Advice discounting occurs when a person revises an initial judgment less toward another person’s estimate than the evidence would justify. In the judge–advisor research paradigm, the judge makes an estimate, receives advice and then provides a final estimate. Researchers can measure how far the final estimate moves toward the advisor. The effect is not simply stubbornness: private access to one’s own reasoning, uncertainty about the advisor, social relationships and prior feedback all influence how competing information is weighted.
The practical improvement is to record an independent initial view, establish what evidence the advisor actually has, compare reliability and independence rather than confidence alone, calculate the movement from the original answer and then check performance against outcomes. Taking every piece of advice at face value would be just as poor as dismissing all advice. Sound intelligence is the ability to update judgments proportionately while respecting different expertise, incentives, correlated sources and genuine uncertainty.
Contents: Why a second opinion changes the cognitive problem · The judge–advisor system and what it measures · Work through one estimate and its denominator · Why the midpoint is a useful baseline, not a command · When precision weighting gives a different answer · Privileged information makes self-confidence feel stronger · Advisor reputation is a dynamic evidence problem · Choosing, averaging and selective correction are different · How advisor distance affects revision · Agreement can conceal correlated mistakes · A teacher–student example with explicit outcomes · A workplace forecasting example with competing incentives · The role of confidence and calibration · Advisory quality is specific to topic and context · A five-stage protocol for better advice use · What to do when advice is algorithmic · Common diagnostic errors and repairs · Changed-context transfer exercise and answer · Frequently asked questions · Evidence, limitations and further reading
Why a second opinion changes the cognitive problem
Before receiving advice, a decision maker is solving a first-order problem: given available evidence, what estimate or action should I select? Once advice appears, the task becomes second-order as well. The judge must decide how much to trust another person’s information, whether the advisor is better calibrated, what assumptions differ and whether the two estimates represent independent evidence. Those questions are not answered by simply comparing which person sounds more confident. A brilliant explanation can rely on weaker data, while a modestly stated estimate can arise from a reliable base rate.
The judge usually remembers why the initial view seemed sensible. They recall the observations, experiences and objections that led to it. They may know little about the advisor’s internal process. That asymmetry creates an apparent evidential imbalance: one’s own reasons feel numerous and concrete, while someone else’s judgment can appear as a bare number. Yaniv and Kleinberger’s 2000 work treats privileged access to one’s own reasons as an important explanation for discounting advice. This is a proposed psychological account grounded in their experiments, not a rule that every disagreement has the same cause.
Decision quality requires separating two questions that often get confused. Which estimate is likely more accurate, given each person’s track record and information? Which estimate feels more defensible to the judge, given the arguments they can explain? Those questions need not produce the same ordering. An educational or organisational setting can reduce the gap by asking advisors to disclose relevant evidence, assumptions and uncertainty without treating elaborate storytelling as proof. The content of the evidence matters more than the sheer number of reasons offered.
The judge–advisor system and what it measures
A typical judge–advisor study has three stages. Participants first provide their own quantitative estimate for a question with an objective or later-checkable answer. They then see an advisor’s estimate, sometimes accompanied by information about its reliability, and finally can revise their judgment. The design can isolate the social-information stage because the initial estimate is recorded before the adviser intervenes. The original experimental work from 2000 involved several studies on advice weighting and advisor reputation, but one should consult the original methodology before generalizing its numerical findings to every workplace.
The common measurement is weight of advice, or WOA: (final estimate minus initial estimate) divided by (advisor estimate minus initial estimate). If the final estimate stays at the initial answer, WOA = 0. If it moves all the way to the advisor’s answer, WOA = 1. If it lands exactly halfway between the two, WOA = 0.5. These reference points let researchers compare shifts across questions measured in different units, but the fraction is meaningful only when initial and advisor estimates actually differ.
The WOA label can be misleading if read too literally. A weight of 0.25 does not mean the participant believed the advisor was exactly 25% likely to be correct. It means the final numerical answer moved one quarter of the distance toward the advisor’s estimate. Judgments can also move away from advice, producing negative WOA, or overshoot it, producing a value above one. Researchers may analyze such cases separately or use conventions that must be stated. The measurement describes expressed revisions, not a mind-reading device for trust.
