AI Match

How to cite

Weidmann, Ben; Rio-Chanona, Maria Del; Piyapromdee, Suphanit; Einsiedler, Johanna. 2026. "AI Match." AEA RCT Registry. AEARCTR-0018555. https://doi.org/10.1257/rct.18555-1.1

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

Ben Weidmann* , Maria Del Rio-Chanona , Suphanit Piyapromdee , Johanna Einsiedler

* Primary investigator

Registry ID
AEARCTR-0018555
DOI
https://doi.org/10.1257/rct.18555-1.1
First registered
Last updated
Status
In development
Trial dates
2026-06-01 – 2027-03-01
Countries
Private
JEL codes
J24, O33, C91, C92
Registry record
http://www.socialscienceregistry.org/trials/18555

Abstract

This laboratory experiment studies how individuals use generative AI tools to complete computer-based, cognitive tasks, and how the productivity effects of AI vary across users with different baseline skill profiles. Participants complete problems in two domains: 'analytical' and 'social'. The experiment has two phases. First, participants work without AI. Second, participants are randomly assigned according to a 2×2 design that crosses AI access (AI vs. no AI) with a brief intervention (practical AI training vs. active control). We measure task performance, beliefs about own performance, AI usage from chat logs, and a range of baseline cognitive and personality measures.

Intervention

The study uses a 2×2 between-subjects design. The first intervention is access to a generative AI tool during the Phase 2 task block (AI vs. no AI). The second intervention, delivered between the Phase 1 (baseline) and Phase 2 (main experimental tasks) is a brief practical training session (vs. an active control activity of equivalent length and structure). All four cells (AI × training; AI × control; no-AI × training; no-AI × control) are equally populated, in expectation.

Experimental design

A laboratory experiment lasting approximately two hours. Participants first complete cognitive and personality measures, then complete baseline (Phase 1) versions of an analytical exercise and a negotiation exercise without AI, and report beliefs about their own performance. They are then randomly assigned to one cell of a 2×2 design (AI access × practical training/active control), receive their assigned training or control activity, and complete post-baseline (Phase 2) versions of both exercises under their assigned AI condition.

Design details

Randomization method
Randomization done in office by a computer. We use simple randomization. Each participant has a 25% chance of being assigned to each cell of the 2x2 design.
Randomization unit
Individual
Clusters
800 individual participants
Observations
800
Arms
We are using simple randomisation. In expectation, we are aiming for 200 participants in each of the 4 treatment arms.
Primary outcomes
Our key outcome variables are 'Phase 2 task performance'. In particular, we measure participant performance on an analytical task and a negotiation task.
Secondary outcomes
Secondary outcomes are constructed primarily from two sources of chat data: (a) the participant's interaction log with the AI tool (b) the negotiation chat between paired participants

Cited by

No citation data available for this registration yet.

Keywords