This article discusses experiments in numerical judgment. Clinical examples are educational, not diagnostic or treatment guidance.
What Is Anchoring Bias?
Anchoring bias is a systematic influence of an initial numerical reference on a later estimate, judgment, or valuation. A person who first considers a high number may subsequently give a higher estimate than another person considering a low number, even when both people receive the same relevant information. The starting value is the anchor; the observable difference between otherwise comparable estimates is the anchoring effect. In experimental psychology, its defining test is whether manipulating the anchor changes responses beyond what the anchor's legitimate informational value warrants. Tversky and Kahneman (1974); Schley and Weingarten (2026).
An anchor may be an asking price, suggested salary, preliminary forecast, sample answer, previous measurement, expert recommendation, or number explicitly described as random. Anchoring becomes especially important when the target quantity is uncertain. It is possible to use a reference number rationally when the source has demonstrated predictive value. The research question is whether the judgment moves toward a number for reasons other than its justified evidentiary contribution.
Consider two independently assigned groups estimating the cost of repairing the same building. One receives the question, “Would the repair exceed $40,000?”; the other asks whether it would exceed $140,000. Both then provide a free numerical estimate. If the second group estimates more, with evidence held constant, the difference supports an anchoring effect. These dollar figures are hypothetical illustrations rather than measured results. The strength of the inference depends on randomized assignment, identical information, an appropriately comparable scale, and transparent analysis.
Anchoring is an empirical judgment phenomenon. A heuristic is a procedure for reaching a judgment, whereas bias denotes a systematic divergence from an identified benchmark. An anchor can guide an adaptive approximation, and an anchoring effect can also produce an unjustified estimate. A logical fallacy is a defect in an argument's structure; a statistical error may occur in numerical computation; a clinical symptom is evaluated in a diagnostic context. None is synonymous with the experimentally measured anchoring effect.
The Classic Experiment: An Arbitrary Number Becomes a Reference
In the influential 1974 study by Amos Tversky and Daniel Kahneman, participants saw a wheel of fortune apparently generate a number that had actually been constrained to stop at 10 or 65. They answered whether the percentage of African countries among United Nations members was higher or lower than the indicated number, then estimated the percentage. The median estimate was 25% after the low anchor and 45% after the high anchor. The 20-percentage-point difference was an observed result from this experimental task, not a constant by which all anchors change all judgments. Tversky and Kahneman (1974); Onuki, Honda, and Ueda (2021).
The wheel mattered because its numerical outcome lacked a meaningful relationship to the percentage being estimated. Participants nevertheless had to compare the target with the arbitrary value before producing their own estimate. This design illustrates two essential research choices: researchers manipulate a reference value, and participants process that number in relation to the target quantity. Merely encountering a number somewhere in the environment is a different intervention and may produce weaker or absent effects.
Tversky and Kahneman interpreted anchoring partly through adjustment from a starting point. Subsequent work revealed that different anchor-generation procedures recruit different cognitive processes. The 1974 experiment is foundational evidence for the phenomenon; it does not establish a single mechanism that explains every anchoring result. Epley and Gilovich (2001); Frederick and Mochon (2012).
Why Anchors Change Numerical Estimates
Insufficient adjustment from a starting value
Under an adjustment account, a person begins with a salient number and moves away from it in the direction required by their knowledge. The process stops when an answer appears plausible, even though more searching or recalculation might lead farther from the initial point. A familiar self-generated reference makes this especially intuitive: when estimating the freezing point of an unfamiliar substance, someone might begin from the known freezing point of water and revise from there. The direction and adequacy of that revision depend on knowledge and effort.
Nicholas Epley and Thomas Gilovich distinguished self-generated anchors from externally supplied ones. Their experiments found process evidence supporting effortful, insufficient adjustment particularly from self-generated starting points. Later experiments indicated that adjustment often ends when a plausible answer is reached unless participants are both able and motivated to search for a more accurate value. These findings caution against treating every externally presented anchor as proof of serial numerical adjustment. Epley and Gilovich (2001); Epley and Gilovich (2006).