Work through one estimate and its denominator
Suppose a project manager initially predicts that a review will require 60 hours of work. A colleague advises 40 hours. After hearing the colleague’s reasons, the manager revises the estimate to 55 hours. The numerator is final minus initial = 55 – 60 = -5 hours. The denominator is advice minus initial = 40 – 60 = -20 hours. WOA = (-5 hours) / (-20 hours) = 0.25, or 25% of the distance toward the advice. Hours cancel, leaving a dimensionless measure of revision.
That 25% describes movement, not improvement. If the project actually takes 50 hours, the original error was |60 – 50| = 10 hours, the advisor error was |40 – 50| = 10 hours, and the revised error is |55 – 50| = 5 hours. In this invented example, the midpoint of 50 hours would be exactly right, but no manager could know that simply from the two estimates. The usefulness of a weight must be judged across many cases, not by one conveniently selected outcome.
Now imagine a final estimate of 65 hours instead. WOA = (65 – 60) / (40 – 60) = 5 / (-20) = -0.25. The judge has moved away from the advice. If the final estimate had been 30 hours, WOA = (30 – 60) / (40 – 60) = 1.5, indicating an overshoot past the advisor. These are mathematically legitimate observations. A recording system that silently clips every weight into the interval zero to one would erase behaviors researchers may need to explain.
When advisor and initial estimates are both 60 hours, the denominator is zero and WOA is undefined. A report must not assign it zero or 100% merely to keep a spreadsheet full. There is no numerical distance between the estimates to weight. The judge may nevertheless receive useful explanatory evidence and become more or less confident. Confidence updating in agreement cases deserves a separate measure. Good quantitative research includes explicit handling rules for zero denominators and flags the number of excluded cases.
Why the midpoint is a useful baseline, not a command
For two equally reliable, unbiased and independent numerical estimates of the same quantity, taking their average can often reduce error. If one observer predicts 60 hours and another predicts 40 hours, the midpoint is (60 + 40) / 2 = 50 hours. When the errors are independent and similarly variable, extreme deviations can partly offset one another. This is the statistical intuition behind the wisdom of an independent pair rather than a guarantee that every compromise is correct. Independence and comparable reliability are substantive assumptions, not decorative qualifications.
Consider a different hypothetical task: the initial estimate is 48 kilometres, advice is 30 kilometres, and the true distance is 45 kilometres. The initial absolute error is 3 km; the advisor’s is 15 km. The simple midpoint is 39 kilometres and has an error of 6 km, worse than keeping the original estimate. This does not disprove the usefulness of averaging over suitable populations. It demonstrates why one realization is insufficient and why advisor quality matters. Reliable revision asks whether the information sources are comparably accurate across relevant cases.
A third case exposes dependence. Two colleagues both read the same outdated schedule and independently report the same 35-hour estimate, but one merely repeats the other’s number. Counting them as two independent confirmations gives the copied estimate excess influence. A panel of ten people who all rely on one source can have less independent information than two people who inspect different records. Social agreement is not automatically statistical independence, and a decision procedure should record provenance before assigning weights to multiple opinions.
When precision weighting gives a different answer
If two estimates are conditionally independent, unbiased and have known error variances under an appropriate model, the precision-weighted average gives more influence to the lower-variance estimate. Precision is the inverse of variance, not the inverse of standard deviation. These conditions are demanding: real human estimates can be biased, dependent, non-Gaussian or affected by strategic incentives. Nevertheless, the simple model is valuable for understanding why a rational weight need not be one half merely because two people contributed.
Suppose the judge estimates 70 units with estimated error standard deviation 10 units and an independent advisor estimates 50 units with standard deviation 5 units. Judge variance is 100 squared units; advisor variance is 25 squared units. Their precisions are 1/100 = 0.01 and 1/25 = 0.04 in inverse squared units. The combined estimate is (70 × 0.01 + 50 × 0.04) / (0.01 + 0.04) = (0.7 + 2.0) / 0.05 = 54 units.
