What Is Decision-Making Under Risk and Uncertainty?
Decision-making under risk and uncertainty is the study of how people choose among actions whose outcomes are not fully known when the choice is made. In the classic distinction, risk involves a specified or reasonably estimable probability distribution over possible outcomes, whereas uncertainty involves probabilities that are incomplete, disputed, or unavailable. Ambiguity is a particularly important form of uncertainty: the decision maker can identify candidate outcomes but lacks a unique, trusted probability distribution over them. This distinction changes both the mathematics of a decision and the psychological processes available for evaluating it. Ellsberg, 1961; Höppner, 2025.
A choice is represented by alternatives, possible states of the world, consequences, beliefs about those states, and a criterion for comparing the consequences. Consider a medical test, a business investment, a flood warning, or a software security decision. Each has consequences whose importance differs for the person or institution involved. If the chances of those consequences are well characterized, expected-value or expected-utility methods can organize the choice. If those chances are uncertain, the decision additionally requires assumptions about missing information, the reliability of estimates, and how cautiously to respond to model uncertainty. Psychology investigates how people actually make such judgments, including how they learn probabilities, respond to consequences, and change their preferences across formats and settings. Fischhoff and Broomell, 2020.
The central scientific question is therefore not simply whether a person is willing to take a risk. It is how beliefs about uncertain events, evaluations of potential outcomes, information-search strategies, reference points, emotions, and decision procedures interact to generate a particular choice. An observed choice alone often cannot identify the underlying mechanism: selecting a safer option can reflect a concave utility function, pessimistic beliefs, regret, institutional constraints, a lack of trust in the stated probabilities, or a strategy suited to the environment.
Risk, Ambiguity, and Uncertainty: Three Distinct Decision Environments
Risk: probabilities are specified
A decision under risk includes outcomes and their probabilities. A fair coin paying $100 on heads and $0 on tails has two outcomes, each with probability 0.5, and an expected monetary value of $50. The expected value is the probability-weighted average payoff: EV = sum over all outcomes of probability times monetary outcome. It is useful for repeated exposure and accounting comparisons, but maximizing money-weighted expectation is only one possible decision criterion. An individual with a pressing need to cover an essential expense may rationally place a different utility on $50 certain than on a gamble averaging $50. Karni, 2014.
Ambiguity: probabilities are incompletely known
Imagine two sealed urns, one with exactly 50 red and 50 black balls and another containing 100 balls in an undisclosed red-black proportion. A decision to bet on red from the first urn comes with an explicit one-half probability. The corresponding choice from the second urn involves ambiguity because the red-ball proportion is unknown. A person who prefers the first urn may dislike imprecise odds, distrust the unknown process, or apply a conservative decision rule. The Ellsberg paradigm studies patterns of such choices, including combinations that challenge a single additive subjective-probability representation. Ellsberg, 1961; Epstein and Schneider, 2010.
Deep uncertainty: outcomes, models, or probabilities can change
Some practical problems go beyond a known menu of possible events with unknown odds. Novel technologies, complex ecological changes, and rare institutional disruptions may involve competing models, evolving conditions, and outcomes that have not yet been articulated. A single numerical probability then risks making uncertainty appear more sharply measured than it is. Decision analysts may compare several plausible scenarios, identify options that perform acceptably across them, and specify triggers for reconsidering a course of action. Such methods answer a different question from estimating one exact probability. Höppner, 2025.
Statistical risk, perceived risk, and risk preference
Statistical risk can refer to a probability distribution, dispersion, expected loss, or a particular tail event, depending on discipline and model. Perceived risk is a person's appraisal of threat, likelihood, severity, unfamiliarity, and control. Risk preference denotes how alternatives with uncertain outcomes are ranked. A person can accurately perceive a high probability of harm yet accept it to achieve an important goal; another may judge a hazard unlikely but still reject it because the downside is unacceptable. Keeping these constructs distinct prevents an overly simple interpretation of behavior. Mishra, 2014.
A Decision Example: Why Expected Value and Expected Utility Differ
Suppose someone chooses between receiving $80 for certain and a lottery that pays $200 with probability 0.5 and $0 otherwise. The lottery's expected monetary value is $100, compared with $80 for the certain option. A risk-neutral decision maker whose utility is linear in money chooses the lottery. A decision maker who values each additional dollar less than the previous dollar may prefer the guaranteed $80. The preference can be consistent with maximizing expected utility; the lottery's higher mean does not prove it is better for every person.
For an illustrative concave utility function u(x) = square root of x over nonnegative payoffs, the sure option produces utility sqrt(80), approximately 8.94, whereas the lottery produces 0.5 sqrt(200) + 0.5 sqrt(0), approximately 7.07. This constructed numerical example shows how a defined utility function generates risk aversion. It is not an empirical estimate of anyone's utility function, and changing initial wealth, stakes, or the domain may change the comparison.
Expected utility evaluates a lottery L using EU(L) = sum over outcomes i of p(i)u(x(i)), where the probabilities are nonnegative and add to one, and u represents how the decision maker ranks consequences. Expected value instead uses u(x) = x. The curvature of the utility function provides a conventional local description of risk attitude under expected utility: concavity implies preference for the certain expected payoff over a nondegenerate lottery with the same mean. This conclusion depends on the domain, definitions, and maintained representation assumptions. Karni, 2014.