Selective accessibility and hypothesis-consistent testing
Under a selective-accessibility account, a comparison question such as “Is the answer higher or lower than 80?” activates information compatible with the proposed number. Considering a high anchor makes high-value examples, attributes, and scenarios relatively accessible, and these activated considerations influence the ensuing absolute estimate. Thus the anchor may change which knowledge is retrieved rather than simply setting the starting point of an arithmetic procedure.
Mussweiler and Strack examined the role of comparison processes, relevant knowledge, and anchor plausibility. Their research found that anchoring depended on what people knew about the target and the extent to which the comparison made plausible, anchor-consistent information accessible. The account predicts important interactions with task wording and memory availability. It also motivates an intervention: deliberately generating evidence that contradicts an anchor can reduce assimilation to it. Mussweiler and Strack (2000); Mussweiler, Strack, and Pfeiffer (2000).
Response-scale distortion
A further account concerns the scale on which people express an estimate. Exposure to an unusually high or low reference can alter what a particular number feels like on a response scale, even if the underlying object is judged similarly. If an anchor changes the interpretation of “heavy” in pounds, later answers in pounds may move without an equivalent shift in a separate way of describing the same object.
Frederick and Mochon's scale-distortion experiments showed that anchor effects can be scale-specific. Their evidence challenges the assumption that all numerical shifts reflect changed beliefs about an object's properties. Researchers therefore need to distinguish transformations of the represented quantity from changes in how a response is expressed. A useful experiment can elicit the same target through distinct units or response formats while testing whether the induced shift follows the scale. Frederick and Mochon (2012).
Deliberation, source credibility, and mixed mechanisms
Externally supplied anchors can affect more elaborated judgments as well as relatively rapid responses. A number from a credible expert, an unverified web page, a negotiation opponent, or a random device can serve different roles because perceived relevance and credibility change how people interpret it. Multiple mechanisms may coexist: comparison activates supporting considerations, the scale acquires a temporary reference, and subsequent adjustment may be incomplete.
Theories of numerical anchoring should be tested with discriminating predictions rather than inferred from the existence of a high-versus-low difference. Wegener and colleagues emphasized the influence of elaboration and thoughtful processing on anchoring. The modern synthesis argues that findings must be organized by the precise operation used to present the anchor and by its relation to the target, rather than compressed into a single universal explanation. Wegener et al. (2010); Schley and Weingarten (2026).
Anchoring Bias Versus Rational Use of Information
A numerical suggestion is not automatically irrelevant. Suppose a laboratory instrument has produced a well-calibrated estimate with a known error distribution. Updating one's own judgment toward that estimate may improve accuracy. In Bayesian terms, a person's posterior belief should depend on the prior belief and the likelihood of the observed evidence, including the external estimate's reliability and dependence on other evidence. A movement toward reliable independent evidence may be normatively appropriate. By contrast, movement toward a wheel-generated number unrelated to the target lacks that information-based justification.
The benchmark must match the task. When the correct factual quantity is known, accuracy can be assessed against it. When the answer is unknowable during the experiment, random assignment and the high–low anchor contrast identify causal influence, while independent calibration data help evaluate whether the resulting estimates are well grounded. For pricing tasks, value may depend on tastes, markets, constraints, and strategic interaction; the experimenter's chosen anchor alone does not establish the uniquely correct market price.
Anchoring is also distinct from a reference point in prospect theory. A reference point may separate perceived gains from losses and affect preferences over outcomes, whereas numerical anchoring principally concerns an earlier value shifting a later quantitative judgment. The two processes can interact in negotiation or valuation, but they should be modeled and measured separately. Likewise, a first offer may carry bargaining information and strategic intent as well as incidental anchoring influence.
