How Much Advice People Actually Take

Answer to the question

How strong is advice discounting, really?

About two-fifths of it. Pooling 346 effect sizes from 129 independent datasets and 17,296 people, the average weight of advice was about 0.39, where 0.50 would mean treating your own estimate and the adviser's as equally good. That is advice discounting, and it is a shortfall rather than a refusal. On the laboratory evidence, taking more of the advice would usually have been more accurate.

Key takeaways

• Across 346 effect sizes, 129 independent datasets and 17,296 people, the average weight placed on advice was about 0.39 where even weighting would be 0.50.
• Information suggesting adviser quality was the only unique predictor, moving adjustment from 32 per cent to 48 per cent.
• An identical estimate carried more weight when framed as advice rather than as an opinion, across six preregistered experiments with 3,411 people.
• On a Norwegian game show with real stakes, 49 contestants weighted advice at 0.58, above their own estimates.
• Four separate searches found no study measuring advice discounting against a real purchase.

The question everyone asks is why people reject the advice they went to the trouble of requesting. It is the wrong question, because rejection is not what the measurements show. The better question is how much of it they use, and the shortfall has a name: advice discounting. The answer is about two-fifths when half would be defensible.

The five instructions that follow are what the measurements permit, and each one is argued for below:

1. Average your estimate with the adviser's rather than choosing between the two.
2. Ask what the adviser knows, and how they arrived at their number.
3. Give more weight to anything offered as help than to anything offered as an opinion.
4. Discount your own expertise, because experts discount advice hardest and lose accuracy by it. That step rests on one experiment with seventy experts in it.
5. Do not write off an adviser who has made one visible mistake, because the evidence says that withdrawal is the costly move. That step is read off the algorithm studies, the only work here that stages a visible error.

Two of the five run into evidence that reverses, and those reversals are the second half of this piece.

Move further towards advice than feels right

The literature runs on one task. You give a numerical estimate, see somebody else's, then give a final estimate. The weight of advice is how far you moved as a fraction of the gap: zero means you ignored the adviser, one means you adopted their number wholesale, and 0.50 means you treated both estimates as equally good. Phoebe Bailey and four colleagues pooled 346 effect sizes from 129 independent datasets and 17,296 people and put the average at about 0.39 where even weighting would be 0.50. Advice discounting is the word for that gap, chosen carefully, because the behaviour is a shortfall and not a refusal.

Advice discounting is the word for that gap, chosen carefully, because the behaviour is a shortfall and not a refusal.

Ilan Yaniv's own figures are sharper. In one study the mean weight of advice was 0.27 while the best achievable weight on the same data was 0.44. Taking the advice still helped: average error fell by about a sixth in his more knowledgeable group and about a fifth in the less knowledgeable one. He also prints a warning: when the advice was mechanically constructed rather than drawn from real people, the accuracy gain vanished entirely.

Kim Scott puts the governing idea well in Radical Candor, arguing that candour gets measured not at the speaker's mouth but at the listener's ear. The book is a practitioner account and is cited here for that framing, with the evidence coming from elsewhere.

Ask what the adviser has told you

Advice discounting is not indiscriminate, and the meta-analysis names the one thing that reliably moves it. The only unique predictor of the pooled weight was information about the adviser suggestive of advice quality: people adjusted 32 per cent of the way for advisers described in ways suggesting low quality, 37 per cent for neutral and 48 per cent for high. Nothing else survived, including age, gender and the individualism measure the authors used for culture, which they report as a plain null.

The founders explain it by information. Ilan Yaniv and Eli Kleinberger attribute the shortfall to differential information: decision makers have privileged access to their internal reasons for holding their own opinion, but not to the advisers' internal reasons. They considered overconfidence, perseverance and confirmation bias and set them aside as less parsimonious, since their task used neutral questions, isolated terminals and payment for accuracy. An account of advice discounting that reaches for ego is reaching past the simpler one.