The movement from 70 toward advice of 50 is (54 – 70) / (50 – 70) = (-16) / (-20) = 0.80, so this simple precision model assigns an 80% advice weight. Why? The advisor’s error variance is one quarter of the judge’s, making its precision four times as large. The result is not a universal psychological prescription. It shows that a weight below 0.5 may be too small when the advisor has demonstrably better reliability, while the same weight could be too large when the advisor has weaker or dependent evidence.
One can also derive the weight directly. If the judge variance is vJ and the advisor variance is vA, under the stated independent unbiased assumptions the ideal coefficient on the advisor estimate in a linear combination is vJ / (vJ + vA). With vJ = 100 and vA = 25, this becomes 100/125 = 0.8. If both variances are 36, the coefficient is 36/72 = 0.5. The formulas reveal what evidence must be collected: comparative accuracy, error variability and source dependence, rather than impressions of seniority.
Privileged information makes self-confidence feel stronger
People have detailed access to their own reasoning, even when some of it is weak. They recall that they examined one example, consulted an informal source or considered a plausible causal story. The advisor’s compressed answer may not reveal equally rich internal deliberation. This can produce egocentric weighting without conscious arrogance. The person sees an evidence trail for the initial judgment but receives only a conclusion for the alternative. Improving the information exchange can therefore change weights without asking anyone to adopt a submissive or deferential social posture.
Yet transparency has limits. A long explanation can be carefully selected, reconstructed after the decision or shaped to persuade. Someone may offer five confident reasons that are all downstream of the same unreliable assumption. A better evidence-sharing format includes the original observations, their source, dates, relevant comparison cases and what would change the estimate. If the advisor cannot share sensitive data, the judge should record that constraint rather than automatically treating missing detail as incompetence or automatically accepting an opaque assertion.
In classrooms, a student may feel certain that a comprehension answer is correct because they can recount the thought process that produced it. A teacher or peer may suggest a different answer but initially provide only the final choice. Learning improves when both identify the supporting lines in the passage, the inference step and the exact wording of the question. This is not a contest over whose confidence sounds stronger. It is a transition from unobservable private reasons toward inspectable shared evidence.
Advisor reputation is a dynamic evidence problem
An advisor’s track record matters, but reputation is not one stable numerical fact. Yaniv and Kleinberger studied how people formed reputations for advisors and reported an asymmetry: negative information could damage a favorable reputation more readily than positive information built it. Such effects create path dependence. A single conspicuous error may be remembered more vividly than several ordinary correct suggestions, especially when feedback is sparse. This is a descriptive research finding; it is not a recommended personnel policy or a universal law for every culture.
A practical decision log should therefore separate observed accuracy from remembered anecdotes. If an expert made 18 accurate predictions out of 20 comparable forecasts, the recorded frequency is 18/20 = 90%, with a small sample and uncertain transfer to new topics. If another advisor made 9 correct out of 10, its visible rate is also 90%, but the evidence base is smaller. Neither figure alone proves skill: task difficulty, chance, selection and definitions of success can change the interpretation.
For continuous judgments, absolute error, squared error and calibration across matched tasks may be more informative than a correct/incorrect count. An advisor who predicts a delivery date within two days consistently may be useful even when the exact day is rarely right. Another may excel on routine tasks but fail under disruption. Reputation systems should retain task families, evidence dates and observed outcomes so that confidence can improve or decline with relevant evidence instead of being permanently fixed by one dramatic story.
Choosing, averaging and selective correction are different
Research by Soll and Larrick examined ways people revise quantitative estimates and compared choosing between the two estimates with averaging. Their work illustrates why the apparently cautious act of choosing the more credible single person need not outperform combining independent, imperfect information. Averaging can use information that either individual estimate leaves out. Nevertheless, the effectiveness of averaging depends on the error structure and sampling environment; the lesson is to test combination strategies, not to average all opinions in every circumstance.
A person can also make a partial correction without explicit averaging. If a tutor says the answer must be smaller, a student might move from 72 to 68 without considering how the advice was generated. If a manager starts from a sales forecast of 1,000 units and sees an independently produced forecast of 700, revising to 950 uses relatively little of the new information. Recording the two inputs and the final judgment permits analysis of whether these small moves were appropriate given subsequent outcomes.
