Pith. sign in

REVIEW 3 major objections 5 minor 262 references

When AI gives qualitative advice from a black box, people barely move extreme confirming priors, move a lot when advice contradicts them, and move less at intermediate priors—better matched by contraction and inertial rules than by quasi-Ba

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 11:32 UTC pith:EB5TIWJK

load-bearing objection Large, clean experiment on belief updating under qualitative AI advice; the descriptive patterns are solid and the model ranking is useful if you accept the maintained Bayesian benchmark. the 3 major comments →

arxiv 2607.10460 v1 pith:EB5TIWJK submitted 2026-07-11 econ.GN q-fin.EC

Learning from an Unknown DGP: Experimental Evidence on Belief Updating with AI Recommendations

Matthew Kovach , Daniel Martin , Gerelt Tserenjigmid This is my paper
classification econ.GN q-fin.EC
keywords belief updatingunknown data-generating processAI recommendationsqualitative informationinertial updatingcontraction rulequasi-Bayesianexperimental economics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks how people revise beliefs when they get qualitative AI recommendations—“more likely over 21” or “more likely under 21”—from a process they do not know. In a large online experiment using face images and incentivized prior and posterior reports, the authors document three regular patterns: almost no updating when advice confirms an extreme prior, large updating when advice contradicts an extreme prior, and smaller updating for middling priors. Those patterns motivate four properties—consistency with the recommendation, monotonicity in the prior, reactionary updating, and threshold non-updating—which hold in the aggregate and for most individuals on the weak properties. Providing overall accuracy and state-contingent rates does not change the shape of updating. Reduced-form models that treat the recommendation as a qualitative information set (contraction and weighted inertial updating) fit held-out data better than a quasi-Bayesian rule that warps a likelihood ratio under signal-dependence neglect. The practical stake is simple: black-box qualitative AI advice may not be integrated the way classical signal models assume, so interface design and welfare calculations need different updating rules.

Core claim

Across 60,252 prior–posterior pairs, belief updating after qualitative AI recommendations from an unknown DGP shows three patterns: near-zero updates when recommendations confirm extreme priors, larger updates when they contradict extreme priors, and smaller updates for intermediate priors. These patterns support consistency, monotonicity, reactionary updating, and aggregate threshold updating; contraction and weighted (subjective) inertial updating capture the data better out of sample than quasi-Bayesian updating.

What carries the argument

Four testable properties of belief updating (consistency, monotonicity, reactionary updating, threshold updating) estimated with observation-level regressions and hinge regressions for non-updating thresholds, then used to discipline comparisons among quasi-Bayesian, Contraction Rule, and weighted Inertial Updating (including a subjective-threshold version).

Load-bearing premise

The paper treats a quasi-Bayesian model that neglects dependence between the image and the AI recommendation (or variants that still force extreme priors to stay extreme) as the right Bayesian foil, so poorer fit is read as evidence against likelihood-ratio updating rather than against a misspecified conditional advice process people might hold after seeing the face.

What would settle it

In the same binary qualitative-advice design, if out-of-sample posterior MSE favored a properly conditioned Bayesian or quasi-Bayesian model over contraction and inertial rules, or if hinge non-updating thresholds collapsed to the degenerate boundaries once people got local feedback or fully conditional reliability, the central claim would fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Global accuracy rates alone need not reshape how people use qualitative AI recommendations if average beliefs about accuracy are already roughly right.
  • Models of human–AI collaboration and welfare from advice should allow inertial or contraction rules rather than only distorted Bayes with known likelihood ratios.
  • Interface formats—qualitative labels, calibrated probabilities, confidence, local explanations—can change which updating rule people apply.
  • Shared qualitative AI advice need not produce Bayesian consensus across people with different priors, because the same message is confirming for some and contradicting for others.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same asymmetric pattern appears for expert human advisors whose recommendations are held fixed, source label (human vs AI) may matter less than qualitative format and opacity of the process.
  • Training that teaches local dependence between own judgment and the model, rather than only average accuracy, is a natural next design to test whether people can be moved toward likelihood-ratio integration.
  • The high share of exact non-updating is not pure inattention if response times remain positive on confirming rounds; it is a candidate feature of the decision rule itself.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper reports a large incentivized experiment (377 U.S. Prolific participants, 160 rounds of the Bouncer age-classification task, 60,252 prior–posterior pairs) on belief updating after qualitative AI recommendations when the DGP is unknown. Participants report a prior that a face is over 21, receive an Over/Under AI recommendation, then report a posterior; treatments vary whether state-contingent AI accuracy is disclosed (INFO vs NOINFO). The authors document three aggregate patterns—near-zero updates when advice confirms extreme priors, larger updates when it contradicts extreme priors, and smaller updates for intermediate priors—and formalize four testable properties (consistency, monotonicity, reactionary updating, threshold updating), which hold in aggregate and across many splits (Tables 1–3). They then compare quasi-Bayesian (Grether-style) updating under signal-dependence neglect to the Contraction Rule and weighted Inertial Updating (objective and subjective), finding that CR and wIU/wsIU have lower out-of-sample MSE and higher completeness than qB, with wIU most restrictive (Tables 4–7, Figure 7).