Their studies also show how quickly a reputation moves. A good adviser, more accurate than 98 per cent of respondents, still earned only equal weight. When adviser quality fell from good to average the weight placed on the adviser's number collapsed, while a rise from poor to average barely shifted it.

That shortfall costs something, and Richard Larrick and Jack Soll supply the reason. Where two estimates fall on opposite sides of the truth, which they call bracketing, their average must beat the average of the two judges. People do not believe this, and offered two numbers they tend to choose one rather than combine them. Neither full text was available to us, which is why this paragraph carries no figures.

Call it advice, not an opinion

The single most useful result in this literature is a labelling effect. Maxim Milyavsky and Yaniv Gvili ran six preregistered experiments with 3,411 people in total and found that an identical estimate carried more weight, and commanded a higher price, when it was presented as advice than when it was presented as somebody's opinion. Their stated conclusion is that advice framing substantially reduced egocentric discounting, across judgment types and for both good and ecologically sampled advice. The mechanism they test is the helping intention ascribed to the source, and thereby its trustworthiness.

Two adjacent questions have duller answers. Francesca Gino and Don Moore found that people overweight advice on difficult tasks and underweight it on easy ones, and that this held whether the advice arrived automatically or had to be sought out, so soliciting it changed nothing. Janet Sniezek, Gunnar Schrah and Reeshad Dalal found that committing money for expert, but not novice, advice raised both the use of that advice and estimation accuracy. Paying for novice advice did neither, so what the money buys is attention to an expert.

So the label does work that soliciting does not. Whoever is advising should say they are trying to help, and say what they know.

Settle the argument about algorithmic advice

Two literatures look as though they contradict each other, and both deserve their full case. Jennifer Logg, Julia Minson and Don Moore ran six experiments and found that lay people relied more on identical advice when they believed it came from an algorithm than from a person: 0.45 against 0.30 on estimating a person's weight from a photograph, with similar gaps on forecasting song popularity and romantic attraction. When 119 active researchers were asked to predict the result, they predicted the opposite. The paper's own qualification belongs with the finding: participants underweighted advice in both conditions, so appreciation shrinks the shortfall without removing it.

Berkeley Dietvorst, Joseph Simmons and Cade Massey ran five studies in which people chose whether to tie a bonus to a statistical model's forecasts or to a human forecaster's, after watching the model perform, the human perform, both or neither. The model beat the humans in all five studies, and every participant was better off choosing it. In every experiment where experience was manipulated, the people who had watched the model err became significantly less likely to bet on it. Watching the human err did not produce the same withdrawal, and the humans' mistakes were larger.

Mischa Seiter and colleagues replicated that comparison with 1,840 participants and the direction held: 67 per cent chose the algorithm with no performance feedback and 47 per cent after seeing it. They print the original figures beside their own, 54 per cent and 42 per cent, so the gap has widened since 2015.

Both results stand, because the two sides measure different things. Logg's team labels a source and measures how much of its advice is taken, with no performance history. Dietvorst's team supplies a history and measures which forecaster people bet on. The same labelling question, asked about goods rather than about advice, is [what an AI label does to a product](https://thewhyweproject.com/articles/what-an-ai-label-does-to-a-product).

The bridge between them is our reading, and we could not establish whether anyone has published it. In the control conditions of the Dietvorst studies, where nobody had watched the model work, a majority still tied their bonus to the model rather than to the human forecaster, which is appreciation measured inside the aversion paper. The baseline is favourable, and one visible error withdraws the favour faster than it would be withdrawn from a person. Logg and Schlund, in a working paper that has not been refereed, put the contradiction down instead to the difference between answering a hypothetical question and making a real judgement.

The same appreciation paper carries the most uncomfortable result here. Seventy United States national-security professionals and 301 online participants worked through four forecasting tasks, and the experts took substantially less advice than the lay people, weighting it below the thirty per cent the paper says advice research typically observes. The cost was measurable: lay people scored better than the experts when the advice came from an algorithm, because the experts discounted it. The paper's own phrase is that algorithmic advice falls on deaf expert ears, with a cost to their accuracy.