How Researchers Measure the Anchoring Effect
Randomized high-anchor and low-anchor experiments
The standard design randomly assigns participants to evaluate a high or low number, then asks for an absolute estimate of the same target. Researchers compare group means, medians, confidence intervals, and distributional overlap. The estimated causal contrast is the expected response under the high anchor minus the expected response under the low anchor, assuming randomization and consistent procedures. For a positive anchoring effect, the contrast points in the anchor's direction. The result should be reported in the original unit as well as a standardized effect size when useful.
An example is the independent-groups estimator difference: estimated anchoring shift = mean(high-anchor estimates) − mean(low-anchor estimates). If both answers are measured in dollars, the difference is in dollars; if they are probabilities, the difference can be reported in percentage points. Changing the distance between the anchors can change the observed difference. Comparing dollar differences across unrelated tasks without accounting for scale, uncertainty, and anchor separation is misleading.
Anchor selection and calibration
Arbitrary anchors are informative only when the experimental manipulation has the intended interpretation. Values that are physically impossible, outside a meaningful scale, or visibly irrelevant may prompt rejection rather than comparison. Extremely plausible anchors can also be informative rather than merely intrusive. Researchers can pretest anchor plausibility and use independent comparison groups to characterize baseline estimate distributions.
Jacowitz and Kahneman proposed using the 15th and 85th percentiles of an unanchored calibration sample as low and high anchors, then locating anchored responses within that baseline distribution. This permits comparison across different estimation questions. Their anchoring index expresses the high–low shift relative to the separation of anchors; its meaning depends on the selected tasks and anchors. It should never be assumed to be a stable personal trait score. Jacowitz and Kahneman (1995).
Process measures and alternative outcomes
A useful study may collect response time, confidence, remembered comparison information, reasons for an estimate, independent accuracy criteria, and the participant's understanding of the anchor's source. Eye movements, sequential response trajectories, and counterfactual explanations can also inform theories. Yet response time or self-reported effort alone does not reveal a unique mental mechanism. Mechanism claims require experimental manipulations that distinguish adjustment, knowledge accessibility, and scale use.
Studies of valuations often measure willingness to pay, willingness to accept, offer amounts, or negotiated outcomes. These are choice-related variables with incentives and constraints, not interchangeable with factual estimates. A researcher should predefine the dependent measure and avoid treating a statistical shift in a hypothetical survey answer as equivalent to realized financial loss.
How Large and Reproducible Is Anchoring Bias?
The 2026 meta-analysis
A major synthesis by Schley and Weingarten, published online May 19, 2026, assembled 2,601 effect sizes from anchoring research, including 1,280 high-versus-low anchor comparisons. For that contrast, the reported pooled standardized result was Hedges' g = 0.825, with a 95% confidence interval from 0.765 to 0.884. The analysis reported I² = 93.73%, meaning variation in effects across studies greatly exceeded what sampling uncertainty alone would explain. The authors report that the overall result remains substantial after analyses addressing publication bias. Schley and Weingarten (2026).
The numerical summary describes the assembled study designs, not the probability that any given person will be anchored or the expected percentage error in every practical decision. A standardized effect of 0.825 is a group comparison on a study-specific scale. The exceptionally high heterogeneity makes moderators and conditions central to interpretation: one should expect very different magnitudes across source types, domains, anchors, procedures, and populations.
The synthesis found reduced or sometimes null anchoring when numbers were merely incidentally present, when the anchor and target referred to different dimensions, or when anchors were introduced as random numbers, as well as under certain incentive and debiasing conditions. It also highlighted whether the anchor supplies directional information. Consequently, the phrase “any random number will anchor any decision” is an overstatement of the evidence. Schley and Weingarten (2026).
Multilaboratory replication and boundary conditions
The Many Labs project coordinated replication of 13 psychological effects across 36 samples, with 6,344 participants overall. The researchers included several anchoring estimates; the anchoring manipulations produced consistent shifts in the expected direction across participating samples, although effect magnitudes differed substantially. These results support the replicability of particular anchoring procedures under coordinated protocols while leaving room for other procedures to have smaller, unstable, or absent effects. Klein et al. (2014).