Significance. If the descriptive patterns and model ranking hold, the paper fills a genuine gap: most experimental belief-updating work uses transparent DGPs, while many real AI interfaces deliver qualitative recommendations from opaque processes. The design strengths are real—equal incentive weight on prior and posterior, binarized scoring, preregistration, two-way clustering by participant and image, many robustness splits, out-of-sample MSE, and completeness/restrictiveness comparisons in the spirit of Fudenberg et al. The contribution is primarily empirical and comparative: it shows that likelihood-ratio-style qB under the paper’s maintained benchmarks struggles with reactionary and threshold updating, while CR and especially wIU/wsIU capture the shape of updating more directly. That is useful for human–AI interaction and for non-Bayesian updating theory, even if microfoundations remain open.

major comments (3)
  1. Sections 4.1 and 6.1 and Table 9: the central claim that CR/wIU outperform “Bayesian/quasi-Bayesian” updating rests on signal-dependence neglect (or prior-conditioned variants that still force posteriors at 0/1 to stay at 0/1). The paper correctly notes that unrestricted Pr(R|s,X) makes interior pairs essentially untestable while preserving endpoint implications that conflict with large moves from extreme contradicting priors. That is a fair defense of the ranking, but the abstract and introduction still read as if poorer qB fit is evidence against likelihood-ratio updating in general. The manuscript should state more sharply, up front, that the comparison is to qB under these maintained benchmarks—not to every Bayesian model with private visual information—and that the decisive empirical fact against endpoint-preserving Bayes is the large updating from extreme contradicting priors.
  2. Section 3.2–3.5 and Appendix A: the four properties and hinge non-updating thresholds are the paper’s main descriptive contribution, yet Appendix A acknowledges they were not preregistered; they are motivated by the observed patterns and then tested. Aggregate support is strong (Table 1; Table 3 across splits), but individual threshold updating holds for only 46.7% of participants, and the sample (188/189 per treatment) is below the preregistered 200–300 per arm. The central claim should be framed as a carefully documented exploratory characterization of updating shape, with preregistered model comparison as the confirmatory piece, rather than as four fully confirmatory properties at both aggregate and individual levels.
  3. Section 3.4 and the model discussion of non-updating: 78.6% of observations have identical prior and posterior, and 8.5% of participants never update. The paper argues this is concentrated where priors already satisfy the recommendation and is consistent with wIU’s predicted non-updating range given accuracy levels. That is plausible, but the high non-update rate is also consistent with slider cost, fatigue, or weak incentives to revise. Appendix B.8’s response-time evidence helps, yet a load-bearing robustness check would be to show that the three patterns and the CR/wIU ranking survive when restricting to rounds with material movement or excluding high no-update participants more aggressively than Table 10 already does—and to report how much of the MSE advantage of wIU/wsIU is mechanical from predicting exact zeros.
minor comments (5)
  1. Figure 1 and related figures: vertical bars are ±1 SE of the posterior mean; with two-way clustering elsewhere, a brief note on whether bin-level SEs are clustered would help readers reconcile the visual and regression evidence.
  2. Equation (2) and footnote on endpoint mapping: the 0.005/0.995 winsorization for Grether estimation is standard but material for extreme priors; report sensitivity to alternative floors/ceilings in the appendix.
  3. Table 4 vs Table 5: aggregate vs individual MSE rankings differ (qB wins a non-trivial individual share). A short paragraph reconciling population completeness with type heterogeneity would improve clarity.
  4. Appendix A is admirably transparent about deviations from AsPredicted #228821; consider moving a one-paragraph version of that disclosure into the main text near the property tests.
  5. Minor polish: “quasi-Bayesian (qB)” is introduced cleanly, but “automation neglect” vs “underreaction” language could be aligned more consistently with Agarwal et al. when comparing β1 and β2.