Normative, Descriptive, and Prescriptive Models
Normative models articulate standards for consistent or goal-directed choices under explicit assumptions. Expected utility and Bayesian probability are examples when their relevant axioms and probability judgments apply. Descriptive models seek to explain or predict observed behavior, including patterns that violate certain normative axioms. Prescriptive approaches develop practical methods to help a particular decision maker reach outcomes that better reflect stated goals and the available evidence. These are related scientific tasks, and one should not silently substitute for another. Fischhoff and Broomell, 2020.
A claim that choice behavior is a 'bias' must specify its comparison standard. Choosing a certain amount instead of a higher expected monetary payoff differs from the expected-value benchmark, but it can satisfy expected utility for concave utility. In contrast, a preference reversal generated by describing logically equivalent outcomes differently raises a question about description invariance if the decision maker regards the descriptions as equivalent. Likewise, incoherent betting choices can be evaluated under the axioms of subjective probability, but the interpretation depends on whether the person's information and understanding match the experimenter's assumptions. Researchers therefore need to name the mathematical or behavioral criterion before classifying a departure from it. Starmer, 2000.
Expected Utility Theory: Axioms and Interpretation
Expected utility theory gives a representation of preferences over risky lotteries. Under suitable axioms, including completeness and transitivity of preferences and an independence condition, a numerical utility function can represent those preferences through probability-weighted utility. The independence condition says, in a standard formulation, that adding the same probabilistic component to each of two lotteries should not reverse their ranking. These axioms define the representation; researchers can separately ask whether people satisfy them and whether the model predicts choices outside the estimation task. Karni, 2014; Starmer, 2000.
Expected utility is sometimes introduced as a theory that people carefully calculate probabilities and utilities. Its axiomatic interpretation is more precise: if an individual's preference ordering satisfies the axioms, it can be represented as maximizing expected utility, regardless of whether the person consciously computes the formula. A psychological process account requires additional evidence about attention, memory, information acquisition, reaction time, or neural and behavioral dynamics. Preference representation and mental implementation are distinct research claims.
The value of expected utility is analytical clarity. It distinguishes differences in probabilities from differences in consequences and can support transparent comparisons when the options and utility evaluations are well specified. Its empirical limits are equally central: experimental patterns, especially common-consequence and common-ratio effects, often lead researchers to examine violations of independence. Different models may fit different samples or choice environments, and an axiom violation in a questionnaire is not by itself a diagnosis of defective reasoning. Starmer, 2000.
The Allais Paradox: A Challenge to Independence
The Allais paradox concerns patterns of preference between lotteries involving certainty and a common component. In a simplified example, imagine choosing between A: $1 million guaranteed, and B: a 10% chance of $5 million, an 89% chance of $1 million, and a 1% chance of nothing. Many people favor the certainty of A. In a second choice, C offers an 11% chance of $1 million and D offers a 10% chance of $5 million, with the remaining probabilities yielding nothing; some people choose D. These two selections can violate the expected-utility independence axiom under the assumed common-consequence structure.
The conceptual point is that a shared outcome or probability component should cancel from the comparison under the axiom, while psychological evaluation may give special weight to the certainty of an outcome. The classic paradigm motivated the development of non-expected-utility theories, but its interpretation also depends on incentive schemes, comprehension, stakes, and how lotteries are presented. One pair of hypothetical answers does not establish a stable individual trait. Starmer, 2000; Höppner, 2025.
Subjective Expected Utility: Decisions When Odds Must Be Inferred
In many decisions the probabilities are beliefs rather than frequencies supplied by an experimenter. Subjective expected utility models represent choices by weighting outcome utilities with the decision maker's subjective probabilities. For an act a whose consequence depends on the state s, the illustrative representation is SEU(a) = sum over states s of P(s)u(x(a,s)). Under the relevant Savage-style axioms, the subjective probabilities and utility representation can be inferred from patterns of preference. This gives a formal account of coherent uncertain choice when people behave as if there is one consistent probability distribution. Karni, 2014; Echenique, 2020.
The representation does not ensure a person's belief is empirically accurate. An individual can maximize subjective expected utility using badly calibrated probabilities. Separately, someone may give accurate probabilities but evaluate possible outcomes according to objectives a researcher has not captured. Good assessment therefore distinguishes probability quality, preference coherence, and eventual decision outcomes. These domains support different interventions: better forecasts can address beliefs; clearer trade-off discussions can address preferences; and improved feedback can address learning.
Ellsberg's Paradox: Why Ambiguity Matters
Ellsberg's three-color urn illustrates a violation of the idea that every ambiguous act is evaluated using one stable additive distribution over states. In one famous setup, an urn contains 30 red balls and 60 balls in an unknown mixture of black and yellow. The participant first chooses whether to bet on red or black, then whether to bet on red-or-yellow or black-or-yellow. A common preference pattern favors red over black in the first comparison and black-or-yellow over red-or-yellow in the second. Those rankings cannot both be represented by a single additive subjective probability over the three colors with otherwise equal payoffs. Ellsberg, 1961.
Ambiguity aversion is often a useful description of a tendency to prefer a known probability distribution over an unspecified one, holding other relevant aspects appropriately controlled. The phenomenon cannot always be reduced to ordinary risk aversion. A person may accept a high-variance lottery with well-defined odds while avoiding a lower-variance gamble with unknown odds. Experimental designs must nevertheless identify whether the participant sees the ambiguous source as less trustworthy, believes the experimenter has an informational advantage, or interprets the unknown mixture as adversarial. Epstein and Schneider, 2010; Boun My et al., 2024.
