JMIR Publications

ABSTRACT

Is more perceived empathy better? Richter and colleagues offer a careful, PRISMA-based systematic review and meta-analysis of how large language model (LLM)–based tools affect physician-patient communication. Their headline finding is notable: across studies, LLM-generated responses were rated as more empathetic than physicians’, with a large pooled effect on empathy ratings and consistently favorable judgments of clarity and completeness. For health-system leaders responsible for deploying and governing these tools, that finding warrants caution as much as enthusiasm. We read this synthesis not as a verdict, but as a precise inventory of what we still do not know and what we recommend in response. The evidence is almost entirely cross-sectional, simulated, and asynchronous. There are no randomized trials, no real-world deployments, no longitudinal follow-up, and no direct measurement of trust or safety, the outcomes that matter most when deciding whether a communication tool is safe to scale. Patients also could not reliably detect inaccurate or potentially harmful LLM answers, and sometimes rated flawed responses highly. That is the safety problem. Empathy helps build trust, but it is not enough;. Human oversight must be built into design and governance. Before deploying these tools, health systems need local validation, ongoing monitoring for accuracy and equity, and pragmatic trials that measure trust, safety, and outcomes alongside perceived empathy. Empathy that generalizes across raters and settings is a genuine capability, but it carries a real risk of mistrust when accuracy has not been established.