Circularity Check

1 steps flagged

Empirical paper with no load-bearing circular derivation; only mild same-sample formalization of observed patterns into property tests.

specific steps
  1. other [Abstract; §1 intro of four properties; §3.2 Table 1; Appendix A]
    "These three behavioral patterns suggest four testable properties of belief updating, which we assess at the aggregate and individual levels. ... Several analyses in the current paper were not specified in the preregistration. These include the four behavioral property tests, the non-updating threshold estimates... These additions are motivated by the observed updating patterns"

    The four properties are not independent theoretical predictions tested on fresh data; they are formal restatements of the three patterns already seen in the same 60,252 pairs, then re-tested on those pairs. This is mild exploratory circularity (data → properties → same-data tests), not a by-construction algebraic reduction of a fitted parameter renamed as prediction.

full rationale

The paper’s central claims are descriptive experimental patterns (near-zero confirming updates at extremes, larger contradicting updates at extremes, smaller intermediate updates) and comparative model fit (qB vs CR vs wIU/wsIU) under held-out MSE, completeness, and restrictiveness. Model parameters are estimated and then evaluated out of sample (100 train/test splits; participant-level splits), not treated as independent first-principles predictions. Self-citations to Dominiak–Kovach–Tserenjigmid (wIU) and related theory papers propose candidate functional forms that are then ranked against alternatives on new data; that is ordinary model competition, not a uniqueness chain that forces the empirical ranking. The only mild circularity is that the four “testable properties” are explicitly suggested by the three patterns in the same dataset and then assessed on that dataset (and were not preregistered). That is exploratory formalization, not an algebraic identity of inputs and outputs. No equation reduces a claimed prediction to a fitted input by construction, and relaxing the Bayesian benchmark is discussed openly rather than smuggled in. Score 1 reflects that minor same-sample theory-from-data step only.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The load-bearing content is experimental measurement plus reduced-form model comparison. Background math is standard probability and regression. Domain assumptions include binary age state, qualitative recommendation as an information set, and signal-dependence neglect as the baseline Bayesian/qB benchmark. Free parameters are the fitted updating coefficients. Invented/operational entities are the four behavioral properties and hinge non-updating thresholds used to organize the data.

free parameters (5)
  • qB parameters (β1, β2, α) = β1=0.51, β2=0.78, α=0.14
    Fitted to prior–posterior pairs under Grether-style quasi-Bayes; baseline estimates β1≈0.51, β2≈0.78, α≈0.14 on full sample.
  • CR parameters (ϵ_R, ρ_R) by recommendation = ϵ_Over=0.70, ϵ_Under=0.77, ρ_Over=0.85, ρ_Under=0.14
    Four free parameters for contraction weights and representative beliefs; aggregate estimates ϵ_Over=0.70, ϵ_Under=0.77, ρ_Over=0.85, ρ_Under=0.14.
  • wIU prior weight w = w=0.44
    Single free parameter in objective weighted inertial updating; aggregate estimate w=0.44.
  • wsIU parameters (w, ϕ_Over, ϕ_Under) = ϕ_Over=0.68, ϕ_Under=0.41, w=0.59
    Subjective thresholds and prior weight fitted per aggregate/individual; aggregate ϕ_Over=0.68, ϕ_Under=0.41, w=0.59.
  • Hinge non-updating thresholds = Over≈0.620; Under≈0.310
    Estimated locations where average updating becomes zero after each recommendation; used as behavioral parameters for threshold updating.
axioms (5)
  • domain assumption Binary payoff-relevant state (Over 21 vs Under 21) and qualitative recommendations as sets of consistent beliefs.
    Section 4 reduces the image task to a binary state and maps Over/Under advice to information sets I(R).
  • domain assumption Baseline Bayesian/qB comparison uses signal-dependence neglect: Pr(R|s,X)≈Pr(R|s).
    Section 4.1 states INFO provides Pr(R|s) but not Pr(R|s,X); baseline qB follows Agarwal et al.-style neglect.
  • domain assumption Truthful reporting under binarized scoring rule and equal selection of prior/posterior for payment.
    Section 2.1 uses Hossain–Okui BSR and random selection to incentivize beliefs.
  • standard math Standard regression/MSE/completeness-restrictiveness machinery for property tests and model comparison.
    Sections 3 and 5 use observation-level regressions, out-of-sample MSE, and Fudenberg et al. measures.
  • ad hoc to paper Objective wIU information sets are half-spaces [1/2,1] and [0,1/2] for Over/Under recommendations.
    Section 4.3 chooses these sets as the natural qualitative content of “more likely” advice; subjective ϕ relaxes them.
invented entities (2)
  • Four testable updating properties (consistency, monotonicity, reactionary updating, threshold updating) no independent evidence
    purpose: Organize the three observed patterns into falsifiable aggregate and individual tests.
    Introduced from the data patterns in Sections 1 and 3.2; not standard textbook axioms of Bayesian updating.
  • Non-updating threshold (hinge location by recommendation) no independent evidence
    purpose: Summarize where confirming priors stop moving on average.
    Estimated via hinge regressions; used as a behavioral parameter distinguishing high/low priors.

