Generative AI and Task Choice: Experimental Evidence on Task Choice and Income Gap

How to cite

Park, Junghyun. 2025. "Generative AI and Task Choice: Experimental Evidence on Task Choice and Income Gap." AEA RCT Registry. AEARCTR-0016197. https://doi.org/10.1257/rct.16197-1.0

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

Junghyun Park*

* Primary investigator

Registry ID
AEARCTR-0016197
DOI
https://doi.org/10.1257/rct.16197-1.0
First registered
Last updated
Status
On going
Trial dates
2025-05-13 – 2025-06-24
Countries
Private
JEL codes
J24, O33
Registry record
http://www.socialscienceregistry.org/trials/16197

Abstract

This study investigates the causal effects of generative AI (e.g., ChatGPT) on individuals’ task choices and resulting income disparities using a randomized controlled trial (RCT) within an experimental economics framework. Participants first complete two tasks—decoding (low-skill) and programming (high-skill)—without AI assistance during the Baseline stage. In the Task Round, participants are randomly assigned to one of four groups. The control group receives no AI assistance. All three treatment groups are granted access to AI but differ in additional design features: the basic AI treatment group performs the task with AI; the information treatment group receives summary statistics of prior participants’ task choices and earnings before making their own decision; and the repeated-task group completes two consecutive task rounds with AI. The primary outcome variables include the rate of high-skilled task selection (task choice), the rate of selecting a task misaligned with one’s baseline ability (task mismatch), and the income difference between the Task Round and Baseline stages. This design allows us to disentangle the effects of AI availability, informational nudges, and repeated exposure on task selection and income inequality.

Intervention

Participants are randomly assigned to one of four groups. One group performs tasks without AI. The remaining three treatment groups are given access to ChatGPT under different conditions: standard AI access, AI with informational support (past participants’ data), and AI with repeated task rounds. The intervention aims to examine how AI affects task choice and income inequality.

Experimental design

This study uses a randomized controlled trial (RCT) to investigate how generative AI affects individuals’ task choice and income. Participants first complete two tasks without AI support in the Baseline stage. They are then randomly assigned to one of four groups: one control group (no AI) and three treatment groups with different AI conditions. All groups choose and perform a task in the Task Round. The study aims to measure task selection, task mismatch, and income difference across groups.

Design details

Randomization method
Randomization was conducted using a computer-based algorithm implemented through the oTree platform.
Randomization unit
The unit of randomization varies by treatment arm. For the control group and the AI-only treatment group, randomization was conducted at the individual level. For the information treatment group and the repetition treatment group, randomization was conducted at the session level, with all participants in a given session receiving the same treatment. This mixed randomization structure was implemented to accommodate interventions that required shared information or session-wide instructions.
Clusters
We plan to conduct 13 experimental sessions, each involving approximately 20 participants. For the AI + Information and AI + Repetition groups, treatment will be randomized at the session level, with each session constituting one cluster. In the Control and AI-only groups, randomization will occur at the individual level.
Observations
250 participants
Arms
Approximately 60 participants per group
Primary outcomes
Task choice, Task mismatch, Income difference

Cited by

No citation data available for this registration yet.

Keywords