Combining structured evidence often performs differently from combining naked numbers. Suppose one expert has a reliable model of normal demand and another detects a verified holiday closure. A sensible final forecast may not be a fixed 50/50 blend, because the holiday information changes the underlying context in a systematic way. Before applying a generic weight, ask whether the advisor contributes a new causal variable, duplicates known information, corrects a measurement error or merely expresses a different preference.
How advisor distance affects revision
Distance is the gap between the judge’s initial estimate and the advice. Advice ten units away may feel like a plausible adjustment; advice one hundred units away may prompt doubts about competence or assumptions. But extreme advice can also be the first signal that the original model is missing something important. A decision maker should not treat numerical distance alone as proof of poor advice. Instead, it should trigger an evidence check: are the two people estimating the same quantity, horizon, units and population?
Consider a project with a manager predicting 20 calendar days and an adviser predicting 60 working days. The numbers differ drastically, but much of the conflict may be definitional rather than evidential. Another case might compare 60 minutes with 60 hours by accident. Effective judgment revision begins with unit normalization, a shared reference class and consistent inclusion rules. Only after aligning the target can the gap be interpreted psychologically or statistically. Otherwise a measured WOA can be numerically precise while describing a meaningless comparison.
When large disagreement remains after definitions are aligned, investigate competing hypotheses. One estimate may incorporate supplier delays that the other omits; one may be an optimistic best case, the other a median. Ask both parties to describe what observations would decisively support their number. That method yields discriminating evidence and a testable forecast, instead of turning a substantive disagreement into a negotiation over whose status commands deference.
Agreement can conceal correlated mistakes
If two colleagues agree, there is no numerical movement to measure with WOA. Their agreement may still be informative when they have independent, accurate methods. But if they reviewed the same spreadsheet cell, an identical answer is not a second independent measurement. A team that simply counts matching statements can mistake repetition for confirmation. Independence requires tracing source lineage, shared assumptions and communication between advisors before they generated their estimates. This is particularly important when many reports repeat one original dataset.
The practical remedy is independent-first assessment. Ask participants to record estimates before group discussion, then disclose sources and compare assumptions. If the group receives a new common fact afterward, everyone may rationally converge. The key distinction is whether convergence arose from shared valid evidence or from social imitation. Meeting notes should record reasons and unresolved questions rather than only the final consensus number. This supports later learning when an outcome reveals which assumptions were right or wrong.
For teams using several AI assistants, the same principle applies. Three model responses can rely on one copied web page or one common training-derived claim. It would be mistaken to treat their agreement as three independently verified sources. Ask for primary records and independently check decisive quantities. The mechanism of advice discounting is about human revision; the independence problem also matters for machine-mediated advice, but the two should not be conflated into one psychological diagnosis.
A teacher–student example with explicit outcomes
Imagine a learner predicts that a word problem’s answer is 36 metres, while a tutor suggests 24 metres. The learner finally writes 33 metres. The measured movement is (33 – 36) / (24 – 36) = -3 / -12 = 0.25, or one quarter of the distance toward the tutor’s answer. This tells us about revision, not whether either number is right. If the units should have been square metres, both estimates are categorically invalid even though the calculation of WOA is arithmetically correct.
The tutor can change the conversation by asking the learner to show the diagram, state the quantity sought and identify the operation used. Suppose the class discovers that the question asked for a perimeter and the figure contains six equal sides of four metres each. Then 6 × 4 = 24 metres is justified by inspectable structure. The reason to revise to 24 is not the tutor’s position of authority but the correct interpretation of the mathematical model. A good teaching practice distinguishes evidence-supported update from simple conformity.
On the next question, the tutor should change the shape or numerical arrangement. If the student correctly derives a new perimeter without relying on teacher hints, it demonstrates transfer of the underlying reasoning. One correct revised answer alone cannot establish deep learning because the learner may have copied the advisor. This distinction helps tutors and parents avoid rewarding superficial compliance while overlooking whether independent judgment improved.