pith-pipeline@v1.1.0-grok45 · 38130 in / 3700 out tokens · 49935 ms · 2026-07-14T11:32:26.251394+00:00 · methodology

0 comments
read the original abstract

We use a controlled experiment to study how beliefs are updated after receiving qualitative information (AI recommendations) from an unknown data-generating process (DGP). Across 60,252 pairs of prior and posterior beliefs, we document three behavioral patterns: updates close to zero when recommendations confirm extreme priors, larger updates when recommendations contradict extreme priors, and smaller updates for intermediate priors. These three behavioral patterns suggest four testable properties of belief updating, which we assess at the aggregate and individual levels. Finally, we examine how well updates are captured by three models of belief updating.

Figures

Figures reproduced from arXiv: 2607.10460 by Daniel Martin, Gerelt Tserenjigmid, Matthew Kovach.

Figure 1
Figure 1. Figure 1: Aggregate average beliefs and estimated predictions for all models by AI recom [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Prior beliefs and binary accuracy in the aggregate data (pooled across participants, [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Updating by treatment. of prior beliefs is above the median image value of 0.267. The round split separates rounds 81–160 from rounds 1–80. The response time split is at the participant level and classifies participants as “fast” if their median response time for prior beliefs is below the sample median of 3.5 seconds. The remaining splits at the participant level classify participants as women, older than… view at source ↗
Figure 4
Figure 4. Figure 4: Predictions of qB assuming δ(R1|s1) = 0.825 and δ(R2|s2) = 0.700, as in our experiment. in the recommendation, distort the prior, and have a reduced-form bias toward one state: µR(s1) = δ(R|s1) β1 µ(s1) β2 e α δ(R|s1) β1 µ(s1) β2 e α + δ(R|s2) β1 µ(s2) β2 (2) , where β1 ≥ 0 controls how strongly the recommendation process enters the posterior, β2 ≥ 0 controls how strongly the prior enters the posterior, an… view at source ↗
Figure 5
Figure 5. Figure 5: Predictions of CR. The parameter ϵR is the weight on the DM’s prior. The corresponding recommendation or AI weight is 1 − ϵR. Because the parameters ρR and ϵR are allowed to vary across recommendations, in our setting CR has four free parameters, with two parameters (ρRi , ϵRi ) for each recommendation Ri [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Predictions of wIU. are disconfirming, the DM takes the minimum belief consistent with the recommendation (here 0.5) and mixes it with their prior. Thus, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Completeness–restrictiveness frontier. Dashed lines connect the nondominated [PITH_FULL_IMAGE:figures/full_fig_p031_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Average posterior at binned priors for different bins of AI advice for radiologist [PITH_FULL_IMAGE:figures/full_fig_p036_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Updating on advice labeled as human or AI from [PITH_FULL_IMAGE:figures/full_fig_p037_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ex post updating value under signal-dependence neglect, with bars showing one [PITH_FULL_IMAGE:figures/full_fig_p038_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Updating by binary directional correctness of the AI recommendation. [PITH_FULL_IMAGE:figures/full_fig_p051_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Updating by AI confidence [PITH_FULL_IMAGE:figures/full_fig_p051_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Updating by image difficulty. 51 [PITH_FULL_IMAGE:figures/full_fig_p051_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Updating by round (rounds 1–80 or rounds 81–160). [PITH_FULL_IMAGE:figures/full_fig_p052_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Updating by response time (above or below the median prior response time by [PITH_FULL_IMAGE:figures/full_fig_p052_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Updating by gender (self-reported as a woman or another category). [PITH_FULL_IMAGE:figures/full_fig_p052_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Updating by age (above or below the median age by participant). [PITH_FULL_IMAGE:figures/full_fig_p053_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Updating by participant ability (lower or higher MSE of prior beliefs by partici [PITH_FULL_IMAGE:figures/full_fig_p053_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Updating by belief about AI accuracy (above or below median belief about how [PITH_FULL_IMAGE:figures/full_fig_p054_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Updating by confidence in own prior (above or below median self-reported confi [PITH_FULL_IMAGE:figures/full_fig_p054_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Updating for participants with higher beliefs about AI accuracy and lower confi [PITH_FULL_IMAGE:figures/full_fig_p055_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Distribution of beliefs about AI accuracy. [PITH_FULL_IMAGE:figures/full_fig_p055_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Distribution of confidence in own prior by gender. [PITH_FULL_IMAGE:figures/full_fig_p056_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Update time by prior belief µ(s1) [PITH_FULL_IMAGE:figures/full_fig_p057_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Distribution of estimated AI weight (1 [PITH_FULL_IMAGE:figures/full_fig_p058_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Instruction screen. 59 [PITH_FULL_IMAGE:figures/full_fig_p059_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: AI accuracy information screen (text in red box for INFO treatment). [PITH_FULL_IMAGE:figures/full_fig_p060_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Payment instruction screen. 61 [PITH_FULL_IMAGE:figures/full_fig_p061_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Helpful tips screen. 62 [PITH_FULL_IMAGE:figures/full_fig_p062_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Prior belief elicitation screen. 63 [PITH_FULL_IMAGE:figures/full_fig_p063_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Belief elicitation screen with pop-up. 64 [PITH_FULL_IMAGE:figures/full_fig_p064_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Zoomed belief elicitation pop-up screen (text in pink box for INFO treatment). [PITH_FULL_IMAGE:figures/full_fig_p065_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Posterior belief elicitation screen. 66 [PITH_FULL_IMAGE:figures/full_fig_p066_33.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

