Persuasion Through AI-Enhanced Communication

How to cite

Vargas, Kristian López; Grasso, Gonzalo Respighi; Iriarte, Julian Martinez; Dube, Valerie. 2025. "Persuasion Through AI-Enhanced Communication." AEA RCT Registry. AEARCTR-0016988. https://doi.org/10.1257/rct.16988-1.0

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

Kristian López Vargas* , Gonzalo Respighi Grasso , Julian Martinez Iriarte , Valerie Dube

* Primary investigator

Registry ID
AEARCTR-0016988
DOI
https://doi.org/10.1257/rct.16988-1.0
First registered
Last updated
Status
On going
Trial dates
2025-10-09 – 2026-10-09
Countries
Private
JEL codes
C91, D03, O33
Registry record
http://www.socialscienceregistry.org/trials/16988

Abstract

This laboratory experiment investigates how communication—both human and AI-assisted—affects generosity in the dictator game. In the baseline condition, Player A (the dictator) decides how much money to share with Player B (the recipient) without any communication. In a second condition, Player B can write a message to Player A before the decision. In the third condition, Player B can use an AI chatbot to help craft that message. By comparing outcomes across these conditions, the study explores whether access to AI tools enhances persuasive communication and increases altruistic behavior.

Intervention

Participants are recruited on Prolific to play a one-shot Dictator Game (DG) online. Each participant is assigned a fixed role (Dictator = Player A; Receiver = Player B) for the entire session. We implement three pre-decision communication conditions:
T1 — No communication (baseline): A decides the allocation with no message from B.
T2 — Human message: Before A’s decision, B can send one short message. Messages must be non-identifying and non-offensive; they are shown verbatim to A.
T3 — AI-assisted message: Before A’s decision, B can draft a message using an embedded LLM assistant. B edits and chooses the final text to send.
The interface enforces length/content constraints and records the final message and any AI interaction metadata (e.g., whether the assistant was opened).

Experimental design

We use a within-subjects, block-randomized design. Each participant experiences all three treatment conditions in three blocks; block order is randomized at the session level. Within each block, there are multiple independent rounds with random, anonymous rematching in every round. Roles are fixed across blocks/rounds.
Unit of analysis (primary): Individual decision makers.
Compliance: In T2/T3, Player B may choose to send nothing; we define intent-to-treat (ITT) by assigned block and possibly report contrasts restricted to rounds in which a message was sent.

Sample size (plan): 360 participants total (balanced between Player As and Player Bs), all exposed to all three blocks.
180 in each role. Each interacts in 12 rounds (4 for each block/treatment). That is, 2,160 observations (720 observations for treatment).

Randomization: computer-implemented; block order randomized; matchings randomized each round.
Blinding: Participants only see counterpart’s role (and any standard, non-identifying info if used).

Design details

Randomization method
Randomization: computer-implemented; block order randomized; matchings randomized each round.
Randomization unit
Randomization: computer-implemented; block order randomized; matchings randomized each round.
Clusters
360 participants total (balanced between Player As and Player Bs), all subjects will be exposed to all three treatments (in random order). 180 participants in each role, Player A and Player B.
Observations
360 participants total (balanced between Player As and Player Bs), all subjects will be exposed to all three treatments (in random order). 180 participants in each role, Player A and Player B. Each participant interacts in 12 rounds (4 rounds per block/treatment). That is, 180*12 = 2,160 observations Since we have three treatments, we have: 2,160 /3 = 720 observations per treatment.
Arms
It is a within-subject design. All participants experience all treatments. We will have 180 clusters (participants) in each role.
Primary outcomes
The primary outcome is the amount transferred by Player A (the dictator) to Player B (the recipient) in the dictator game. This captures generosity or altruism under different communication conditions.
Secondary outcomes
Secondary outcomes involve text-based analysis of messages written under different conditions, focusing on linguistic features, tone, structure, and content differences across human-written and AI-assisted messages.

Cited by

No citation data available for this registration yet.

Keywords