Embedded Nudge: (De)Biased AI and Earnings Forecasts
How to cite
Papadopoulos, Lazaros. 2026. "Embedded Nudge: (De)Biased AI and Earnings Forecasts." AEA RCT Registry. AEARCTR-0019013. https://doi.org/10.1257/rct.19013-2.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
In this pre-registered online experiment I investigate how a design choice made by the vendor of an AI assistant shapes financial judgment when the underlying information is held fixed. Participants forecast a firm's future earnings and allocate a hypothetical investment after reading an earnings conference-call transcript. Using a correspondence design, I hold all financial content constant and vary only the demographic characteristics signaled by the executives' names. I cross this with random assignment to AI assistance whose content reflects different vendor design choices, including a no-assistance control. The design aims to identify whether the identity signaled by a firm's executives moves earnings forecasts and investment when fundamentals are held constant, and whether the vendor's design choice then amplifies, leaves unchanged, or attenuates that gap. Secondary analyses examine participants' perceptions of and reliance on the assistance. The results speak to how the design choices embedded in widely used AI tools shape earnings forecasts and investment.
Intervention
The intervention is what each participant receives while forecasting a firm's earnings from an earnings conference-call transcript, and it has two randomized components. (1) AI assistance: participants receive AI-generated analytical assistance whose content reflects a vendor's design choice, or no assistance. (2) Identity context: the demographic characteristics signaled by the firm's executives' names are randomly varied while all financial content is held identical. The study treats the vendor's design choice as an embedded nudge in the tool's choice architecture. The factorial structure, session procedure, and outcome measures are described under Experimental Design.
Experimental design
A 2 × 3 between-subjects randomized factorial crossing the two manipulations under Intervention (identity × AI assistance), with six equally allocated arms randomized at the individual level. In a single online session, participants give consent, complete a short demographics and financial-literacy questionnaire, and pass a comprehension check. They then read the earnings conference-call transcript (key terms hoverable; reading time recorded) and, in the AI arms, may consult the assistant. They next provide the outcome measures: a categorical and a continuous earnings forecast, a confidence rating, a hypothetical investment allocation, and a brief written rationale. Finally, they complete the perception-and-reliance questionnaires and end-of-session manipulation checks, followed by debriefing. The design isolates whether executives' signaled identity affects judgment when fundamentals are fixed, and how the vendor's AI design choice moderates it.
Design details
Cited by
No citation data available for this registration yet.