Do not assume the laboratory transfers

Almost every figure above comes from somebody revising a date, a weight or a probability for a small bonus. Nobody bought anything. We looked in four places for the commercial test, namely field experiments on advice discounting against real purchases, retail and salesperson studies measuring weight of advice, online reviews set against transaction data, and financial-adviser field studies, and no study measuring advice discounting against a real purchase came back.

What came back instead was one genuine field study. Erik Lohre and Torleif Halkjelsvik found it: five seasons of the Norwegian game show Alle mot 1, with 49 contestants making 382 judgments, soliciting advice on 147 of them, and stakes rising towards about ten thousand United States dollars. The average weight of advice was 0.58. At high stakes, in public, having asked for it, contestants weighted advice above their own estimates, which is the reverse of advice discounting.

At high stakes, in public, having asked for it, contestants weighted advice above their own estimates, which is the reverse of advice discounting.

> The results suggest that highly controlled lab studies may not give an accurate representation of advice taking in high stakes and highly public settings.

Carry the qualifications with the figure. The sample is small and was screened by a television production company, and the authors say their contestants probably score higher on extraversion and openness than the Norwegian population. They also report that estimates did not improve much after advice taking, and that the gain available from averaging was negligible in that setting, which cuts against the cost story as hard as the 0.58 cuts against the discounting story.

Everything here is a judge-advisor estimation task, and in that paradigm advice is a number to be averaged, while a shop assistant's recommendation or an expert review is an argument, a product and a social act. A machine's recommendations are a different question, taken up in [How Much Recommendations Really Sell](https://thewhyweproject.com/articles/how-much-recommendations-really-sell), and ratings in [Why We Trust Online Ratings Anyway](https://thewhyweproject.com/articles/why-we-trust-online-ratings-anyway).

So the rule is narrow, and inside its scope it holds. Average rather than choose, ask what the adviser knows, treat offered help as help, doubt your own expertise before you doubt theirs, and keep an adviser who has erred once. Take about half.

Common questions

How strong is advice discounting, really?

About two-fifths of it. Pooling 346 effect sizes from 129 independent datasets and 17,296 people, the average weight of advice was about 0.39, where 0.50 would mean treating your own estimate and the adviser's as equally good. That is advice discounting: people do move, and they stop short. On the laboratory evidence, taking more of the advice would usually have been more accurate.

Why does advice discounting cost accuracy?

Because averaging usually beats choosing. Where two estimates fall on opposite sides of the truth, their average must beat the average of the two judges, so combining buys accuracy for nothing. People do not believe this and tend to pick one estimate instead. In one study the mean weight of advice was 0.27 where the best achievable weight on the same data was 0.44.

Do people prefer advice from algorithms or from people?

Both findings are real. Lay people relied more on identical advice when told it came from an algorithm, though they still underweighted it in either condition. People who watched a statistical model make a visible error then avoided it, even though the model beat the human forecasters. We reconcile the two ourselves, by way of Dietvorst's own control conditions: the baseline is favourable, and one visible error withdraws it.

