Decision making following advice from humans and AI
How to cite
Thunstrom, Linda; DePaemelere, Collin; Mollerstrom, Johanna. 2026. "Decision making following advice from humans and AI." AEA RCT Registry. AEARCTR-0019360. https://doi.org/10.1257/rct.19360-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
This study examines how advice affects decision-making in an allocation task and how people evaluate the social appropriateness of different allocation choices. It also examines whether these effects differ depending on whether the advice comes from another person or from AI.
Intervention
The project consists of 3 related studies.
Study 1 (N=950): Participants are asked to rate the social appropriateness of dictator game (DG) allocations (splits of $10 between the proposer and recipient, in $1-increments). Participants are randomized into one of 5 conditions:
- control (no advice): the participant is asked to rate the social appropriateness of the possible splits in the DG
- selfish-human: the participant needs to rate the social appropriateness in the scenario where the proposer in the DG gets advice from a human to choose a $8/$2 split between themselves/recipient
- fair-human: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from a human to choose a $5/$5 split between themselves/recipient
- selfish-AI: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from an AI to choose a $8/$2 split between themselves/recipient
- fair-AI: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from an AI to choose a $5/$5 split between themselves/recipient
Study 2 (N=1200): Participants are asked to choose dictator game (DG) allocations (splits of $10 between themselves and a recipient, in $1-increments). Participants are randomized into one of 5 conditions:
- control (no advice): the participant gets no advice before choosing their allocation
- selfish-human: before choosing their allocation, the participant gets advice from a human to choose a $8/$2 split between themselves/recipient
- fair-human: before choosing their allocation, the participant gets advice from a human to choose a $5/$5 split between themselves/recipient
- selfish-AI: before choosing their allocation, the participant gets advice from an AI to choose a $8/$2 split between themselves/recipient
- fair-AI: before choosing their allocation, the participant gets advice from an AI to choose a $5/$5 split between themselves/recipient
Study 3 (N=120): All participants are asked to predict the allocation most commonly chosen by participants in the selfish-AI-treatment in Study 2.
Experimental design
The project consists of 3 related studies.
Study 1 (N=950): Participants are asked to rate the social appropriateness of dictator game (DG) allocations (splits of $10 between the proposer and recipient, in $1-increments). Participants are randomized into one of 5 conditions:
- control (no advice): the participant is asked to rate the social appropriateness of the possible splits in the DG
- selfish-human: the participant needs to rate the social appropriateness in the scenario where the proposer in the DG gets advice from a human to choose a $8/$2 split between themselves/recipient
- fair-human: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from a human to choose a $5/$5 split between themselves/recipient
- selfish-AI: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from an AI to choose a $8/$2 split between themselves/recipient
- fair-AI: the participant is asked to rate the social appropriateness in the scenario where the proposer in the DG gets advice from an AI to choose a $5/$5 split between themselves/recipient
Study 2 (N=1200): Participants are asked to choose dictator game (DG) allocations (splits of $10 between themselves and a recipient, in $1-increments). Participants are randomized into one of 5 conditions:
- control (no advice): the participant gets no advice before choosing their allocation
- selfish-human: before choosing their allocation, the participant gets advice from a human to choose a $8/$2 split between themselves/recipient
- fair-human: before choosing their allocation, the participant gets advice from a human to choose a $5/$5 split between themselves/recipient
- selfish-AI: before choosing their allocation, the participant gets advice from an AI to choose a $8/$2 split between themselves/recipient
- fair-AI: before choosing their allocation, the participant gets advice from an AI to choose a $5/$5 split between themselves/recipient
Study 3 (N=120): All participants are asked to predict the allocation most commonly chosen by participants in the selfish-AI-treatment in Study 2.
Design details
Cited by
No citation data available for this registration yet.