Beyond a Single Probability: Multiple-Priors and Other Ambiguity Models
Multiple-priors models let a decision maker consider a set of plausible probability distributions instead of committing to one. Under maxmin expected utility, the person evaluates an action by its least favorable expected utility across the allowed distributions and chooses the action with the best of those worst-case values. A simplified expression is V(a) = min over P in the plausible set of E_P[u(x(a,s))]. This decision rule formalizes caution about model uncertainty, but its appropriateness depends on the cost of disastrous outcomes and whether the probability set is credible. Epstein and Schneider, 2010.
Alternative ambiguity frameworks permit smooth ambiguity attitudes, differing responses to the source of uncertainty, or beliefs that change with learning. They can explain why ambiguity aversion is strong in one task, weak in another, or replaced by ambiguity seeking. The central modeling challenge is identification: similar observed choices can arise from different beliefs and valuations, and a model with many adjustable parameters may describe in-sample data without generalizing well. Karni, Maccheroni, and Marinacci, 2015; Echenique, 2020.
Prospect Theory and Reference-Dependent Valuation
Prospect theory explains important regularities in choices over gains and losses. Instead of always evaluating consequences as changes in final wealth, people often evaluate outcomes relative to a reference point. The value function in the theory is typically concave over gains, convex over losses, and steeper for losses than for corresponding gains under selected measurement conditions. Probabilities also enter choice through decision weights that need not equal objective probabilities. Kahneman and Tversky, 1979.
Cumulative prospect theory extends this approach using cumulative, rank-dependent weights and separate weighting functions for gains and losses. It offers a way to describe a fourfold pattern: risk aversion for high-probability gains, risk seeking for low-probability gains, risk seeking for high-probability losses, and risk aversion for low-probability losses. These are theoretical patterns supported in particular experimental designs, not universal responses across every population, stake size, or decision domain. Tversky and Kahneman, 1992.
Consider two objectively equivalent presentations of a loss. A decision maker may respond differently when an outcome is presented as a loss relative to what they expected versus a gain relative to a different baseline. The resulting choice is shaped by how the reference point is established and whether the comparison itself remains psychologically equivalent. The concept deserves a dedicated treatment because its predictions and debates are more specific than the general distinction between risk and uncertainty. Kahneman and Tversky, 1979; Starmer, 2000.
Why loss aversion requires careful interpretation
Loss aversion is often summarized by a single coefficient representing stronger responses to losses than to equal gains. Such a coefficient depends on the functional form, reference point, elicitation procedure, and whether the choice involves mixed gambles, consumption outcomes, or prices. A coefficient obtained in one setting is not a universal constant of human psychology. Risk aversion, probability weighting, loss aversion, regret, and selection into the task may yield similar behavioral patterns. Scientific interpretation therefore requires designs that distinguish these explanations rather than simply labeling every cautious choice 'loss averse.' Kahneman and Tversky, 1979; Höppner, 2025.
How Probabilities Are Represented and Weighted
A probability is a number between zero and one indicating the likelihood of an event under a specified model. A subjective belief is a person's assessment of that likelihood, which can be calibrated against eventual outcomes when sufficiently many forecasts are recorded. Probability weighting in a choice model is different: it concerns the influence of stated or believed probabilities on preference. Confusing probability judgment with decision weight makes it difficult to identify a mechanism. Tversky and Kahneman, 1992.
Rare events are particularly informative. In description-based tasks, a stated one-percent chance of a large loss may receive disproportionate weight relative to its objective frequency within some prospect-theory models. Yet real-world decisions often depend on personal experience rather than explicit descriptions. An event never encountered in a small sample may exert little influence even if its underlying probability is important. Researchers call this divergence the description–experience gap. Hertwig and Erev, 2009.
Decisions from description versus experience
In a description condition, the participant may read that a lottery pays a reward with a 1% probability. In an experience condition, they may sample outcomes before choosing, perhaps never seeing that reward. Differences in search length, sampling error, recency, and learning mechanisms can create choice differences without assuming stable, irrational probability weights. Evidence reviews find that rare events tend to have less impact in many experience-based tasks than in corresponding description-based tasks, with qualifications depending on feedback, sampling rules, and incentives. Hertwig et al., 2004; Hertwig and Erev, 2009.
This distinction matters for insurance, security, maintenance, and preparedness. A person can understand a probability stated on a page while personal memories still shape perceived urgency. Improving a decision may require ensuring that a rare but consequential outcome is included in the comparison, clarifying denominators and time horizons, and collecting sufficient evidence. Whether an intervention changes actual behavior or long-term outcomes is a further empirical question.
Psychological Mechanisms in Risky and Uncertain Choice
Attention and selective information acquisition
Decision makers seldom process every item of information equally. Which outcomes appear first, how many alternatives are compared, how probabilities are displayed, and whether a person can sample more observations all influence the evidence available at the moment of choice. A fast heuristic can be efficient when information is costly and the environment rewards simple discriminating cues. A more elaborate calculation can help when stakes are large and inputs are trustworthy. The usefulness of a strategy depends jointly on the person, task, and structure of the environment. Gigerenzer and Gaissmaier, 2011.
Memory, sampling, and learning
Beliefs about uncertainty are frequently constructed from remembered cases, recent outcomes, social testimony, and repeated feedback. Samples can be unrepresentative, particularly when rare events occur infrequently or data collection is selectively stopped. Recency weighting may track a changing environment efficiently, but it may also make a stable process look more volatile than it is. Feedback can teach useful probabilities if the learner sees reliable outcome labels, has enough repetitions, and knows which actions caused which consequences. Decisions made once, without feedback, present a different learning problem from repeated lottery tasks. Hertwig and Erev, 2009.