Sources

Yaniv, I., & Kleinberger, E. (2000). 'Advice Taking in Decision Making: Egocentric Discounting and Reputation Formation.' Organizational Behavior and Human Decision Processes, 83(2), 260-281. doi:10.1006/obhd.2000.2909. [FULL TEXT READ, from the Hebrew University Center for the Study of Rationality discussion-paper copy, Discussion Paper 212, which reproduces the published article including journal, volume and page line.] The founding demonstration. Four studies, all on computers, all with a cash bonus for accuracy, all using general-knowledge questions on dates of historical events within the last 300 years. Study 1, 25 undergraduates: mean weight of own estimate 0.71, significantly above the 0.50 equal weighting implies, with 21 of 25 above 0.50. Across all 375 individual trials the median weight of own estimate was 0.98 and 53 per cent of trials fell in the 0.9 to 1.0 band, so the modal response to a piece of advice was to ignore it entirely, which must not be conflated with the average. The exact correct answer was given on only 4.4 per cent of first-phase trials. MECHANISM, in the authors' words: differential information, 'the notion that decision makers have privileged access to their internal reasons for holding their own opinion, but not to the advisors' internal reasons', with overconfidence, perseverance and confirmation bias explicitly set aside as less parsimonious given neutral questions, isolation at a terminal and payment for accuracy. Study 2, 80 respondents: weight of own estimate 0.74 for a poor adviser and 0.48 for a good one, the good adviser having been more accurate than 98 per cent of respondents and still earning only equal weight. Study 3, 104 respondents, quality switching mid-sequence: 0.82 with a poor adviser, 0.41 with a good one, with the asymmetry the paper is known for, a fall from good to average pushing weight of own estimate from 0.41 to 0.72 against a rise from poor to average moving it only from 0.82 to 0.74. Study 4, 92 respondents choosing whether to buy advice: bought from the good adviser about 70 per cent of the time and from the poor adviser about 20 per cent. Author order confirmed at the Crossref item record: Yaniv first, Kleinberger second, both Hebrew University of Jerusalem.

Yaniv, I. (2004). 'Receiving other people's advice: Influence and benefit.' Organizational Behavior and Human Decision Processes, 93(1), 1-13. [FULL TEXT READ, from the Hebrew University discussion-paper copy, Discussion Paper 405.] Sole author, and the cleanest accuracy numbers in the literature. Study 1, 30 respondents: mean weight of advice 0.27 against the 0.50 equal weighting implies; more knowledgeable judges discounted more, 0.20 against 0.33. Accuracy improved in both groups: mean absolute error fell from 46.3 to 38.9 years for the high-knowledge group, about a sixth, and from 66.0 to 50.7 for the low-knowledge group, about a fifth. THE FIGURE THE ARTICLE USES: the paper derives the best achievable weight on its own data and reports an average optimal empirical weight of advice of 0.44 against an actual 0.27. Study 2, 48 respondents: weight fell with distance from the initial opinion and overall error fell about a fifth. Study 3, 76 respondents, mechanically constructed rather than ecologically sampled advice: THE ACCURACY GAIN VANISHED ENTIRELY, error moving only from 65.6 to 63.4 years and not significant, which is the paper's own warning that the benefit depends on the advice being real.

Bailey, P. E., Leon, T., Ebner, N. C., Moustafa, A. A., & Weidemann, G. (2022). 'A meta-analysis of the weight of advice in decision-making.' Current Psychology, 42(28), 24516-24541. doi:10.1007/s12144-022-03573-2. Open access, CC-BY. [ABSTRACT, INTRODUCTION, METHOD SUMMARY AND POOLED-EFFECT SECTION READ IN FULL at the Springer open-access page.] The single most important source in this file and the number the article leads on. A three-level multi-level meta-analysis: a search across PsycINFO, PubMed, Web of Science and Scopus leaving 53 articles comprising 129 independent datasets with 17,296 participants, from which 346 effect sizes were extracted. THE OVERALL POOLED WEIGHT OF ADVICE WAS 0.39, 95 per cent confidence interval 0.37 to 0.42, meaning individuals on average adjusted their estimates to be 39 per cent closer to the advised estimate. Two outlier effects at 0.93 and 0.92 were identified and excluding them changed the pooled effect not at all. THE ONLY UNIQUE PREDICTOR was information about the adviser suggestive of advice quality: adjustment of 32 per cent, 37 per cent and 48 per cent for advisers described in ways suggesting low, neutral and high quality. The authors' own warning, used in substance: the benefits of compromise and averaging may be lost if accurate advice is perceived to be low quality, or if too much weight is given to inaccurate advice perceived as high quality. DEMOGRAPHIC AND CULTURAL SCOPE, reported in the article as a null rather than as a finding: sample characteristics had no effect, and the authors state they provide no evidence that age, gender or individualism, their operationalisation of culture, influences the weight of advice. The paper also states the field's baseline, that incorporating 50 per cent of an advised estimate would represent a rational use of advice. EXTERNAL VALIDITY NOTE: every one of the 53 articles is a judge-advisor paradigm study, and the paper's own framing concedes that a revisable numerical estimate 'represents one of the simplest forms of advice-taking'. Bailey first and corresponding, at the Graduate School of Health, University of Technology Sydney.