A set of five experiments reported by Onuki, Honda, and Ueda, with 493 participants in total, investigated which features of a stimulus are required to elicit anchoring. In their tested arrangements, an isolated number and purely semantic cues without a number did not elicit the predicted anchoring; the combined numerical and semantic comparison context mattered. This is a task-specific finding, and the 2026 meta-analysis offers a broader synthesis consistent with treating incidental numerical priming differently from explicit comparative anchoring. Onuki, Honda, and Ueda (2021).
Replications that qualify broad valuation claims
Not every classic valuation result generalizes. Maniadis, Tufano, and List revisited influential claims about anchoring and economically meaningful valuations and found reasons to be cautious about the strength and generality of some initial results. Their paper also explains why research priors, statistical power, independent replication, and the possibility of publication selection matter. A robust laboratory phenomenon can coexist with failures to reproduce a particular downstream application. Maniadis, Tufano, and List (2014).
Replication should therefore be evaluated at multiple levels: whether a manipulation reliably shifts an estimate; whether the same proposed mechanism predicts the shift; whether the difference survives incentives and knowledge; and whether a consequential decision changes outside the laboratory. These are distinct empirical questions. A meta-analytic average offers a high-quality summary of one collection of contrasts, but it does not substitute for fit-for-purpose evidence in each applied domain.
When Is Anchoring Stronger or Weaker?
Target uncertainty and subject-matter knowledge
Anchoring is especially plausible when an estimate must be made with weak or incomplete knowledge. Knowledge can reduce the set of plausible answers and help someone identify an unreasonable reference. However, expertise is not equivalent to immunity. Mussweiler and Strack found that the influence of anchor plausibility depended on participants' available knowledge. Experienced professionals can also anchor under some domain-specific conditions, as studies of property valuation and judicial sentencing illustrate. Mussweiler and Strack (2000); Northcraft and Neale (1987).
To assess expertise fairly, researchers should match the task to what the expert actually knows. An attorney's general legal expertise does not imply precise knowledge of an arbitrary experimentally constructed benchmark; a real-estate specialist may have local market expertise while still responding to a presented listing price. Effective expertise matters most when it supplies precise independent evidence or a tested decision rule.
Source relevance, comparison format, and numerical dimensions
Anchors that appear in a direct higher-or-lower comparison can work differently from incidental numbers encountered before a task. A proposed salary, a forecast, and a random phone number are psychologically different inputs. If the proposed number comes from a relevant measurement process, a shift may include appropriate learning. If it is unrelated to the target's unit or category, the influence may be reduced or absent. These distinctions emerged prominently in the 2026 synthesis. Schley and Weingarten (2026).
Language also matters because asking “Could it be more than 80?” may trigger consideration of cases that would make 80 plausible, whereas showing 80 as background visual noise need not do so. Researchers should record whether an anchor is externally provided, self-generated, task-relevant, random, or embedded in another question. Rewriting an experimental prompt may alter the psychological operation and its result.
Motivation, incentives, and cognitive effort
More effort sometimes helps, particularly when self-generated estimates can be checked against better information. Yet a person can think carefully about anchor-consistent reasons and reinforce the initial number. The mere instruction to “think harder” therefore lacks a general guarantee. The 2026 meta-analysis identified incentives and explicit debiasing conditions among moderators associated with diminished effects in some paradigms, while other experimental findings demonstrate continued anchoring among knowledgeable participants. Epley and Gilovich (2006); Schley and Weingarten (2026).
Individual variation also depends on familiarity with the scale, numeracy, confidence, stakes, and whether usable independent evidence is available. Claims that one personality type or age group is generally immune to anchoring require comparisons across adequately powered tasks and settings. Cultural and language differences deserve targeted investigation rather than extrapolation from university samples.