262 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Organizational Behavior and Human Decision Processes , volume =

    The detrimental effects of power on confidence, advice taking, and accuracy , author =. Organizational Behavior and Human Decision Processes , volume =. 2011 , doi =

    2011

  2. [2]

    Theoretical Economics , volume=

    Pseudo-Bayesian updating , author=. Theoretical Economics , volume=. 2022 , publisher=

    2022

  3. [3]

    arXiv preprint arXiv:2109.07007 , year=

    The Empirical Content of Bayesianism , author=. arXiv preprint arXiv:2109.07007 , year=

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2603.02076 , year=

    When an AI Judges Your Work: The Hidden Costs of Algorithmic Assessment , author=. arXiv preprint arXiv:2603.02076 , year=

    arXiv

  5. [5]

    Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society , series =

    Vodrahalli, Kailas and Daneshjou, Roxana and Gerstenberg, Tobias and Zou, James , title =. Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society , series =. 2022 , publisher =. doi:10.1145/3514094.3534150 , url =

    doi:10.1145/3514094.3534150 2022

  6. [6]

    Nature Human Behaviour , year =

    Vaccaro, Michelle and Almaatouq, Abdullah and Malone, Thomas , title =. Nature Human Behaviour , year =

  7. [7]

    and Smith, Angela M

    Holt, Charles A. and Smith, Angela M. , title =. Journal of Economic Behavior & Organization , volume =. 2009 , doi =

    2009

  8. [8]

    arXiv preprint arXiv:2511.18582 , year=

    Barriers to AI adoption: Image concerns at work , author=. arXiv preprint arXiv:2511.18582 , year=

    arXiv

  9. [9]

    PLoS computational biology , volume=

    Hierarchical models in the brain , author=. PLoS computational biology , volume=. 2008 , publisher=

    2008

  10. [10]

    Econometrica , volume=

    A practical guide to updating beliefs from contradictory evidence , author=. Econometrica , volume=. 2021 , publisher=

    2021

  11. [11]

    Econometrica , volume =

    Bruhin, Adrian and Fehr-Duda, Helga and Epper, Thomas , title =. Econometrica , volume =. 2010 , doi =

    2010

  12. [12]

    Review of Economic Studies , volume=

    Dynamic opinion aggregation: long-run stability and disagreement , author=. Review of Economic Studies , volume=. 2024 , publisher=

    2024

  13. [13]

    Journal of consumer Research , volume=

    On the external validity of experiments in consumer research , author=. Journal of consumer Research , volume=. 1982 , publisher=