Larrick, R. P., & Soll, J. B. (2006). 'Intuitions about combining opinions: Misappreciation of the averaging principle.' Management Science, 52(1), 111-127. doi:10.1287/mnsc.1050.0459. [ABSTRACT AND SUMMARY LEVEL ONLY; full text not retrieved. No sample size or effect size may be quoted, and the draft says in prose that both averaging papers were read at abstract level only.] Larrick first, Soll second, both Fuqua School of Business, Duke University. The paper that explains why discounting costs anything. It defines bracketing, the case where two judges' estimates fall on opposite sides of the true value, and shows that under a convex loss function such as mean absolute deviation the average of two bracketing judges must beat the average judge's error. The finding is that people do not see this: they wrongly believe averaging yields no accuracy gain over the average judge, and the misconception held across experiments using summary data, specific instances and conceptual reasoning. An erratum at Management Science 52(2), 309-310 corrects a typesetting error and does not affect the finding.

Soll, J. B., & Larrick, R. P. (2009). 'Strategies for revising judgment: How (and how well) people use others' opinions.' Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(3), 780-805. doi:10.1037/a0015145. [ABSTRACT AND SUMMARY LEVEL ONLY; full text not retrieved. No sample size or effect size quoted.] Soll first, Larrick second, both Fuqua School of Business, Duke University. Identifies the two strategies people use when revising a numerical estimate after seeing another person's, choosing between the two estimates or averaging them, builds a probability, accuracy, redundancy model to compare them, and reports that averaging was the more effective strategy across a wide range of commonly encountered environments while people nonetheless lean towards choosing, most participants having been more accurate had they always averaged.

Sniezek, J. A., Schrah, G. E., & Dalal, R. S. (2004). 'Improving judgement with prepaid expert advice.' Journal of Behavioral Decision Making, 17(3), 173-190. doi:10.1002/bdm.468. [FULL DEPOSITED ABSTRACT READ, reconstructed from the OpenAlex inverted index, which carries it complete.] The primary source for the cost-of-advice moderator, and the earlier, conditional and independently reported one. A laboratory experiment varying adviser expertise and the opportunity for monetary rewards, the stated main contribution being the manipulation of the timing of monetary rewards. The abstract's own finding: 'We found, as predicted, that committing money for expert, but not novice, advice increases Judges' use of their subsequent advice and their estimation accuracy.' Sniezek first, Schrah second, Dalal last and corresponding. No sample size was retrieved and none is quoted.

Gino, F. (2008). 'Do we listen to advice just because we paid for it? The impact of advice cost on its use.' Organizational Behavior and Human Decision Processes, 107(2), 234-245. doi:10.1016/j.obhdp.2008.03.001. [ABSTRACT-LEVEL ONLY; full text not retrieved and no repository copy found. No sample size or effect size quoted, and the draft states in prose that it was read at abstract level only.] Sole author. Three judge-advisor experiments investigating whether the cost of advice, independent of its quality, affects how people use it, reporting that paid advice was weighted more heavily than free advice of identical quality, attributed to the forces documented in the literature on the sunk-cost fallacy. Carried in the article as consistent with Sniezek, Schrah and Dalal 2004 rather than as the load-bearing source. No independent replication of this specific result was found in this pass. The criticism is of an evidence base of one paper, not of the author.