Examples in Real Decisions
Negotiations and first offers
Negotiators often exchange numbers before agreement on the underlying value has been established. A first offer may establish a focal point for subsequent discussion, but it also communicates an aspiration, information, or bargaining strategy. To show psychological anchoring, a study must separate that informational and strategic content from the effect of the number itself. A very aggressive opening figure might increase later counteroffers in one setting but damage credibility or end a negotiation in another.
A practical safeguard is to develop an independent estimate of the bargaining range and a documented reservation value before hearing the other party's preferred figure. If the other party provides a number, ask for the evidence supporting it. The research supports scrutiny of an early number; it does not establish a universal prescription always to make the first offer.
Prices, property estimates, and consumer valuation
Listing prices and suggested retail prices may become psychological reference values. Northcraft and Neale had students and professional real-estate agents evaluate properties after exposure to manipulated asking prices; their findings were consistent with anchoring in an information-rich valuation task. Nevertheless, market prices depend on comparable transactions, property condition, demand, location, and financing. The initial price can contain genuine market information, and one experiment cannot specify the size of financial harm for all property transactions. Northcraft and Neale (1987).
The boundary with consumer and economic psychology is important: anchoring provides a general mechanism for numerical judgment, while specialized studies of purchase intent, pricing strategy, willingness to pay, and actual transactions require their own evidence. An apparent discount from an initial high price is also partly a framing problem that must be distinguished experimentally from anchoring on a numerical value.
Medical and clinical numerical judgments
A preliminary risk estimate, projected waiting time, symptom score, or proposed treatment cost may shape a later judgment. In a clinical environment, some numerical reference values are essential and evidence-based, such as validated risk prediction models or age-adjusted reference intervals. Their use should be evaluated by calibration, discrimination, measurement validity, and clinical consequences. The anchoring literature justifies a question about influence, but it does not establish that any specific diagnosis or treatment decision was distorted.
Clinicians and patients can compare an early estimate with independent measurements, base rates, alternative hypotheses, and validated decision aids. Such safeguards also address other forms of cognitive error. The presence of anchoring in a hypothetical estimation experiment is not evidence of a psychiatric condition and is not sufficient to judge a clinician's competence.
Legal judgments and sentencing
Experiments by Englich, Mussweiler, and Strack reported that experienced legal professionals' sentencing responses shifted toward irrelevant proposed penalties, including numbers explicitly generated by chance in the experimental materials. The authors also observed greater accessibility of arguments consistent with the anchor. These experiments make the potential mechanism concrete while remaining bounded by their tasks, samples, and legal contexts. Englich, Mussweiler, and Strack (2006).
This general anchoring chapter connects to applied investigations in Judge Decision-Making: Psychology, Expertise, and Bias and Sentencing Psychology: What Influences Punishment Decisions. Those articles address legal procedure, expert responsibility, and the broader structure of sentencing decisions. The experimental finding should not be translated directly into a claim about any particular court verdict.
Scientific forecasts and organizational planning
Teams routinely begin project forecasts from an early estimate, a previous budget, or a suggested deadline. Numbers can become anchors when later revisions are insufficiently responsive to new evidence. Yet forecast error can also arise from incomplete task breakdowns, incentives, optimistic assumptions, or changing external conditions. Identifying anchoring specifically requires evidence that the early number causally shifts the forecast relative to an appropriate comparator.
An organization can request independent estimates, blind reviewers to one another's preliminary figures, document assumptions, and update forecasts using observed base rates. These procedures are sensible safeguards, but their effectiveness should be judged by actual forecasting accuracy, calibration, and decision outcomes rather than by whether participants say they feel less biased.
Human–AI judgment and algorithmic suggestions
A numerical response produced by an AI system can serve as an anchor when a person estimates the same quantity afterward. This is a psychologically plausible application of numerical anchoring, especially when the AI estimate is displayed before independent deliberation. Yet the relevant evidence must distinguish a system's predictive validity from the user's susceptibility to suggestion. If the AI's estimate is well calibrated and informative, movement toward it can be beneficial; an unsupported AI number may mislead.