    1982

  14. [14]

    , author=

    Effects of cue consistency and value on base-rate utilization. , author=. Journal of Personality and Social Psychology , volume=. 1989 , publisher=

    1989

  15. [15]

    The Quarterly Journal of Economics , volume=

    Cognitive uncertainty , author=. The Quarterly Journal of Economics , volume=. 2023 , publisher=

    2023

  16. [16]

    2023 , institution=

    Contingent belief updating , author=. 2023 , institution=

    2023

  17. [17]

    American Economic Review , volume=

    Modeling the change of paradigm: Non-Bayesian reactions to unexpected news , author=. American Economic Review , volume=. 2012 , publisher=

    2012

  18. [18]

    2025 , note =

    Over- and Underreaction to Information: Belief Updating with Cognitive Constraints , author=. 2025 , note =

    2025

  19. [19]

    Available at SSRN 6482578 , year=

    Communicating with Data-Generating Processes: An Experimental Analysis , author=. Available at SSRN 6482578 , year=

  20. [20]

    Risk, Decision and Policy , volume=

    An experimental study of updating ambiguous beliefs , author=. Risk, Decision and Policy , volume=. 2000 , publisher=

    2000

  21. [21]

    Organizational behavior and human decision processes , volume=

    Cueing and cognitive conflict in judge-advisor decision making , author=. Organizational behavior and human decision processes , volume=. 1995 , publisher=

    1995

  22. [22]

    The Quarterly Journal of Economics , volume=

    Overreaction in expectations: Evidence and theory , author=. The Quarterly Journal of Economics , volume=. 2023 , publisher=

    2023

  23. [23]

    , title =

    Danz, David and Vesterlund, Lise and Wilson, Alistair J. , title =. American Economic Review , volume =. 2022 , doi =

    2022

  24. [24]

    2020 , institution=

    Non est disputandum de generalizability? A glimpse into the external validity trial , author=. 2020 , institution=

    2020

  25. [25]

    Experimental Economics , volume =

    Coutts, Alexander , title =. Experimental Economics , volume =. 2019 , doi =

    2019

  26. [26]

    Intention-Based Reciprocity and Signaling of Intentions , journal =

    Toussaert, S. Intention-Based Reciprocity and Signaling of Intentions , journal =. 2017 , doi =

    2017

  27. [27]

    and Schneider, Martin , title =

    Epstein, Larry G. and Schneider, Martin , title =. The Review of Economic Studies , volume =. 2007 , doi =

    2007

  28. [28]

    Journal of Mathematical Economics , volume =

    Cheng, Xiaoyu , title =. Journal of Mathematical Economics , volume =. 2022 , doi =

    2022

  29. [29]

    Management Science , volume =

    Liang, Yucheng , title =. Management Science , volume =. 2025 , doi =

    2025

  30. [30]

    and Harms, Philipp and Jackson, Matthew O

    Fryer, Roland G., Jr. and Harms, Philipp and Jackson, Matthew O. , title =. Journal of the European Economic Association , volume =

  31. [31]

    and Huang, R

    Moehring, Anne and Kutwal, M. and Huang, R. and Banerjee, O. and Jacobi, A. and Eber, C. and Mendoza, D. and Chung, M. and Dayan, E. and Gupta, Y. and Bui, T. D. T. and Truong, S. Q. H. and Pareek, A. and Langlotz, C. P. and Lungren, M. P. and Agarwal, N. and Rajpurkar, P. and Salz, T. , title =. Scientific Data , volume =. 2025 , doi =

    2025

  32. [32]

    http://www.nber.org/papers/w33949

    Agarwal, Nikhil and Moehring, Alex and Wolitzky, Alexander. Designing Human-AI Collaboration: A Sufficient-Statistic Approach. 2025. doi:10.3386/w33949 , URL = "http://www.nber.org/papers/w33949", abstract =

    work page doi:10.3386/w33949 2025

  33. [33]

    2023 , howpublished =

    Moehring, Anne and others , title =. 2023 , howpublished =. doi:10.17605/OSF.IO/Z7APQ , url =

    work page doi:10.17605/osf.io/z7apq 2023

  34. [34]

    , title =

    Bick, Alexander and Blandin, Adam and Deming, David J. , title =. Management Science , year =. doi:10.1287/mnsc.2025.02523 , note =

    doi:10.1287/mnsc.2025.02523 2025

  35. [35]

    and Dinlersoz, Emin and Foster, Lucia S

    Bonney, Kathryn and Breaux, Cory L. and Dinlersoz, Emin and Foster, Lucia S. and Haltiwanger, John C. and Pande, Aditya A. , title =. 2026 , doi =

