The role of AI in trust: An experimental study using ChatGPT

How to cite

WU, JIABIN; Holdahl, Ethan; Wiegand, Connor; Bivins, Tanner. 2023. "The role of AI in trust: An experimental study using ChatGPT." AEA RCT Registry. AEARCTR-0011748. https://doi.org/10.1257/rct.11748-1.0

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

JIABIN WU* , Ethan Holdahl , Connor Wiegand , Tanner Bivins

* Primary investigator

Registry ID
AEARCTR-0011748
DOI
https://doi.org/10.1257/rct.11748-1.0
First registered
Last updated
Status
In development
Trial dates
2023-07-10 – 2025-06-30
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/11748

Abstract

This paper examines the function of artificial intelligence (AI) as an assistant in a two-player trust game. Prior to the commencement of the game, the second player has the option to transmit a message to the first player. With the aid of AI, the second player can potentially enhance and rephrase the message before sending it. Similarly, the first player can utilize AI to comprehend the message received from the second player. To investigate the impact of AI on players' communication and decision-making in this strategic scenario, we conduct an experiment with various conditions: with or without AI assistance for the second player, the first player, or both, and whether the players have knowledge about their opponents receiving AI assistance or not.

Experimental design

In the experiment, subjects are randomly matched in pairs to play a two-player two-stage binary trust game. Before the game starts, the second mover has the option to send a free-form message to the first mover and either one or both or none of the two players in each pair have an AI assistant to help with elaboration of the message (for the sender) and interpretation of the message (for the receiver).

Treatments: Benchmark with communication; AI for first mover only (public information); AI for second mover only (public information); AI for both players (public information); AI for first mover only (private information); AI for second mover only (private information); AI for both players (private information)

The experiment is programmed with oTree and incorporate GPT 3.5 Turbo as the AI.

Design details

Randomization method
The treatment randomization is done at the session level. All other randomizations are done by computer.
Randomization unit
Treatment is randomized at the session level. Pair matching and the game are done at the individual level.
Clusters
210 pairs
Observations
420 subjects
Arms
Each treatment with 30 pairs of players.
Primary outcomes
Second mover's original message. AI's rephrase of second mover's original message, AI's interpretation of second mover's sent message; First mover's trusting behavior; Second mover's trustworthiness behavior; First mover's belief of second mover's trustworthiness behavior; Second mover's belief of first mover's belief.

Cited by

No citation data available for this registration yet.

Keywords