Gino, F., & Moore, D. A. (2007). 'Effects of task difficulty on use of advice.' Journal of Behavioral Decision Making, 20(1), 21-35. doi:10.1002/bdm.539. [ABSTRACT AND SUMMARY LEVEL ONLY; full text not retrieved. No sample size or effect size quoted.] Gino first, Moore second, both at the Tepper School of Business, Carnegie Mellon University at the time. THE FINDING THAT BREAKS THE OLD TITLE: two studies found that people give more weight to advice on difficult tasks and less on easy ones, and that this held regardless of whether advice was automatically provided or had to be sought out, so solicitation by itself did not change the weighting. CITATION CAUTION: OpenAlex carries the publication year as 2006 against a volume 20 issue 1 dated 2007, so the article cites it without a year, by authors and finding. The verified DOI is 10.1002/bdm.539.

Milyavsky, M., & Gvili, Y. (2024). 'Advice taking vs. combining opinions: Framing social information as advice increases source's perceived helping intentions, trust, and influence.' Organizational Behavior and Human Decision Processes, 183, article 104328. doi:10.1016/j.obhdp.2024.104328. [FULL PUBLISHER ABSTRACT READ at the Ben-Gurion University research record, which carries it entire.] The most useful single finding in this file for a commercial reader. Six preregistered experiments, 3,411 participants in total. In the authors' own words, participants 'placed greater weight on and paid more for others' estimates when presented as advice than when presented as opinions', and 'this advice framing effect substantially reduced egocentric discounting and held across various types of judgments, and for both good and ecological advice.' The proposed mechanism, supported by mediational analyses and by experimental evidence, is higher helping intentions and thereby trustworthiness ascribed to the source of advice rather than of opinions. Milyavsky first and corresponding, Gvili second, both at the Faculty of Business Administration, Ono Academic College. NOTE: Yaniv Gvili is not Ilan Yaniv, the first author of the two founding papers. Different people, not to be merged.

Logg, J. M., Minson, J. A., & Moore, D. A. (2019). 'Algorithm appreciation: People prefer algorithmic to human judgment.' Organizational Behavior and Human Decision Processes, 151, 90-103. doi:10.1016/j.obhdp.2018.12.005. Open access at eScholarship, CC-BY-NC-ND. [FULL TEXT READ in substantial part from the eScholarship author-accepted copy, including the method and results of every experiment and the general discussion.] SIDE ONE OF THE DISPUTE. Six experiments, all using weight on advice as the dependent variable, all incentivised. Experiment 1A, 202 participants estimating a person's weight from a photograph: weight of advice 0.45 for algorithmic advice against 0.30 for human advice. Experiment 1B, 215 participants forecasting Billboard song popularity: 0.37 against 0.21. Experiment 1C, 286 participants forecasting romantic attraction: 0.38 against 0.26. Experiment 1D asked 119 active judgment-and-decision-making researchers to predict the result and they predicted algorithm aversion, the opposite of the finding. Experiment 2, 149 participants: the preference survived separate evaluation, 0.50 against 0.35, and 75 per cent preferred the algorithm under joint evaluation. Experiment 3 found the preference weakened when the human alternative was the participant's own estimate, though more than half still chose the algorithm in both conditions. THE HONEST QUALIFICATION, STATED IN THE PAPER: participants underweighted advice in both conditions, adjusting less than half-way towards it, so algorithm appreciation is a smaller shortfall rather than the absence of one. Logg first, Minson second, Moore third. EXPERIMENT 4 OF THE SAME PAPER, read separately. [FULL METHOD, RESULTS AND DISCUSSION SECTIONS READ.] THE EXPERT FINDING. 301 Mechanical Turk participants against 70 United States national-security professionals, 371 in all, on four tasks as repeated measures: a weight estimate, a Tesla delivery forecast, a cyber-sanctions forecast and a Brexit Article 50 forecast, with advice drawn from a follow-up tournament to the Good Judgment Project and of high quality. Experts took substantially less advice than lay people, and the paper's own sentence, quoted in the article in substance: 'the experts weighted advice less than the 30% WOA observed in advice-taking research.' THE COST: on Brier scores there was an effect of sample expertise and an interaction, and lay people achieved greater accuracy than experts on algorithmic advice because the experts discounted it. The discussion's own words, quoted in the article: 'algorithmic advice falls on deaf expert ears, with a cost to their accuracy.' The paper also relays a converging result that expert attorneys weight advice less than law students in a verdict-estimation task; that paper was not read and is not used. SCOPE THE AUTHORS STATE: the two samples likely differ in many respects beyond forecasting expertise, and the Mechanical Turk responses serve as a benchmark rather than a matched control.

Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). 'Algorithm aversion: People erroneously avoid algorithms after seeing them err.' Journal of Experimental Psychology: General, 144(1), 114-126. doi:10.1037/xge0000033. [FULL TEXT READ in substantial part from the Wharton-hosted copy of the accepted manuscript. LIMITATION: the condition-by-condition choice percentages are carried in Figure 3 and could not be read, so NO PERCENTAGE from Studies 1, 2 or 4 is quoted from this source.] SIDE TWO OF THE DISPUTE. Five studies in which participants chose whether to tie an incentive to a statistical model's forecasts or to human forecasts, after being assigned to see the model perform, the human perform, both, or neither. Studies 3a and 3b forecast states' ranks by departing airline passengers at final samples of 410 and 1,036, and in these the model produced about half as much error as the human forecasters. RESULTS: the model outperformed participants in all five studies and participants in every condition were better off choosing the model, yet in every experiment where experience was manipulated those who saw the model perform, and therefore err, were significantly less likely to tie their bonuses to it. THE ASYMMETRY THAT IS THE MECHANISM: seeing the model err decreased the tendency to choose the model, while seeing the human err did not significantly decrease the tendency to choose the human, even though the humans' mistakes were larger. Dietvorst first, Simmons second, Massey third.

Seiter, M., and colleagues (2021). 'Towards a better understanding on mitigating algorithm aversion in forecasting: an experimental study.' Journal of Management Control. doi:10.1007/s00187-021-00326-3. Open access at Springer. [ABSTRACT AND RESULTS SECTION 5.1 READ IN FULL. LIMITATION: the page metadata exposed only Mischa Seiter of Ulm University as an author and the volume, issue and page numbers were not retrieved, so the article cites it as 'Seiter and colleagues' with the journal and DOI only.] THE REPLICATION STATUS OF DIETVORST. An experimental study with 1,840 participants overall, drawing on Dietvorst, Simmons and Massey 2015 and adding time pressure, 'do your best' goals and forecasters' data input decision rights. THE REPLICATION HELD AND THE DIRECTION WAS CONFIRMED: 67 per cent chose the algorithm in the appreciation replication condition and 47 per cent after performance feedback. THE COMPARISON WITH THE ORIGINAL, AND THE ONLY LICENCE FOR THE 2015 PERCENTAGES: the paper reports the share choosing the algorithm without feedback as 54 per cent in Dietvorst and colleagues against 67 per cent here, and with feedback 42 against 47, so the delta representing algorithm aversion has grown from 12 to 19 percentage points. Their own substantive finding was support for time pressure mitigating algorithm aversion and no support for 'do your best' goals or data input decision rights. CITING RULE OBSERVED IN THE DRAFT: the 54 and 42 per cent figures are attributed to this paper's comparison.