    2026

  36. [36]

    On Prior Confidence and Belief Updating

    Chan, Kenneth and Charness, Gary and Dave, Chetan and Reddinger, J. Lucas , title =. arXiv preprint arXiv:2412.10662 , year =. doi:10.48550/arXiv.2412.10662 , note =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.10662

  37. [37]

    Games and Economic Behavior , volume =

    Coutts, Alexander , title =. Games and Economic Behavior , volume =

  38. [38]

    Learning under ambiguity: An experiment in gradual information processing , journal =

    Ngangou. Learning under ambiguity: An experiment in gradual information processing , journal =

  39. [39]

    American Economic Review , volume =

    Chen, Jaden Yang , title =. American Economic Review , volume =. 2026 , doi =

    2026

  40. [40]

    Journal of Economic Theory , volume =

    De Filippis, Roberta and Guarino, Antonio and Jehiel, Philippe and Kitagawa, Toru , title =. Journal of Economic Theory , volume =. 2022 , doi =

    2022

  41. [41]

    Journal of the European Economic Association , volume =

    Epstein, Larry G and Halevy, Yoram , title =. Journal of the European Economic Association , volume =

  42. [42]

    Journal of Economic Theory , volume =

    Shishkin, Denis and Ortoleva, Pietro , title =. Journal of Economic Theory , volume =. 2023 , doi =

    2023

  43. [43]

    2025 , note =

    Aina, Chiara , title =. 2025 , note =

    2025

  44. [44]

    , title =

    Aina, Chiara and Schneider, Florian H. , title =. 2025 , doi =

    2025

  45. [45]

    Journal of the European Economic Association , volume =

    Charness, Gary and Oprea, Ryan and Yuksel, Sevgi , title =. Journal of the European Economic Association , volume =. 2021 , doi =

    2021

  46. [46]

    American Economic Review , volume =

    Esponda, Ignacio and Vespa, Emanuel and Yuksel, Sevgi , title =. American Economic Review , volume =. 2024 , doi =

    2024

  47. [47]

    The Quarterly Journal of Economics , volume =

    Augenblick, Ned and Lazarus, Eben and Thaler, Michael , title =. The Quarterly Journal of Economics , volume =

  48. [48]

    Available at SSRN 5286198 , year =

    Hoong, Ruru and Dreyfuss, Bnaya , title =. Available at SSRN 5286198 , year =. doi:10.2139/ssrn.5286198 , note =

    work page doi:10.2139/ssrn.5286198

  49. [49]

    The Review of Economics and Statistics , volume =

    Fudenberg, Drew and Gao, Wayne and Liang, Annie , title =. The Review of Economics and Statistics , volume =. 2026 , doi =

    2026

  50. [50]

    Annual Review of Economics , volume =

    Ortoleva, Pietro , title =. Annual Review of Economics , volume =. 2024 , doi =

    2024

  51. [51]

    Journal of the European Economic Association , volume =

    Benjamin, Daniel J and Rabin, Matthew and Raymond, Collin , title =. Journal of the European Economic Association , volume =

  52. [52]

    The Review of Economic Studies , volume =

    Rabin, Matthew and Vayanos, Dimitri , title =. The Review of Economic Studies , volume =

  53. [53]

    The Quarterly Journal of Economics , volume =

    Rabin, Matthew , title =. The Quarterly Journal of Economics , volume =

  54. [54]

    American Economic Review , volume =

    Shmaya, Eran and Yariv, Leeat , title =. American Economic Review , volume =

  55. [55]

    Game-Theoretic Analyses of Trading Processes , booktitle =

    Wilson, Robert , editor =. Game-Theoretic Analyses of Trading Processes , booktitle =. 1987 , publisher =

    1987

  56. [56]

    2012 , note =

    Compte, Olivier and Postlewaite, Andrew , title =. 2012 , note =

    2012

  57. [57]

    Econometrica , pages =

    Bergemann, Dirk and Morris, Stephen , title =. Econometrica , pages =

  58. [58]

    arXiv preprint arXiv:2502.00958 , year =

    Dominiak, Adam and Kovach, Matthew and Tserenjigmid, Gerelt , title =. arXiv preprint arXiv:2502.00958 , year =