Affect, dread, and personal importance
A decision has emotional significance because outcomes differ in what they mean for safety, relationships, identity, and future opportunities. Anxiety about potential loss can change information search or the attractiveness of delay; positive affect can make gains salient; anticipated regret can favor options that are easier to justify later. These influences require measurement in context. Emotional intensity is not interchangeable with inaccurate probability estimation, and a cautious choice may be well aligned with the decision maker's values. The integration of affect and objective attributes is an empirical issue examined in psychological risk research. Mishra, 2014.
Motivation, control, and institutional incentives
People choose within constraints. A worker, manager, patient, or student may be sensitive to how rewards, responsibility, deadlines, and available resources are distributed. A manager rewarded for avoiding visible failures may prefer a different risk profile from an owner who bears the long-run consequences. Apparent inconsistency may reflect a change in goals, background wealth, option availability, or accountability. Good experiments control these factors where possible; good applications ask whose utility and whose probabilities the analysis represents.
Why a Single Risk-Taking Trait Is Too Simple
Risk preference is often measured with hypothetical gambles, incentivized lotteries, self-reports, observed behaviors, or domain-specific questionnaires. These instruments need not measure one interchangeable latent trait. A 2025 systematic review and individual-participant-data meta-analysis synthesized 358 risk preference measures from 33 panels, 57 samples, and 579,114 respondents. The authors found substantial heterogeneity, greater temporal stability for some self-report measures than for behavioral measures, and low convergence across different measures. The finding cautions against treating one laboratory gamble as a comprehensive assessment of a person's decisions in finance, health, work, and relationships. Bagaïni et al., 2025.
Risk attitudes also vary with consequences. People may accept uncertainty to help another person while avoiding it with money, or tolerate a business gamble yet avoid an unfamiliar medical procedure. Age, knowledge, financial security, trust, stakes, and familiarity can change the choice context. Researchers must examine measurement reliability, repeated decisions, and within-person variation before inferring a stable disposition. Cultural comparisons add translation, institution, wealth, and sampling questions; national or demographic labels alone do not explain individual decisions.
Measuring Risk Preference, Beliefs, and Decision Quality
Multiple price lists and incentivized lotteries
A common design presents a sequence of choices between a relatively safe lottery and a riskier one, varying the probability or payoff until the participant switches options. Analysts use the switching point to estimate a risk preference parameter under a maintained utility model. Incentivizing choices with real monetary consequences can improve relevance, but stakes, payment rules, task comprehension, and the distribution of possible choices still matter. A model-based estimate is conditional on the validity of the model and on the person's interpretation of the task.
Certainty equivalents and probability equivalents
A certainty equivalent is a guaranteed payoff judged equally attractive to a given lottery. For an equal-probability gamble paying $0 or $100, a certainty equivalent of $40 suggests a risk premium of $10 compared with the gamble's $50 expected monetary value. This observation is compatible with risk aversion in that task but does not uniquely identify a psychological cause. Probability-equivalent procedures instead vary the chance of an outcome until the participant expresses indifference. These methods can be affected by response noise, rounding, and construction of the question.
Eliciting subjective probabilities
An assessor can ask participants to state a probability for an event, record repeated forecasts, and compare them with subsequent observed frequencies. Calibration concerns agreement between stated confidence levels and observed frequencies over an appropriate set of events. Resolution or discrimination concerns whether forecasts distinguish events that occur from those that do not. A scoring rule can reward honest probabilities under specified incentives and utility assumptions. Forecast accuracy can improve without directly revealing a person's risk preference, because belief accuracy and consequence valuation remain separate.
Estimating ambiguity preferences
Ambiguity studies compare wagers with known and unknown probability structures, sometimes systematically varying the available information. Researchers can test whether a participant pays to avoid an ambiguous source or reverses a preference when descriptions change. Estimation must separate ambiguity attitudes from distrust of the source, optimistic versus pessimistic beliefs, risk preferences, and misunderstanding. A 2024 experiment examining relations between risk and ambiguity preferences found imperfect alignment and a sample-specific negative correlation; the authors report that 34% of their sample showed oppositely classified risk and ambiguity attitudes. That percentage describes their operational measures and participants, not a population prevalence. Boun My et al., 2024.
Process measures and real decisions
Choice rates reveal preferences but may conceal how the choice was reached. Response time, eye movements, information search, repeated sampling, confidence, and think-aloud protocols can provide convergent process evidence. Field outcomes and consequential choices increase external relevance, although they introduce confounding factors and missing counterfactuals. Random assignment supports causal identification of framing or information interventions; preregistration and independent replication help evaluate how stable an effect is. No single metric simultaneously establishes choice quality, mechanism, and real-world transfer. Fischhoff and Broomell, 2020.
What Makes a Decision Good?
A good decision is evaluated against a stated objective using the information available at the time it was made. An unfortunate outcome does not necessarily prove the preceding decision was poor: even a well-calibrated 90% forecast will be wrong on some occasions. Conversely, a lucky outcome does not validate a flawed procedure. Decision quality may refer to coherence with explicitly accepted axioms, accuracy of probability forecasts, consistency with the decision maker's values, robustness to plausible models, fairness of the process, or long-run performance. These criteria can conflict and must be made explicit.
