AI Framing and Review Behavior: Evidence from HR and Finance Professionals

How to cite

Wiles, Emma; Hsu, Megan; Kropp, Matthew; Bedard, Julie. 2026. "AI Framing and Review Behavior: Evidence from HR and Finance Professionals." AEA RCT Registry. AEARCTR-0017596. https://doi.org/10.1257/rct.17596-1.0

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

Emma Wiles* , Megan Hsu , Matthew Kropp , Julie Bedard

* Primary investigator

Registry ID
AEARCTR-0017596
DOI
https://doi.org/10.1257/rct.17596-1.0
First registered
Last updated
Status
In development
Trial dates
2026-01-07 – 2026-01-30
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17596

Abstract

We study how framing an assistant as an AI tool, a human employee, or an AI employee affects (i) error detection in document review, (ii) escalation / delegation behavior, (iii) confidence in sign‑off, (iv) preferences over firm governance for such systems, and (v) adoption and sentiment toward AI.
Participants review a set of documents (HR or Finance) that contain seeded errors, after seeing an assistant framed in one of three ways. We then measure performance and post‑task attitudes and governance preferences.

Intervention

Participants are randomly assigned at the individual level to one of three framing conditions:
T1: AI Tool – The drafts are presented as produced by an AI tool.
T2: Human Employee – The drafts are presented as produced by a human employee the participant supervises.
T3: AI Employee – The drafts are presented as produced by “ALEX-3, an AI employee” the participant supervises.

In addition, participants are assigned to review documents in one domain:
HR (e.g., job descriptions), or
Finance (e.g., expense/budget documents),

with random assignment stratified below.
Within a domain, participants review a sequence of documents (D1–D5). Document order is randomized using a Latin square design so that each document appears in each position equally often.
All participants see the same underlying drafts and seeded errors; only the framing of the assistant differs.

Experimental design

See attached Pre-Analysis Plan

Design details

Randomization method
Stratification and randomization: Randomization to framing conditions (T1–T3) is stratified by: Domain: HR vs Finance. Review frequency: reviews others’ work at least several times per week vs less often. AI usage at work: uses GenAI tools daily vs less often. We will include stratum fixed effects in the main analyses.
Randomization unit
Individual
Clusters
NA
Observations
1400 Registered, expected N around 790
Arms
260ish per treatment arm.
Primary outcomes
(i) error detection in document review, (ii) escalation / delegation behavior, (iii) confidence in sign‑off, (iv) preferences over firm governance for such systems, and (v) adoption and sentiment toward AI. See Pre Analysis Plan attached

Cited by

No citation data available for this registration yet.

Keywords