A workplace forecasting example with competing incentives
Imagine a service team predicts the duration of a planned outage. The operations manager estimates 80 minutes, while an engineer advises 40 minutes. The manager revises to 70 minutes, giving WOA = (70 – 80) / (40 – 80) = (-10) / (-40) = 0.25. If the actual outage lasts 50 minutes, the original error is 30 minutes, the advisor’s error is 10 minutes and the revised error is 20 minutes. The example shows underuse of more accurate advice on this one case; it does not by itself establish a general pattern.
The organisation should record whether the engineer had direct measurements that the manager lacked, whether the manager considered hidden approval delays, and whether either party faced incentives to present a reassuring number. A manager may deliberately include contingency in a published service promise, which is a different objective from estimating the expected technical duration. Before diagnosing cognitive bias, establish whether both estimates represent the same target: a mean, median, worst-case bound or contractual deadline.
Later feedback should be evaluated by decision objective. If the team committed to restore within 70 minutes and finished in 50, the commitment was satisfied, but that does not prove the underlying time estimate was well calibrated. Conversely, a forecast mean of 50 could be accurate across many events while leaving little contingency for a high-consequence outage. Forecast accuracy, confidence interval coverage and service commitments are distinct variables. Good institutional intelligence keeps them separate.
The role of confidence and calibration
Someone can be confident but inaccurate, or tentative but well calibrated. Confidence is useful only when it meaningfully predicts correctness or uncertainty across comparable questions. If a forecaster assigns 80% confidence to many binary events, a basic calibration check asks whether roughly 80% of those events occur over a large, appropriately chosen set. Such data cannot be inferred from the tone of a single explanation. Confidence wording alone is therefore a weak basis for weighing advice.
For numerical estimates, ask for a plausible interval with an explicit definition: for example, a central 80% interval for the quantity being predicted. Intervals give judges a way to compare information and disagreement more richly than two points. If the judge predicts 60 hours with an 80% interval of 45 to 90, and the advisor predicts 40 with an 80% interval of 35 to 50, their overlap near 45 to 50 is informative but does not resolve which model is better. Evaluate calibration and sharpness over repeated outcomes.
Do not punish advisors solely for admitting uncertainty. If a process promotes the most assertive speaker and later labels cautious speakers indecisive, it can select for poor calibration. Encourage estimates that include conditions and identify what new observation would change the view. Structured uncertainty reports can reduce empty persuasion, although they take work and may not be proportionate for trivial decisions. The goal is not maximal quantification of every conversation; it is better weighting when decisions matter.
Advisory quality is specific to topic and context
An experienced project manager may forecast staffing accurately but misjudge network latency. A skilled tutor may diagnose student misconceptions but be less informed about a new regulatory deadline. Accuracy in one domain should not be transferred automatically into unrelated domains. Relevance matters because advisor quality is conditional on task, evidence and environment. Treat professional credentials as a useful prior clue rather than a substitute for checking pertinent, recent performance.
Context shifts also affect the judge. Someone who knows an older process thoroughly might overvalue familiar assumptions after the organisation changes software. An advisor with direct experience of the new system could be more accurate even with less seniority. Conversely, a newcomer may underestimate historical constraints. A thoughtful update begins by writing down what changed and which parts of each person’s evidence transfer. This is a mechanism for organisational learning rather than a personality ranking.
A good after-action record asks three distinct questions: Was the initial estimate close to the outcome? Was the advice better? Did revision improve the decision? If an advisor was less accurate, ignoring them might have been sensible. If advice was more accurate and ignored repeatedly, the team may be leaving information unused. If the combined view performed best across a stable class of tasks, a structured combination rule may help. One dataset rarely answers all three without careful comparisons.
A five-stage protocol for better advice use
Stage one is independent commitment. Record your estimate, the specific question, units, deadline and reasons before hearing others. This reduces retrospective reconstruction, where a person later remembers having predicted the outcome. It also lets a team distinguish genuine convergence from contamination through prior discussion. For a school project, a learner can write an initial answer and why it seems valid. For a team forecast, preserve the versioned input rather than overwriting it when advice arrives.