Where outcomes can be repeated and measured, decisions can be evaluated using prediction scores, expected losses, or regret relative to a benchmark strategy. In one-off situations, transparency about beliefs, alternatives, and trade-offs may be more informative than treating one outcome as a definitive test. Good decision-making requires that a benchmark be appropriate for the actual decision problem; maximizing average dollars is an incomplete objective when financial ruin, health, or irreversibility matters.
Comparing the Main Decision Models
Expected value
Expected value computes the probability-weighted mean of outcomes measured on a chosen numerical scale. It is directly useful in accounting, insurance pricing, and repeated decisions where this scale reflects the objective. Its key limitation as a rule for individuals is that the mean alone omits how consequences are valued and distributed. A low-probability catastrophic loss may matter even if the mean payoff is positive.
Expected utility
Expected utility weights utilities, not raw monetary amounts, by specified objective probabilities. It captures aversion to or attraction toward risk through the utility representation, given the relevant axioms. It does not by itself explain all experimental violations of independence, reference-dependent choice, or ambiguous odds. Karni, 2014.
Subjective expected utility
Subjective expected utility replaces stipulated odds with coherent subjective beliefs, under the relevant axiomatic representation. It is helpful when uncertainty can be represented through one person's probability distribution and consequences can be ranked. The Ellsberg paradigm shows why a single additive prior may fail to represent important ambiguity-sensitive choices. Ellsberg, 1961.
Cumulative prospect theory
Cumulative prospect theory models reference dependence, different valuations of gains and losses, and rank-dependent probability weighting. It can accommodate choice patterns not explained by standard expected utility, but parameter estimation, reference-point specification, and out-of-sample prediction remain central issues. Its descriptive strength varies by task and available data. Tversky and Kahneman, 1992.
Ambiguity-sensitive and robust models
Multiple-priors and related models explicitly represent uncertainty about the probability distribution. Robust approaches compare actions over credible sets of models or scenarios and pay attention to worst-case performance and reversibility. They are particularly relevant when a single precise forecast is unjustified. Their outcomes depend substantially on how the plausible model set and loss criteria are chosen. Epstein and Schneider, 2010.
Heuristic and experience-based models
Heuristic models specify limited-information strategies such as examining a few cues, using thresholds, or learning from sampled outcomes. They can outperform more complex rules when data are sparse, relationships shift, or computation is expensive, but they can also omit consequential evidence. Experience-based models illuminate learning, feedback, rare events, and sequential sampling. The question is which strategy predicts and supports good decisions in the specific environment. Gigerenzer and Gaissmaier, 2011; Hertwig and Erev, 2009.
Worked Example: Insurance and the Meaning of a Small Probability
Suppose a household faces a 1% annual probability of a $20,000 loss. The expected annual monetary loss is $200. A risk-neutral household with no other costs may regard an annual insurance premium substantially exceeding $200 as unattractive. A risk-averse household may rationally pay more to avoid a potentially disruptive loss. Deductibles, coverage exclusions, claims administration, tax treatment, correlated losses, and existing wealth alter the comparison. The correct benchmark is therefore the expected utility of the available complete options, including affordability and consequences, rather than the $200 number alone.
Now suppose the 1% probability is estimated from sparse data and experts disagree about whether it is 0.5% or 3%. The decision has ambiguity as well as risk. At those bounds the expected annual monetary loss would be $100 or $600, respectively, before other terms. Presenting only '1%' hides the model sensitivity. A defensible comparison reports the plausible range, underlying evidence, sensitivity to uncertain assumptions, and what protection or reserve would be sustainable. These figures are illustrative calculations, not advice about actual insurance premiums.
Worked Example: Medical Screening and Conditional Probabilities
Suppose a condition affects 1% of a hypothetical population, a screening test detects 90% of cases, and it returns a positive result in 5% of people who do not have the condition. Among 10,000 screened people, roughly 100 would have the condition and 90 of them would test positive; among 9,900 without the condition, about 495 would also test positive. Thus, approximately 90 of 585 positive results would be true positives, giving a positive predictive value near 15.4% in this particular hypothetical population. The calculation illustrates why a test's sensitivity does not equal the probability of having the condition after a positive result.
The arithmetic is a Bayesian update using a stated prior prevalence and test characteristics. A real clinical decision also depends on the consequences of false positives and false negatives, confirmatory testing, treatment risks, and the person's values. Population-specific prevalence and test performance must come from clinical evidence, and healthcare decisions should rely on qualified clinicians and current practice guidance. Research on decision aids indicates that structured information and value clarification can improve knowledge and accurate risk perceptions for many health choices, although it does not demonstrate that every aid improves every downstream outcome. Stacey et al., 2024.
Worked Example: Business Investment Under Ambiguity
A company is considering launching a product with a possible large return and a possible substantial loss. Forecasts differ because demand, competitors, regulation, and production costs may change together. A decision based on one expected-return estimate could overlook how sensitive the outcome is to those assumptions. A scenario analysis can compare an immediate full launch, a limited pilot, staged investment, and postponement, specifying the consequence of each action under plausible demand and cost states.
The decision can include the value of information: a small pilot may reveal customer behavior before the organization commits to a costly irreversible launch. The value of an experiment depends on its cost, whether it changes the probability of later states or merely reveals them, and whether the resulting information changes the preferred action. A robust strategy may sacrifice some best-case return to avoid unacceptable outcomes under several credible scenarios. This is a framework for analysis; it does not determine the optimal strategy without the organization's actual evidence and objectives.