Lohre, E., & Halkjelsvik, T. (2024). 'Advice taking when the stakes are high: Evidence from a game show.' Judgment and Decision Making, 19, e25. doi:10.1017/jdm.2024.4. Open access, CC-BY. [FULL ABSTRACT AND PARTICIPANTS, OPEN SCIENCE AND PROCEDURE SECTIONS READ at the Cambridge Core open-access page.] THE ONLY GENUINE FIELD STUDY OF ADVICE TAKING FOUND IN THIS PASS, AND IT REVERSES THE EFFECT. Five seasons, 2018 to 2022, of the Norwegian television game show 'Alle mot 1'. 49 contestants, 26 women and 23 men aged 19 to 59, making 382 judgments and soliciting advice on 147 of them, with stakes rising through each episode to a theoretical maximum of 100,000 Norwegian kroner, about 10,000 United States dollars at the time. The design is a judge-advisor system in everything but the setting, with contestants, viewers and advisers all answering independently inside a thirty-second limit. THE HEADLINE: the average weight of advice was 0.58, so contestants weighted advice more than their own estimates. Solicitation was associated with higher stakes; own estimation success was associated with less weight on advice; no association was detected with the source of the advice. TWO QUALIFICATIONS CARRIED IN THE ARTICLE: estimates did not improve much after advice taking and the potential for improvement by averaging was negligible in this setting; and the sample is small and self-selected, screened by the production company, which the authors say probably means higher extraversion and openness than the Norwegian population. THE SENTENCE QUOTED IN THE DRAFT AS A BLOCK QUOTE: 'The results suggest that highly controlled lab studies may not give an accurate representation of advice taking in high stakes and highly public settings.' AUTHOR ORDER CAUTION: the Cambridge page metadata lists only Torleif Halkjelsvik and attaches an ORCID belonging to the other author. The OpenAlex record is authoritative: Erik Lohre of BI Norwegian Business School is first author and corresponding; Halkjelsvik of Simula Metropolitan Center for Digital Engineering is second. Credit Lohre.

Logg, J. M., & Schlund, R. (2024). 'A Simple Explanation Reconciles "Algorithm Aversion" and "Algorithm Appreciation": Hypotheticals vs. Real Judgments.' SSRN working paper, abstract id 4687557. [NOT READ beyond a summary of its abstract; no full text retrieved and author order and date not verified at a bibliographic record.] The one attempt at reconciling the two sides that this pass located, named in the draft as an unrefereed working paper so that the article does not appear to invent a synthesis that already exists. The argument reported is that the contradiction arises from how preferences are elicited, with real judgments producing consistent use of algorithmic advice while hypothetical questions produce unstable preferences. The article's own reconciliation rests on the primary designs rather than on this paper and is presented as our reading.

ABSENCE CLAIM, LICENSED BY THE FACT FILE. Four separate searches were run and recorded: field experiments on advice discounting with real consumer purchases; salesperson and retail field studies measuring weight of advice; online reviews and ratings treated as advice against transaction data; and financial-adviser field studies. No study measuring advice discounting against a real purchase was found. The smallest consumer-side studies located, treating online reviews as probability signals in a discounting design, were at 29 and 25 participants and are too small to carry an effect size, so they are not used. The article states the absence with that provenance and in its own voice.

NOT USED, AND RECORDED SO THE OMISSION IS DELIBERATE. Harvey, N., & Fischer, I. (1997), Organizational Behavior and Human Decision Processes 70(2), 117-133, doi:10.1006/obhd.1997.2697, named the task and supplied the weight-of-advice formula, but was NOT READ, carries no deposited abstract and was found in no repository, so no figure from it appears. Sniezek, J. A., & Van Swol, L. M. (2001), Organizational Behavior and Human Decision Processes 84(2), 288-307, is the standing reference on trust, confidence and expertise in the judge-advisor system and was NOT READ, so the confidence heuristic is left out of the article entirely rather than attributed. The unnamed meta-analysis of interventions against algorithm aversion, located as a conference paper in the Hawaii International Conference on System Sciences repository with an effect of 0.23, has unestablished authors, year and peer-review status and is left out.

Related concepts

• Algorithm Appreciation
• Algorithm Aversion

People

• Jennifer M. Logg
• Kim Scott
• Berkeley J. Dietvorst