Designers can test order effects by randomizing whether users make an initial estimate before or after seeing AI advice, and evaluate both absolute error and appropriate reliance. An AI-generated explanation may alter perceived credibility as well as numerical salience. This domain intersects with AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making, which owns broader questions of algorithmic authority, trust calibration, and automation bias. Those constructs and anchoring must be measured separately.
Can Anchoring Bias Be Reduced?
Independent estimates before exposure
The most direct procedural precaution is to estimate the target before seeing an externally suggested number, whenever that is compatible with the decision context. An independent initial judgment supplies a record for later comparison and prevents an early number from being the first explicit candidate. For teams, separately prepared estimates can reduce shared exposure to one person's suggestion. Their value depends on the independence and quality of the information each estimator uses.
Independent estimation is a reasonable evidence-informed design choice, not a guaranteed debiasing treatment across every environment. If participants already share a visible price, common forecast, or prevailing convention, independent written responses may still incorporate the same anchor. Genuine independence should be assessed and, where possible, experimentally varied.
Consider the opposite and seek disconfirming ranges
Mussweiler, Strack, and Pfeiffer tested a consider-the-opposite strategy in which participants generated arguments against the offered anchor. Their studies found reduced anchoring, consistent with the idea that such a procedure changes which target-relevant facts become accessible. It is particularly useful when the decision maker can identify substantive reasons an opening number is too high or too low. Mussweiler, Strack, and Pfeiffer (2000).
A disciplined version asks: “What specific evidence would make this figure implausibly high, and what evidence would make it implausibly low?” Then calculate an independently justified interval before selecting a point estimate. This is preferable to merely replacing one arbitrary number with another. Overcorrection is also possible if the original anchor was genuinely informative, so the goal is proportional, evidence-based updating.
Use reference classes and transparent calculations
When estimating costs, durations, frequencies, or other measurable quantities, build a reference class from comparable observations and use its distribution rather than one conspicuous case. Record units, denominators, sample selection, and uncertainty. If a problem has a known formula or reliable dataset, calculate the estimate directly and test sensitivity to assumptions. Such procedures create an external criterion against which an anchor's pull can be detected.
In probabilistic judgments, use base rates and conditional probabilities explicitly. For example, the chance of a test result given a condition is distinct from the chance of the condition given that result. Anchoring and base-rate neglect are different errors, but sound quantitative workflow can protect against both. Check that comparison examples genuinely match the target population and that data quality warrants the precision displayed.
Incentives, training, and computerized decision support
The 2026 synthesis identifies debiasing interventions and incentives as conditions under which anchoring is sometimes reduced; it does not imply that a single classroom lesson permanently eliminates numerical anchoring across new settings. A training study should test delayed performance, untrained tasks, and consequential outcomes rather than immediate answers on the practiced exercise alone. Schley and Weingarten (2026).
Software can prompt an independent estimate, reveal historical distributions, separate raw data from suggested values, and record the sequence in which information is shown. Such features are design hypotheses requiring evaluation: a numerical dashboard can introduce its own anchors through highlighted defaults, scale choices, and algorithmic forecasts. The best safeguard is measured accuracy and calibrated reliance under realistic use, rather than an interface slogan about eliminating bias.
What Anchoring Research Does and Does Not Establish
The existence of numerical anchoring is well established across multiple laboratory designs, and the 2026 synthesis confirms a substantial pooled high-versus-low contrast. The precise mechanism in an individual experiment often remains a testable question. Insufficient adjustment, selective accessibility, and response-scale distortion produce overlapping predictions in many tasks, and researchers should collect measurements capable of distinguishing them.