Worked Example: Forecasting an Uncertain Technology
When assessing a new technology, analysts may have evidence about technical performance but limited evidence about adoption, misuse, failure modes, and institutional responses. A single confident forecast combines several uncertainties and can conceal dependence between them. The analysis becomes clearer when it decomposes the question: What technical result is likely? Under what deployment conditions? Which harms or benefits follow from that result? What is known from measurement, what is inferred from analogy, and what remains unresolved?
Decision-support software and AI models can help organize alternatives, simulate scenarios, and summarize evidence. Their output can also introduce new uncertainties about training data, calibration, provenance, and changes in operating conditions. Trust calibration requires evaluating an aid's performance within the task, monitoring failures, and retaining the ability to override recommendations. These human–AI interaction mechanisms are distinct from the general theory of choice under risk and have their own specialized scientific owner in the encyclopedia.
Framing, Context, and Risk Communication
Two messages can refer to the same underlying event while eliciting different judgments. A 90% survival rate and a 10% mortality rate describe equivalent binary outcomes when their denominators, time horizons, and populations are genuinely identical. The different emotional and semantic emphasis may nonetheless change attention and preferences. In risk communication, researchers distinguish logically equivalent frames from descriptions that contain genuinely different information or imply different reference points. This matters for interpreting whether a preference reversal indicates a violation of invariance or a reasonable response to pragmatic meaning. Fischhoff and Broomell, 2020.
Useful communication practices include reporting absolute as well as relative changes, identifying the population and time horizon, distinguishing probabilities from severity, and showing uncertainty intervals when they are estimable. Natural frequencies can make conditional-risk questions more transparent, but no presentation format eliminates all errors in all tasks. Good information design should be tested against comprehension, calibration, and consequential choices rather than assumed effective because a visual looks simple.
Risk, Uncertainty, and Social Context
Risky choices are shaped by institutions, social norms, available information, responsibility for others, and the distribution of harm. An individual investor, a public official, and a physician face different decision objectives even when all are presented with the same nominal probabilities. Fairness may require special attention to who bears the losses and whether vulnerable groups have meaningful choices. A psychologically realistic analysis specifies the decision maker and the affected populations before claiming that one option is universally rational.
Cross-cultural research introduces an additional identification problem. Differences between samples may reflect income, insurance systems, trust, institutions, education, probability numeracy, translation, experimental payment size, or selection into the study. Treating a country mean as a universal characteristic of its residents obscures within-group variation. Evidence-based comparisons require harmonized tasks, representative samples where possible, and uncertainty intervals around group estimates.
Reproducibility, Effect Sizes, and External Validity
The literature contains strong foundational demonstrations of uncertainty-sensitive choice, together with substantial heterogeneity in measured effect sizes and mechanisms. Experimental outcomes depend on instructions, monetary stakes, ambiguity sources, incentive compatibility, comprehension checks, reference points, and the source of probabilities. A dramatic choice pattern in a classic vignette does not establish that the same percentage of people will make the same choice in another country or decade. The proper reporting unit is the study's sampled population, procedure, operational contrast, and uncertainty about the estimate.
Modern comparative work emphasizes fitting models to tasks with appropriately measured choice data and evaluating prediction on new participants and new decision situations. Models with more parameters may describe existing responses more closely while performing less well out of sample. The 2025 review by Höppner organizes model selection around descriptive realism, predictive accuracy, tractability, and ecological validity rather than treating one theory as sufficient for every risk environment. Höppner, 2025.
The 2025 individual-participant-data meta-analysis of risk-preference measures offers a concrete caution: temporal stability and convergence change substantially with instrument and domain. Experimental estimates of ambiguity aversion, prospect-theory parameters, or risk premiums likewise require uncertainty quantification and replication within the measurement design. When a paper does not provide a transferable pooled effect size, reporting a sweeping population estimate would create false precision. Bagaïni et al., 2025.
How to Make Better Decisions Under Risk and Uncertainty
Specify the decision and the time horizon
Write down the alternatives that are genuinely available, the objective, the relevant time horizon, and who experiences each consequence. A choice between two advertised products might become a choice among purchasing now, testing first, waiting, or declining. A decision becomes more manageable when the option set and the costs of delay are explicit.
Separate facts, forecasts, and values
Record which outcomes are observed facts, which are probabilistic forecasts, which are assumptions about ambiguous conditions, and which are personal or institutional values. A disagreement about the probability of a loss requires different work from a disagreement about how much that loss matters. This separation reduces the risk of presenting normative preferences as if they were objective measurements.
Use expected utility when the inputs justify it
For well-specified risks, calculate expected outcomes and examine whether the consequences have nonlinear utility. Where an option threatens irreversible financial or physical harm, include the consequences explicitly instead of treating the mean alone as sufficient. If a utility function is only an illustrative assumption, label it as such and test how recommendations change under plausible alternatives.
Use scenarios and sensitivity analysis for ambiguity
If reasonable experts offer competing probabilities or models, evaluate each available action across those assumptions. Highlight assumptions that reverse the preferred choice, and identify options that remain acceptable under multiple scenarios. When useful, compare the consequences of a reversible pilot with a full commitment. A precise probability may be inappropriate where evidence supports only a wide range or qualitative uncertainty.
Seek information that can change the decision
Additional information has value when it is credible, timely, and capable of changing the preferred action enough to justify the cost. Asking for another forecast that merely repeats an existing opinion may add little. A properly designed experiment, second independent measurement, or targeted comparison can be much more informative. Define in advance what result would change the decision; otherwise information search can become indefinite.
