Do Generative AI Chatbots Increase Dominant-Strategy Play in Matching Mechanisms?

How to cite

Kitadai, Ayato; Nishino, Nariaki. 2026. "Do Generative AI Chatbots Increase Dominant-Strategy Play in Matching Mechanisms?." AEA RCT Registry. AEARCTR-0017736. https://doi.org/10.1257/rct.17736-1.0

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

Ayato Kitadai* , Nariaki Nishino

* Primary investigator

Registry ID
AEARCTR-0017736
DOI
https://doi.org/10.1257/rct.17736-1.0
First registered
Last updated
Status
In development
Trial dates
2026-01-30 – 2026-03-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17736

Abstract

A central challenge in mechanism design is that strategy-proof or otherwise equilibrium-tractable mechanisms may still fail in practice when participants do not understand the incentive structure. In matching environments, truthful reporting is often observed, yet it can reflect a focal default rather than sophisticated dominant-strategy play. This study tests whether generative AI chatbots can reduce mechanism understanding costs and thereby increase dominant-strategy play in matching mechanisms, including cases in which truth-telling is not optimal. We conduct a 2×2 laboratory experiment that cross-randomizes the assignment rule (Deferred Acceptance vs. Reverse Deferred Acceptance) and the availability of a generative AI chatbot (present vs. absent). Participants make a one-shot preference submission in a school-choice environment with priorities and random tie-breaking; they interact with computerized counterparts, and payoffs depend on the assigned school. Instruction time and on-screen layout are held constant across conditions. In the AI condition, participants may ask questions in natural language, and the chatbot is constrained to support comprehension of the provided instructions without recommending specific actions or revealing meta-level properties of the mechanisms. Our primary outcome is dominant-strategy play, defined as truthful reporting under Deferred Acceptance and reverse-truthful reporting under Reverse Deferred Acceptance. We compare dominant-strategy play between the AI-available and AI-unavailable conditions within each mechanism, and estimate treatment effects using the assigned availability of the chatbot. Secondary outcomes include comprehension test performance, decision times, and usage logs, as well as post-experimental measures of perceived demand effects. The design directly tests whether access to generative AI assistance shifts behavior toward dominant strategies even when the dominant strategy deviates from truthful reporting, helping distinguish defaults from comprehension as drivers of equilibrium (non)play.

Experimental design

2×2 between-subjects factorial experiment. Factor A is mechanism (DA vs. RDA). Factor B is AI assistance (AI access vs. no AI). Participants make a one-shot ranking decision over four schools. Each participant is matched with three computer agents. Priorities are determined by residency (local vs. non-local) and a uniform random lottery among non-residents, displayed to participants. Payoffs depend on the assigned school slot via participant-specific payoff tables.

Design details

Randomization method
Randomization done in office by a computer
Randomization unit
Individual participant.
Clusters
200 individuals
Observations
200 individuals
Arms
50 treatments
Primary outcomes
Dominant-strategy play (DSPlay): Indicator variable equal to 1 if the participant submits the dominant-strategy rank-order list under the assigned mechanism (DA or RDA), and 0 otherwise.
Secondary outcomes
1. Dominant-strategy play conditional on submission (DSPlay_Submit): Indicator equal to 1 if the participant submits a list by the deadline and the submitted list matches the dominant strategy under the assigned mechanism; 0 otherwise. Defined among submitters only. 2. Comprehension quiz accuracy (QuizCorrect): Indicator equal to 1 if the incentivized comprehension quiz is answered correctly within the quiz time limit; 0 otherwise. 3. Decision time (DecisionTime): Time in seconds from entering the preference-submission screen to final submission. 4. AI usage measures (AIUseAny, AINumMsg, AITotalChars or AITotalTokens): Existence or number of participant messages and total characters (or total tokens) sent to the chatbot (AI-available arm only). 5. Perceived experimenter demand (DemandFeel, DemandText): DemandFeel is a categorical response (Yes, No, Unsure) to whether the participant felt the experimenter expected a particular behavior; DemandText is a free-response explanation collected if Yes or Unsure. 6. Perceived influence of the right-side panel (PanelInfluence, PanelReason): PanelInfluence is a 5-point scale on whether the right-side panel affected understanding or judgment; PanelReason is a free-response explanation. 7. Prior knowledge of DA (PriorDA): Post-experimental survey response to “Before participating in this experiment, did you already know the Deferred Acceptance (DA) algorithm (or Gale-Shapley school choice mechanism)?” Response options: Yes, No, Unsure.

Cited by

No citation data available for this registration yet.

Keywords