The magnitude and practical importance of an observed effect are conditional. A 20-percentage-point shift in an artificial geography question, a standardized mean difference pooled across studies, and the change in money paid in an actual negotiation are three different outcomes. Heterogeneity, publication bias, unequal anchor distances, researcher degrees of freedom, and differences in task instructions affect generalization. Schley and Weingarten (2026); Maniadis, Tufano, and List (2014).
Sampling and culture also matter. Laboratory participants may differ from professional decision makers in incentives, relevant knowledge, and institutional constraints. Multinational replication can demonstrate an effect across participating sites, but it should not be transformed into a claim that all populations and cultures display identical anchoring. Similarly, a group-level shift does not tell us how each participant arrived at a response or whether the same individual would anchor consistently across topics.
When people encounter repeated numeric suggestions in search engines, social platforms, or AI interfaces, multiple processes may operate together: source credibility, selective attention, confirmation bias, social conformity, learning, and anchoring. The study design must isolate the anchor's contribution. The Artificial Era raises new questions about how machine-generated estimates participate in human judgment, but claims about the philosophical status of Artificial Sapiens do not substitute for experiments on numerical decision support.
Frequently Asked Questions
Is anchoring bias always caused by an irrelevant number?
No. Researchers often use irrelevant or randomized numbers to isolate causal influence from genuine information. In everyday decisions, anchors may carry useful evidence, persuasion, and strategic meaning simultaneously. The empirical task is to estimate how much of a shift reflects valid information and how much reflects a task-induced pull toward the starting value. Schley and Weingarten (2026).
Does anchoring mean that the first number always wins?
No. Anchoring is a difference in expected estimates across defined conditions, not a rule that everyone chooses the first option or copies the displayed figure. Strong knowledge, alternative evidence, extreme implausibility, decision procedures, and task wording can diminish or reverse an individual response; some experimental interventions yield no detectable anchoring effect.
Is anchoring the same as confirmation bias?
They are related but distinct. Anchoring concerns a numerical reference shaping a later quantitative response. Confirmation bias concerns preferential search, interpretation, or retention of information consistent with an existing belief. Selective accessibility offers one reason the phenomena can interact: an anchor may prompt consideration of evidence that makes it seem plausible. Their operational definitions and outcomes remain separate.
Is anchoring a mental disorder or a sign of low intelligence?
Anchoring is studied as a judgment effect in ordinary cognition and appears in samples with varied expertise. An experimental shift toward a reference number does not diagnose a disorder or measure global intelligence. Competence should be assessed by the person's domain knowledge, the task, the available evidence, and validated measures appropriate to the question.
Can an accurate first estimate still be an anchor?
Yes. Accuracy and anchoring are not mutually exclusive. A correct preliminary result can become a reference, and later movement toward it may improve the final estimate. The normative question is whether the information warrants that influence and whether new contrary evidence is adequately incorporated.
Does expertise prevent anchoring?
Expertise can provide better benchmarks and exclude implausible estimates, yet studies have also detected numerical anchoring among real-estate agents and legal professionals in controlled tasks. Findings for one professional task should not be generalized mechanically to every expert decision. Northcraft and Neale (1987); Englich, Mussweiler, and Strack (2006).
What is the best evidence-based way to check an anchored estimate?
Create an independent estimate when feasible, consult an appropriate reference-class distribution, articulate reasons the anchor may be too high or too low, and update using traceable evidence. The consider-the-opposite intervention has direct experimental support for reducing anchoring in some settings. No single method has established universal, durable effects across all consequential decisions. Mussweiler, Strack, and Pfeiffer (2000).
Can AI-generated numbers create anchoring?
An AI output that appears before a user's own estimate is a candidate anchor. The causal effect, if any, depends on the task, quality of the generated estimate, its framing, and the user's reliance. It should be tested with randomized presentation and an independent measure of accuracy. AI assistance can improve judgments when evidence is reliable; an unsupported number may distort them.
Related Articles
Cluster Cognitive Biases, Judgment and Decision-Making — the research chapter covering heuristics, probability judgments, risk, preferences, and decision models.
