Health Data Sharing in the Age of AI
How to cite
Choi, Eseul; Kang, Hyewon; Sung, Won. 2026. "Health Data Sharing in the Age of AI." AEA RCT Registry. AEARCTR-0018237. https://doi.org/10.1257/rct.18237-1.0
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
This study examines how information of artificial intelligence (AI) influences individuals’ willingness to share personal health data and their valuation of privacy protections. Using a 2 × 4 factorial design, Korean adults are randomly assigned to conditions varying in data-sharing institution types (for-profit vs. non-profit) and AI information (none, benefits, risks, or both). For five types of health data that differ in sensitivity, we measure respondents’ willingness to share information using a 7-point Likert scale and elicit their minimum acceptable compensation (willingness to accept, WTA). We further assess preferences for four alternative privacy policies using preference rankings and willingness-to-pay via a double-bounded dichotomous choice format. The findings offer policy-relevant insights for designing proportionate and targeted health data governance frameworks in the era of AI.
Intervention
Our survey experiment employs a 2 × 4 factorial design that varies institutional context and AI information. The first dimension manipulates institution type (for-profit vs. non-profit), which is presented to all participants before the health data–sharing questions. Specifically, respondents are informed that their data would be shared with either (1) a for-profit company or (2) a non-profit research or public institution.
The second dimension varies informational framing about AI through four conditions: none, benefits, risks, and combined benefits–risks. Following the institutional treatment, participants are shown up to two information cards describing (i) the expected benefits of AI use (e.g., personalized medical services) and/or (ii) potential privacy risks (e.g., re-identification). The four AI information conditions are defined as follows:
• T1 – AI Benefits card
• T2 – AI Risks card
• T3 – Both cards: Both the Benefits and Risks cards are presented simultaneously
• Control: Neither card is shown
After the experiment module, all participants proceed to the policy preference survey module, in which both the Benefits and Risks cards are shown again alongside descriptions of four types of AI privacy policies.
Experimental design
Participants are recruited through Hankook Research's panel and complete a 15–20-minute online survey. It proceeds as follows.
(1) Survey: baseline measures of privacy attitudes, AI knowledge, health care experience, and trust in institutions
(2) Experiment: information provision experiment
A. Random information provision treatment: institution type (for-profit vs. non-profit) × AI information (control, T1-benefits, T2-risks, T3-both)
B. Outcome measures: For each of five data types, Intention to Share (7-point Likert scale) and WTA
C. Manipulation check
(3) Survey: policy preference survey
A. Information provision: full disclosure of AI information (same as T3) and four AI-era privacy policies
B. Outcome measures: policy rank by importance and WTP for the top two ranked policies
(4) Survey: socioeconomic status
Design details
Cited by
No citation data available for this registration yet.