Make uncertainty legible
Use clear absolute probabilities, counts, time frames, and distinguish individual and population risks. Show a range when data justify one and make clear whether the range represents sampling uncertainty, forecast uncertainty, or alternative models. Ask decision makers to explain the main downside, likely outcome, and assumptions to verify that the description was understood.
Document the decision process and learn from outcomes
Record forecasts, the evidence used, rejected alternatives, and the reason for choosing. Review decisions after outcomes emerge, allowing for both luck and bad forecasting. Track repeated probabilities in a forecast log and examine calibration rather than celebrating isolated correct guesses. This creates feedback about decision procedures without confusing good results with good prior reasoning.
Use structured decision aids where evidence supports them
Decision aids can display options, consequences, probabilities, and value trade-offs in a consistent way. In healthcare, a 2024 Cochrane systematic review concluded that patient decision aids probably improve informed, values-congruent choice and improve knowledge and accurate risk perception compared with usual care, while finding no difference in decision regret and leaving some downstream outcomes uncertain. Those findings support the promise of structured assistance in the studied clinical contexts; transfer to every consumer, legal, organizational, or AI-mediated decision is a separate empirical claim. Stacey et al., 2024.
Why Debiasing Training Has Limits
Instruction about heuristics, probability, or known biases can improve performance on specific educational tasks, yet durable transfer to unrelated high-stakes decisions is harder to establish. A laboratory demonstration may teach people the correct response to the particular format practiced without changing their everyday evidence search, value judgments, or use of feedback. Practical evaluation should therefore include delayed tests, new task formats, and behavior under actual incentives. Researchers distinguish improved knowledge, improved calibration, reduced experimental bias, and improved real-world outcomes.
The strongest choice-support strategy for many settings may combine an appropriate formal representation with a practical workflow: define objectives, elicit beliefs, test assumptions, compare consequences, discuss uncertainties, and revisit the decision when relevant evidence arrives. Its success depends on reliable inputs, meaningful user comprehension, and an environment in which the decision maker can act on the result. Formal tools enhance judgment when their assumptions remain visible.
Decision-Making Under Risk in the Artificial Era
The Artificial Era introduces decision environments in which people and artificial systems jointly produce forecasts, rankings, and recommendations. These systems can expand access to analysis while also making uncertainty harder to inspect when the reasoning behind an output is opaque. A model-generated probability is an estimate conditioned on data, modeling choices, and deployment context. Its numeric precision should not be confused with independently verified reliability.
A responsible human–AI decision workflow distinguishes the uncertainty of the world from the uncertainty of the model, exposes input assumptions, checks the system against relevant cases, and determines who is accountable for irreversible actions. Where a system recommends an option, the scientific questions include how the human calibrates trust, how the information changes perceived risk, and how an override can be used. The general psychological theory of risk and ambiguity provides a foundation for studying such choices; empirical validation of particular AI systems must be established separately.
The phrase Artificial Era supplies the project's historical and philosophical horizon. Statements about the status of Artificial Sapiens belong to that philosophical framework; expected-utility axioms, behavioral experiments, and forecasting outcomes supply the empirical framework used in this scientific article. The two levels have distinct methods of justification.
Frequently Asked Questions
What is the difference between risk and uncertainty?
In the classic distinction, risk involves sufficiently specified probabilities over outcomes, while uncertainty involves unknown or inadequately established probabilities. Ambiguity is a commonly studied form of uncertainty in which the possible outcomes are recognizable but their probabilities are imprecise or unavailable. Everyday language uses these terms more loosely, so a scientific account should state the adopted definition.
Is choosing a safe option a cognitive bias?
A safe choice can be fully consistent with expected utility if the decision maker places high utility on avoiding a large loss. Calling it biased requires a justified benchmark and evidence that the choice departs from that benchmark under the assumptions of the task. A single preference does not establish a mental disorder or a stable reasoning defect.
How is expected value different from expected utility?
Expected value averages outcome magnitudes using probabilities. Expected utility averages the utility assigned to those outcomes. They agree when utility is linear in the measured outcome, but otherwise a higher average monetary payoff can have lower expected utility for a particular decision maker.
What is the Ellsberg paradox?
The Ellsberg paradox shows that some patterns of preference between wagers with known and unknown probabilities cannot be represented by a single additive subjective probability distribution under the standard assumptions. It helped establish ambiguity-sensitive choice as a distinct area of decision theory.
Does prospect theory prove that everyone fears losses twice as much as gains?
Prospect theory models reference-dependent evaluation and asymmetric gains and losses under particular assumptions. Numerical parameter estimates vary with task, population, stakes, reference point, and model specification. No universal multiplier follows from the foundational theory for every real-world decision.
Why can people react differently to the same probability?
The outcomes may have different personal value; descriptions may establish different reference points; people may distrust the probability source; or they may rely on different experiences and samples. Identical numeric probabilities do not guarantee identical meanings, motivations, or preferences.
Which mathematical model should be used for an uncertain decision?
Choose the model to match the decision environment and objective. Expected utility is appropriate for many decisions with credible specified probabilities and coherent utility comparisons. Subjective expected utility represents uncertainty through a single set of subjective probabilities. Ambiguity-sensitive or robust approaches may be more informative when several probability models remain credible. Predictions should be evaluated against new data and sensitivity to assumptions.
Can decision-making under uncertainty be improved?
People can often improve decisions by defining alternatives, distinguishing facts from beliefs, communicating absolute probabilities, considering consequences and values, and using structured comparisons. Empirical research supports some decision aids in particular settings, especially healthcare. Claims about lasting benefits of debiasing training or transfer across domains require separate evidence.
