski & Joshua Lewis
Management Science, forthcoming
Abstract:
Business managers and policymakers must often communicate magnitudes. Yet conveying large relative magnitudes without desensitizing people to further increases can be challenging because of diminishing sensitivity to large numbers. In this research, we propose that percentage expressions not only make large relative magnitudes (e.g., 500%) appear larger than equivalent non-percentage expressions but also make large increases in relative magnitudes (e.g., from 500% to 600%) appear larger. We posit an explanation: percentages typically have values between 0% and 100%, so when percentages and percentage-point differences reach 100% or more, they seem unusually large. This hypothesis is supported by data scraped from New York Times articles and a series of online experiments employing both management-relevant scenarios and incentive-compatible decisions. Existing theories of magnitude perception either cannot predict all the results of these studies (e.g., numerosity and unitosity) or need further specification to do so (e.g., decision-by-sampling and range-frequency theory). We discuss implications for the theory of magnitude and difference perception and the practice of communicating large magnitudes and changes.
Superficial auditory (dis)fluency biases higher-level social judgment
Robert Walter-Terrill, Joan Danielle Ongchoco & Brian Scholl
Proceedings of the National Academy of Sciences, 1 April 2025
Abstract:
When talking to other people, we naturally form impressions based not only on what they say but also on how they say it -- e.g., how confident they sound. In modern life, however, the sounds of voices are often determined not only by intrinsic qualities (such as vocal anatomy) but also by extrinsic properties (such as videoconferencing microphone quality). Here, we show that such superficial auditory properties can have surprisingly deep consequences for higher-level social judgments. Listeners heard short narrated passages (e.g., from job application essays) and then made various judgments about the speakers. Critically, the recordings were modified to simulate different microphone qualities, while carefully equating listeners’ comprehension of the words. Though the manipulations carried no implications about the speakers themselves, common disfluent auditory signals (as in “tinny” speech) led to decreased judgments of intelligence, hireability, credibility, and romantic desirability. These effects were robust across speaker gender and accent, and they occurred for both human and clearly artificial (computer-synthesized) speech. Thus, just as judgments from written text are influenced by factors such as font fluency, judgments from speech are not only based on its content but also biased by the superficial vehicle through which it is delivered. Such effects may become more relevant as daily communication via videoconferencing becomes increasingly widespread.
Acting Wastefully but Feeling Satisfied: Understanding Waste Aversion
Ro'i Zultan, Ori Weisel & Yaniv Shani
Journal of Behavioral Decision Making, April 2025
Abstract:
Paying more than one could have paid to obtain the same outcome is wasteful. In four experiments, we show that waste aversion can lead people to prefer a more wasteful outcome over a more frugal outcome, as long as it eliminates the feeling of wastefulness. In Study 1, we measured participants' satisfaction with lottery outcomes to find that they are less satisfied with their obtained outcome relative to an inferior, dominated, outcome -- if they are aware of a counter-factual in which they could have paid less to achieve the dominant outcome. Study 2 revealed that responsibility for the decision that led to the outcome does not intensify the effect, suggesting that wastefulness is a more prominent explanation for the effect than regret. Study 3 extended the results from outcome satisfaction to decisions. Participants altered their choice of whether to continue or terminate searching for an apartment based on their awareness of a counterfactual that renders the process leading to the outcome as wasteful or not. Waste aversion leads participants to extend their search beyond what they would do based purely on their preferences and expectations. Study 4 replicated these findings with payoff-relevant decisions. Taken together, these four studies establish that waste aversion leads to higher satisfaction with dominated outcomes in real-world experiences. The effect does not rely on decision regret, and may lead to suboptimal decisions.