The digital therapist? LLMs and the future of clinical dialogue

Large language models (LLMs) are increasingly used in healthcare for triage, documentation, patient education, decision support, and mental health communication. Although these systems can generate fluent, empathic, and clinically plausible language, their safe use depends on clear task boundaries, technical design, and human oversight. This article offers a conceptual narrative review of LLM-mediated clinical dialogue, focusing on three questions: what forms of clinical understanding LLMs can and cannot support; how different types of LLM systems vary in reliability and risk; and how they may be responsibly integrated into real-world clinical workflows. We distinguish general-purpose LLMs, domain-specific medical LLMs, retrieval-augmented systems, multimodal models, and agent-based systems, and align each with appropriate clinical use cases. We argue that LLMs should be treated as assistive tools for documentation, information retrieval, structured reflection, and low-risk patient communication, rather than as autonomous clinicians or therapists. Practical safeguards—including retrieval grounding, uncertainty disclosure, clinician review, audit trails, escalation pathways, and post-deployment monitoring—are necessary to reduce hallucination, bias, and overreliance. The article concludes that LLMs may support healing dialogue only when they expand, rather than replace, the clinician’s interpretive responsibility, relational presence, and moral accountability.