Assess what higher perceived empathy ratings for LLM-generated health care responses establish—and what they leave unresolved about communication quality, safety, and deployment.
Approach:
Evidence reviewed: Discusses a systematic review and meta-analysis of LLM-based tools for physician-patient communication, alongside related studies and emerging real-world evaluations.
Critical perspective: Examines the limits of written-response ratings and considers safety, oversight, workflow, health literacy, equity, and validation needs.
Key Findings:
Across 10 studies, AI responses received higher perceived-empathy ratings than physician responses (standardized mean difference 1.02, 95% CI 0.44–1.60; odds ratio 5.5). AI responses were also longer and clearer, and were frequently judged more useful.
Empathy ratings did not differ in the sole disease-specific study. The review reported substantial heterogeneity (I²=73%) and no randomized trials, longitudinal data, or direct measurement of trust.
In one study, reviewers identified some AI responses as potentially harmful, while patients rated them highly and could not distinguish them.
The review reports that AI-generated text is generally longer and may exceed the reading level suitable for patients with limited health literacy.
Interpretation:
The article argues that perceived empathy in written responses does not establish effective communication within an ongoing clinical relationship. It also notes that empathy ratings do not establish accuracy, comprehension, equity, trust, adherence, or safety. The article calls for oversight supported by clinician expertise, time, tools, and incentives, and for evaluation proportionate to each use case.
Limitations:
The evidence discussed is largely based on simulated, asynchronous, text-based interventions and blinded ratings rather than communication over time.
The review lacks randomized trials, longitudinal data, and direct trust measurement; its results show substantial heterogeneity.
The article notes a gap between the interventions studied and production tools already being used, including draft portal replies and ambient documentation.
Conclusion:
Higher empathy ratings alone do not establish deployment readiness. The article states that readiness requires local validation with patients and message types; longitudinal auditing for hallucination, clinical inconsistency, and drift; readability assessment involving people with low health literacy; subgroup review across condition, language, age, disability or assistive needs, prior knowledge, and question type; and workflow monitoring for clinician burden, additional review work, automation bias, and uneven benefit.
