RPD315 - Does Disclosure of AI Use Change Patient Response to Empathetic Communication?: Early Results of the PALL-AI Trial

Key Message : As AI use in healthcare increases, many are calling for transparent disclosure of AI communication to patients, despite a lack of evidence if disclosure may alter patient perception. PALL-AI, a clinical trial to test this phenomenon, found disclosure did not alter patient preferences toward AI communication.

Abstract:

Introduction: Artificial intelligence (AI) is increasingly integrated into healthcare, including large language models (LLMs), which may augment clinician communication [1]. Ethical frameworks recommend disclosure when LLMs are used, but studies outside medicine suggest disclosure can reduce receptivity to AI communication [2][3]. PALL-AI is the first clinical trial to assess this effect in Hospice and Palliative Medicine.

Methods: PALL AI was a single-center, federally registered clinical trial completed in 2025. The trial enrolled 216 adults with any malignancy who were receiving outpatient palliative care. Each participant reviewed two standardized empathetic statements about coping with cancer— one of three authored by palliative care physicians and one of three by LLMs (ChatGPT 4.0, Bing Copilot, Google Gemini)—all produced from an identical prompt and matched for length. Participants were randomized to a blinded arm (no authorship disclosed) or an unblinded arm (informed which statement was human vs AI). Demographic data was also collected. The primary outcome was patient preference for AI statements in the disclosure vs nondisclosure arm. Secondary analyses assessed overall preference (human vs AI), differences by LLM, and demographic subgroups.

Results: Of 216 patients enrolled, 212 completed the study per protocol. There was no significant difference in AI preference between blinded and unblinded arms (RR=1.08, 95% CI: 0.72-1.64, p=0.72). Overall, AI-authored statements were preferred over human-authored statements (95% CI: 64.1-76.3%, p< 0.001). Preference did not differ significantly across the three LLMs. Preferences did not differ significantly by gender, age ( < 65 vs ≥65), race (White vs non-White), and education (college degree vs not) (all p> 0.05). Findings did not differ significantly between intention-to-treat and per-protocol analysis.

Conclusions: In this trial, disclosure of AI authorship did not reduce patient preference for AI-generated empathetic statements; AI statements were preferred overall. These findings suggest disclosure alone may not diminish receptivity to LLM-assisted communication in palliative care.

References: 1. Rajkomar A, Dean J, Kohane I. Machine Learning in Medicine. N Engl J Med. 2019;380(14):1347-1358. doi:10.1056/NEJMra1814259.
2. Luonuansuu TA, Christensen AR, Hutchinson SZ. Artificial Intelligence/Machine Learning in Palliative Care #492. J Palliat Med. 2024;28(1). doi:10.1089/jpm.2024.0401 Liebert Publishing+1
3. Papneja, H., Yadav, N. Self-disclosure to conversational AI: a literature review, emergent framework, and directions for future research. Pers Ubiquit Comput 29, 119–151 (2025). https://doi.org/10.1007/s00779-024-01823-7