Forensic Evidence and Cognitive Bias: How Context Can Contaminate Judgment — examines contextual influence and safeguards when specialists interpret evidence.
Judge Decision-Making: Psychology, Expertise, and Bias — explores expertise, legal judgment, and experimental evidence on bias.
Sentencing Psychology: What Influences Punishment Decisions — discusses sentencing decisions and the role of proposed numerical penalties.
AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making — separates algorithmic authority and reliance from the specific problem of numerical anchoring.
References
Englich, B., Mussweiler, T., and Strack, F. (2006). Playing dice with criminal sentences: The influence of irrelevant anchors on experts’ judicial decision making. Personality and Social Psychology Bulletin, 32(2), 188–200. https://doi.org/10.1177/0146167205282152
Epley, N., and Gilovich, T. (2001). Putting adjustment back in the anchoring and adjustment heuristic: Differential processing of self-generated and experimenter-provided anchors. Psychological Science, 12(5), 391–396. https://doi.org/10.1111/1467-9280.00372
Epley, N., and Gilovich, T. (2006). The anchoring-and-adjustment heuristic: Why the adjustments are insufficient. Psychological Science, 17(4), 311–318. https://doi.org/10.1111/j.1467-9280.2006.01704.x
Frederick, S. W., and Mochon, D. (2012). A scale distortion theory of anchoring. Journal of Experimental Psychology: General, 141(1), 124–133. https://doi.org/10.1037/a0024006
Furnham, A., and Boo, H. C. (2011). A literature review of the anchoring effect. The Journal of Socio-Economics, 40(1), 35–42. https://doi.org/10.1016/j.socec.2010.10.008
Jacowitz, K. E., and Kahneman, D. (1995). Measures of anchoring in estimation tasks. Personality and Social Psychology Bulletin, 21(11), 1161–1166. https://doi.org/10.1177/01461672952111004
Klein, R. A., et al. (2014). Investigating variation in replicability: A “Many Labs” replication project. Social Psychology, 45(3), 142–152. https://doi.org/10.1027/1864-9335/a000178
Maniadis, Z., Tufano, F., and List, J. A. (2014). One swallow doesn’t make a summer: New evidence on anchoring effects. American Economic Review, 104(1), 277–290. https://doi.org/10.1257/aer.104.1.277
Mussweiler, T., and Strack, F. (2000). Numeric judgments under uncertainty: The role of knowledge in anchoring. Journal of Experimental Social Psychology, 36(5), 495–518. https://doi.org/10.1006/jesp.1999.1414
Mussweiler, T., Strack, F., and Pfeiffer, T. (2000). Overcoming the inevitable anchoring effect: Considering the opposite compensates for selective accessibility. Personality and Social Psychology Bulletin, 26(9), 1142–1150. https://doi.org/10.1177/01461672002611010
Northcraft, G. B., and Neale, M. A. (1987). Experts, amateurs, and real estate: An anchoring-and-adjustment perspective on property pricing decisions. Organizational Behavior and Human Decision Processes, 39(1), 84–97. https://doi.org/10.1016/0749-5978(87)90046-X
Onuki, Y., Honda, H., and Ueda, K. (2021). What stimuli are necessary for anchoring effects to occur? Frontiers in Psychology, 12, 602372. https://doi.org/10.3389/fpsyg.2021.602372
Schley, D. R., and Weingarten, E. (2026). Fifty years of anchoring effects: A theoretical reintegration and meta-analysis. Management Science. https://doi.org/10.1287/mnsc.2023.03238
Tversky, A., and Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. https://doi.org/10.1126/science.185.4157.1124
Wegener, D. T., Petty, R. E., Blankenship, K. L., and Detweiler-Bedell, B. (2010). Elaboration and numerical anchoring: Implications of attitude theories for consumer judgment and decision making. Journal of Consumer Psychology, 20(1), 5–16. https://doi.org/10.1016/j.jcps.2009.12.003