Stage two is advisor evidence. Record who or what supplied the recommendation, what sources it used, whether it is independent of the initial judgment and which assumptions differ. Ask for uncertainty when it would change the decision. If privacy or confidentiality prevents full disclosure, mark the constraint. Do not reward a long persuasive story unless its facts can be checked. The advisor should be able to say what would falsify the suggestion rather than merely restating confidence.
Stage three is explicit revision. For quantitative judgments with different initial and advisor values, calculate WOA from the original and final numbers. If the judge moves away or overshoots, retain that fact rather than coercing every observation into zero to one. For qualitative decisions, document the changed premise or option instead of inventing a numerical weight. The calculation is a diagnostic record, not a target every participant must hit.
Stage four is outcome checking. When the answer becomes known, compare initial, advice, final and possibly a simple average against the same truth and error rule. Preserve a denominator: number of decisions evaluated, not just the subset that ended in success. Split by task type and evidence quality if enough cases exist. An improvement in average error accompanied by major failures in rare safety-critical cases may warrant a different objective and review.
Stage five is rule revision. If independent advisers consistently beat initial forecasts, consider giving them more weight in that class of decisions. If advice repeats a common bad source, fix the evidence pipeline before changing social norms. If decisions are too heterogeneous for one universal weight, use a decision-specific check instead. Institutional intelligence grows when learning becomes a recorded process rather than a vague feeling that the team is getting better.
What to do when advice is algorithmic
An AI assistant may provide an estimate, a recommendation or an explanation with fluent language. The psychological temptation to overdiscount or overaccept advice still appears at the human interface, but the machine’s error characteristics are not identical to those of a human advisor. Its apparent confidence is not necessarily calibrated probability, and its sources may be missing or incorrect. Ask what data the output relies on and what independent checks confirm the most consequential details.
A model that proposes an answer to a mathematics question should be judged through derivation and substitution, not through its language style. In a research task, verify primary sources rather than counting several AI paraphrases as corroboration. In an operational task, distinguish a recommended action from permission to execute it. A person can use AI advice as a hypothesis generator while retaining final accountability, especially for irreversible or sensitive steps.
When measuring advice taking from AI, define the experimental contrast carefully. An algorithm recommendation accompanied by an explanatory rationale can influence users differently from a bare number, and knowledge of the source may change responses. Advice distance, perceived expertise and task stakes may moderate the effect. Do not call every refusal to adopt an AI output irrational bias; the decision may reflect legitimate uncertainty about machine competence or hidden costs.
Common diagnostic errors and repairs
The first error is measuring only agreement after the meeting. A group that ends unanimously may have ignored important evidence, or it may have integrated evidence well. Without independent initial judgments there is no way to tell. Repair this with short private estimates before discussion on high-stakes questions. Agreement is an outcome, not proof of sound weighting. Also preserve dissenting assumptions in minutes where they matter, rather than presenting consensus as though it was inevitable.
The second error is treating a change toward advice as inherently good. An advisor may be misinformed, conflicted or wrong about which quantity is being estimated. Repair this by scoring final decisions against observable outcomes and by checking expertise in the relevant context. A team that follows poor advice enthusiastically is not displaying superior intelligence. Learning requires detecting errors from both the original and external sources.
The third error is interpreting a very small WOA as a stable personality trait. One case can be rationally resistant to advice when the judge has strong independent evidence; another can reflect status pressure or unequal information. Repair this by collecting multiple comparable cases and recording source independence, expertise and framing. Do not turn a descriptive measurement into a diagnostic label for a colleague or student based on one exchange.
The fourth error is treating numerical precision as scientific certainty. A WOA of 0.247 can be calculated exactly from arbitrary estimates, but its psychological interpretation still depends on the experimental design. Rounding to two decimals is usually enough for an illustrative record. Report counts, missing cases and zero denominators. If a dashboard displays average WOA, explain how it handles negative values, overshoots, categories and weights across unlike tasks.
Changed-context transfer exercise and answer
A hospital facilities team, in a hypothetical nonclinical operations scenario, estimates how long a lift repair will take. The supervisor’s initial estimate is 12 hours, an engineer’s advice is 8 hours, and the supervisor’s final estimate is 11 hours. First calculate WOA. Then suppose a second engineer also says 8 hours after copying the first engineer’s worksheet. Should the supervisor treat two matching estimates as two independent observations? Finally suppose the repaired lift becomes available after 9 hours. Compare the three numerical errors.
