Belief Updating under AI Advice: An Experiment

How to cite

XUE, Lian; Khanna, Manshu; Wei, Lijia; Wang, Ziyi. 2025. "Belief Updating under AI Advice: An Experiment." AEA RCT Registry. AEARCTR-0015153. https://doi.org/10.1257/rct.15153-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Lian XUE* , Manshu Khanna , Lijia Wei , Ziyi Wang

* Primary investigator

Registry ID
AEARCTR-0015153
DOI
https://doi.org/10.1257/rct.15153-1.0
First registered
Last updated
Status
On going
Trial dates
2024-12-01 – 2025-05-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/15153

Abstract

The rapid integration of AI into healthcare presents a critical challenge: understanding how medical professionals incorporate AI advice into their diagnostic decision-making. While AI systems demonstrate increasing accuracy in medical diagnosis, their effectiveness ultimately depends on how humans update their beliefs in response to AI recommendations. We conduct a lab-in-the-field experiment examining belief updating processes when participants receive AI-generated medical advice. Our experimental design employs a quadratic scoring rule to elicit truthful reporting of beliefs across multiple treatment scenarios. Participants first provide prior probability estimates for treatment success, then report their expectations about the informativeness of upcoming AI advice, and finally update their beliefs after receiving AI recommendations. This three-stage elicitation, grounded in Bayesian updating frameworks, allows us to examine both the direct impact of AI advice on belief formation and the role of ex-ante trust in AI systems.

Intervention

This study examines how individuals update their beliefs when receiving AI-generated advice in medical decision-making scenarios. Participants evaluate treatment options across multiple hypothetical medical cases. For each case, participants: 1) provide their initial assessment of treatment success probabilities, 2) indicate their expectations about upcoming AI advice, and 3) make final assessments after receiving AI recommendations. The AI system used is ChatGPT-4o with a documented accuracy rate above 70%. All probability assessments are incentivized using proper scoring rules to encourage truthful reporting.

Experimental design

The experiment employs a 2×2 between-subjects factorial design investigating how individuals incorporate AI advice into medical decision-making. Participants evaluate treatment options across 15 hypothetical medical scenarios, providing probability estimates before and after receiving AI recommendations. The design uses incentive-compatible mechanisms to elicit truthful belief reporting. Treatments vary in the format of AI advice provided. All participants complete pre-experiment cognitive assessments and post-experiment questionnaires measuring algorithm literacy and trust.

Design details

Randomization method
Participants will be randomly assigned to one of four treatment conditions using a 2×2 factorial design at the session level. Randomization is implemented through the oTree experimental software platform prior to each session, ensuring balanced assignment across treatment conditions.
Randomization unit
Primary randomization occurs at the individual participant level for the assignment and ordering of medical scenarios. Additional session-level randomization is used to assign different cohorts of participants to experimental timeslots while maintaining balanced characteristics across sessions.
Clusters
48-64 experimental sessions with 5 participants each (12-16 sessions per treatment condition)
Observations
3,600-4,800 participant-scenario observations (240-320 participants × 15 scenarios each)
Arms
Four treatment conditions (2×2 factorial design): Treatment 1 (One suggestion, No explanation): 60-80 participants (12-16 sessions) Treatment 2 (One suggestion, With explanation): 60-80 participants (12-16 sessions) Treatment 3 (Two suggestions, No explanation): 60-80 participants (12-16 sessions) Treatment 4 (Two suggestions, With explanation): 60-80 participants (12-16 sessions)
Primary outcomes
1. Belief updating magnitude: Change in probability estimates between prior and posterior beliefs 2. Belief accuracy: Quadratic score of final probability estimates relative to true treatment outcomes 3. Second-order belief accuracy: Correlation between expected informativeness scores and actual belief updates

Cited by

No citation data available for this registration yet.

Keywords