    Pith/arXiv arXiv

  59. [59]

    Trust and reliance on AI—An experimental study on the extent and costs of overreliance on AI , journal =

    Klingbeil, Artur and Gr. Trust and reliance on AI—An experimental study on the extent and costs of overreliance on AI , journal =

  60. [60]

    2024 , note =

    Harris, Adam and Yellen, Maggie , title =. 2024 , note =

    2024

  61. [61]

    Theoretical Economics , volume =

    Kovach, Matthew , title =. Theoretical Economics , volume =

  62. [62]

    Journal of Economic Theory , volume =

    Ke, Shaowei and Wu, Brian and Zhao, Chen , title =. Journal of Economic Theory , volume =

  63. [63]

    arXiv preprint arXiv:2303.06336 , year =

    Dominiak, Adam and Kovach, Matthew and Tserenjigmid, Gerelt , title =. arXiv preprint arXiv:2303.06336 , year =

    Pith/arXiv arXiv

  64. [64]

    and Li, Shangwen and Martin, Daniel J

    Caplin, Andrew and Deming, David J. and Li, Shangwen and Martin, Daniel J. and Marx, Philip and Weidmann, Ben and Ye, Kadachi Jiada , title =. Management Science , year =. doi:10.1287/mnsc.2024.08994 , note =

    doi:10.1287/mnsc.2024.08994 2024

  65. [65]

    Psychological Science , volume =

    Haddara, Nadia and Rahnev, Dobromir , title =. Psychological Science , volume =

  66. [66]

    arXiv preprint arXiv:2502.08501 , year =

    Grimon, Marie-Pascale and Mills, Christopher , title =. arXiv preprint arXiv:2502.08501 , year =. doi:10.48550/arXiv.2502.08501 , note =

    work page doi:10.48550/arxiv.2502.08501

  67. [67]

    Aversion to Hiring Algorithms: Transparency, Gender Profiling, and Self-Confidence , journal =

    Dargnies, Marie-Pierre and Hakimov, Rustamdjan and K. Aversion to Hiring Algorithms: Transparency, Gender Profiling, and Self-Confidence , journal =. 2026 , doi =

    2026

  68. [68]

    and Monahan, Amy B

    Choi, Jonathan H. and Monahan, Amy B. and Schwarcz, Daniel , title =. Minnesota Law Review , volume =. 2024 , doi =

    2024

  69. [69]

    arXiv preprint arXiv:2409.02391 , year =

    Merali, Ali , title =. arXiv preprint arXiv:2409.02391 , year =

    Pith/arXiv arXiv

  70. [70]

    Journal of experimental psychology: General , volume =

    Dietvorst, Berkeley J and Simmons, Joseph P and Massey, Cade , title =. Journal of experimental psychology: General , volume =

  71. [71]

    Computational Brain & Behavior , volume =

    Tejeda, Heliodoro and Kumar, Aakriti and Smyth, Padhraic and Steyvers, Mark , title =. Computational Brain & Behavior , volume =

  72. [72]

    Proceedings of the ACM on Human-Computer Interaction , volume =

    Green, Ben and Chen, Yiling , title =. Proceedings of the ACM on Human-Computer Interaction , volume =

  73. [73]

    Gruber, Jonathan and Handel, Benjamin R and Kina, Samuel H and Kolstad, Jonathan T , title =

  74. [74]

    and Hauskrecht, Milos , title =

    Naeini, Mahdi Pakdaman and Cooper, Gregory F. and Hauskrecht, Milos , title =. Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence , pages =

  75. [75]

    Journal of Mathematical Psychology , volume =

    Bamber, Donald , title =. Journal of Mathematical Psychology , volume =

  76. [76]

    Radiology , volume =

    Hanley, James A and McNeil, Barbara J , title =. Radiology , volume =

  77. [77]

    Psychological review , volume =

    Moore, Don A and Healy, Paul J , title =. Psychological review , volume =

  78. [78]

    Journal of Political Economy , volume =

    de Clippel, Geoffroy and Zhang, Xu , title =. Journal of Political Economy , volume =

  79. [79]

    Studia Scientiarum Mathematicarum Hungarica , volume =

    Csisz\'ar, Imre , title =. Studia Scientiarum Mathematicarum Hungarica , volume =

  80. [80]

    and Gentzkow, Matthew and Yu, Chuan , title =

    Chan, David C. and Gentzkow, Matthew and Yu, Chuan , title =. Working Paper , year =

Showing first 80 references.