WOA = (11 – 12) / (8 – 12) = (-1) / (-4) = 0.25. The copied second opinion offers no new independent measurement merely because a new person repeats the number. If actual availability was after 9 hours, the initial error is |12 – 9| = 3 hours, the advisor error is |8 – 9| = 1 hour and the revised error is |11 – 9| = 2 hours. In this example more movement would have helped, but another repair might favor the supervisor.
The changed context matters because engineering evidence can vary between buildings and failures. Before institutionalizing an 80% advisor weight, the team should compare several comparable repairs, verify independence and ask what each forecast included. An estimate of time until a mechanic finishes work is not necessarily the time until the lift is inspected and released for service. The exercise tests source provenance and target definition as well as arithmetic; these are the transferable skills that prevent overconfident generalizations.
Frequently asked questions
Is advice discounting always irrational?
No. If the judge has better evidence, the advisor is unreliable or both rely on the same source, giving little weight to advice may be justified. The term usually describes a systematic pattern that can lead to underuse of useful independent information. Whether a particular instance is excessive requires relevant accuracy, dependence and objective-function evidence. One cautious revision cannot establish a general bias.
Does a WOA of 0.5 mean perfect judgment?
No. It means the final numerical estimate is exactly halfway between the initial and advisor estimates. That can be a useful baseline for equally reliable independent estimates, but it can be poor when one source is much more accurate or both are biased together. Evaluate the decision against outcomes and assumptions, not against a mandated midpoint.
Can advice weights be negative or greater than one?
Yes. A final estimate moving away from the advisor produces a negative ratio. Moving beyond the advisor can produce a ratio greater than one. Both are mathematically possible. A research report should say how these cases are analyzed, and never silently clip them into a prettier scale without justification. If initial and advisor estimates match, WOA is undefined because the denominator is zero.
How can a teacher apply this without turning class into statistics?
Ask learners to write an independent answer and a reason before seeing another answer. Compare methods, locate the decisive evidence, and then have them explain exactly what made them revise. Use new problems to check transfer. The WOA calculation is optional; the underlying practice is evidence-based updating rather than copying a stronger voice.
What if an AI assistant provides the advice?
Treat it as one source with task-dependent reliability. Inspect source quality, calculations, data provenance and the cost of a mistake. A fluent answer is not independent corroboration of a claim, especially when several systems may share information sources. Use independent tests and human review when a decision affects safety, rights or irreversible operations.
Evidence, limitations and further reading
The foundational primary study is Yaniv and Kleinberger (2000), Advice Taking in Decision Making: Egocentric Discounting and Reputation Formation, with bibliographic details at PubMed PMID 11056071. Its findings concern studied judgment tasks and should not be presented as universal percentages applying to all occupations, ages or cultures. The explanation involving privileged access to one’s own reasons is a theoretical account supported by the research, not an infallible diagnosis of an individual.
A second original research source is Soll and Larrick (2009), Strategies for Revising Judgment: How (and How Well) People Use Others’ Opinions, also indexed at PubMed PMID 19379049. It examines choosing versus averaging strategies and their effectiveness across judgment environments. This guide’s numerical cases were constructed for instruction rather than extracted from those studies, and every conclusion about them remains bounded by the stated assumptions.
For nearby human-intelligence mechanisms, read How Intelligence Works | Common Knowledge — Public Signals and Nested Beliefs on shared signals, How Intelligence Works | Memory Conformity — Co-Witness Discussion and Independent Recall on social contamination of reports, and How Intelligence Works | Competency Traps — Routines and Organisational Learning on repeated decisions. The How Intelligence Works guide collects the wider learning pathway.
Advice taking is a mechanism of judgment revision, not a complete account of group intelligence. Social power, incentives, communication rules and the quality of available evidence also shape decisions. For consequential choices, bring in domain experts and appropriate independent verification; a conversational method is not a substitute for medical, legal or safety-critical procedures. The most useful habit is precise: ask what changed in the evidence, update by the strength of that change, then learn from what actually happened.