Is risk preference a stable personality trait?
Some risk-preference measures have meaningful stability, but a large 2025 synthesis found substantial differences across domains and measurement approaches and low agreement across different instruments. The more informative question is what kind of risk preference was measured, with which procedure, in which population, and over what interval.
How should I interpret a probability supplied by an AI system?
Treat it as a model-dependent forecast. Check its task-specific calibration, data provenance, validation on relevant examples, and changes in conditions. Compare alternatives and consequences, and assess whether the model's uncertainty estimate reflects unknown events and alternative plausible models. A plausible number alone does not validate a recommendation.
Related Articles
Neuroticism and Decision-Making: Risk, Uncertainty, and Negative Emotion — individual differences in negative emotion and their relevance to decision processes.
Power and Risk Taking: Confidence, Approach Motivation, and Decision-Making — power, accountability, and motivational influences on risk behavior.
AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making — credibility and reliance on automated decision support.
Judge Decision-Making: Psychology, Expertise, and Bias — specialized institutional decisions under evidentiary uncertainty.
Anxiety and Intolerance of Uncertainty: Worry, Avoidance, and Decision-Making — the clinical and affective context of uncertainty sensitivity.
Prospect Theory: How Gains, Losses, and Reference Points Shape Choice — formal and behavioral models of reference-dependent risky choice, probability weighting, and the structure of gains and losses.
Loss Aversion: Definition, Evidence, and Scientific Debate — experimental evidence and theoretical debates about asymmetric responses to gains and losses.
Endowment Effect: Why Ownership Can Change Valuation — Examines how ownership influences stated value and preference.
Ambiguity Aversion: Why Unknown Probabilities Change Preferences — Explains choice when outcome probabilities are incomplete or unreliable.
Regret Aversion: How Anticipated Regret Influences Choices — Studies how anticipated regret influences decisions.
References
Bagaïni, A., Liu, Y., Kapoor, M., Son, G., Bürkner, P.-C., Tisdall, L., et al. (2025). A systematic review and meta-analyses of the temporal stability and convergent validity of risk preference measures. Nature Human Behaviour, 9, 700–712. https://doi.org/10.1038/s41562-024-02085-2.
Boun My, K., Brunette, M., Couture, S., and Van Driessche, S. (2024). Are ambiguity preferences aligned with risk preferences? Journal of Behavioral and Experimental Economics, article 102237. https://doi.org/10.1016/j.socec.2024.102237.
Echenique, F. (2020). New developments in revealed preference theory: Decisions under risk, uncertainty, and intertemporal choice. Annual Review of Economics, 12, 299–316. https://doi.org/10.1146/annurev-economics-082019-110800.
Ellsberg, D. (1961). Risk, ambiguity, and the Savage axioms. Quarterly Journal of Economics, 75(4), 643–669. https://doi.org/10.2307/1884324.
Epstein, L. G., and Schneider, M. (2010). Ambiguity and asset markets. Annual Review of Financial Economics, 2, 315–346. https://doi.org/10.1146/annurev-financial-120209-133940.
Fischhoff, B., and Broomell, S. B. (2020). Judgment and decision making. Annual Review of Psychology, 71, 331–355. https://doi.org/10.1146/annurev-psych-010419-050747.
Gigerenzer, G., and Gaissmaier, W. (2011). Heuristic decision making. Annual Review of Psychology, 62, 451–482. https://doi.org/10.1146/annurev-psych-120709-145346.
Hertwig, R., Barron, G., Weber, E. U., and Erev, I. (2004). Decisions from experience and the effect of rare events in risky choice. Psychological Science, 15(8), 534–539. https://doi.org/10.1111/j.0956-7976.2004.00715.x.
Hertwig, R., and Erev, I. (2009). The description–experience gap in risky choice. Trends in Cognitive Sciences, 13(12), 517–523. https://doi.org/10.1016/j.tics.2009.09.004.
Höppner, M. (2025). Models in decision-making under risk and uncertainty. Journal of Economic Surveys, first published August 19, 2025. https://doi.org/10.1111/joes.70008.
Kahneman, D., and Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291. https://doi.org/10.2307/1914185.
Karni, E. (2014). Axiomatic foundations of expected utility and subjective probability. In Handbook of the Economics of Risk and Uncertainty, Vol. 1, 1–39. https://doi.org/10.1016/B978-0-444-53685-3.00001-5.
Karni, E., Maccheroni, F., and Marinacci, M. (2015). Ambiguity and nonexpected utility. In Handbook of Game Theory with Economic Applications, Vol. 4, 901–947. https://doi.org/10.1016/B978-0-444-53766-9.00017-3.
Mishra, S. (2014). Decision-making under risk: Integrating perspectives from biology, economics, and psychology. Personality and Social Psychology Review, 18(3). https://doi.org/10.1177/1088868314530517.
Stacey, D., Lewis, K. B., Smith, M., et al. (2024). Decision aids for people facing health treatment or screening decisions. Cochrane Database of Systematic Reviews, 2024(1), CD001431. https://doi.org/10.1002/14651858.CD001431.pub6.
Starmer, C. (2000). Developments in non-expected utility theory: The hunt for a descriptive theory of choice under risk. Journal of Economic Literature, 38(2), 332–382. https://doi.org/10.1257/jel.38.2.332.
Tversky, A., and Kahneman, D. (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5, 297–323. https://doi.org/10.1007/BF00122574.