On a busy call shift as a resident doctor at Stanford, I have heard versions of the same sentence more than once: “That’s reassuring. ChatGPT said the same thing.” In clinic, patients increasingly admit they left visits hesitant to ask clarifying questions, then turned to a generative AI tool to explain their diagnosis in plainer language. Some return reassured; others come back more uncertain. But the pattern is unmistakable. Patients are no longer using AI merely as a reference. They are actively comparing their doctors to a chatbot.
This shift matters because it reframes expectations of communication, empathy, and responsiveness in health care. A 2023 study published in JAMA Internal Medicine found that clinicians rated chatbot responses to real patient questions as higher quality and more empathetic than physicians’ responses. While researchers continue to benchmark AI against clinicians using standardized prompts and expert raters, patients have already moved ahead of that debate. In practice, they are benchmarking clinicians against AI, and those comparisons are beginning to shape trust and satisfaction.
The shift in patient expectations
Let’s be honest: Traditional care models, including 15-minute visits, asynchronous portals, and delayed responses, are not meeting patient expectations. Generative AI offers patients instant explanations in plain language, written conversations to revisit, unhurried empathy, and free access. These features can enhance understanding and self-efficacy, particularly for people managing chronic or complex conditions. Yet AI limitations are real. These tools lack access to the full medical record and may generate confident but incorrect guidance. Their fluency can undermine clinician recommendations when the two diverge, especially in situations involving uncertainty, risk trade-offs, or resource constraints.
Navigating the dual reality for clinicians
For clinicians, generative AI presents a dual reality. On one hand, AI-assisted communication reduces documentation burden and cognitive load, freeing time and attention for patient interaction. On the other, evidence that patients perceive AI-generated responses as more empathetic raises uncomfortable questions about how clinical expertise and authority are experienced. When clinicians must explain uncertainty, decline unnecessary testing, or contradict an AI-generated suggestion, they do so within a framework of professional accountability and legal liability that the technology does not share.
Many health systems are piloting generative AI tools to draft responses to non-urgent patient portal messages, allowing clinicians to edit and approve the final communication. This preserves human oversight while standardizing tone and structure, but it quietly establishes AI-mediated communication as the institutional norm. Meanwhile, patients encounter consumer-facing tools that are immediate, patient-focused, and unconstrained by visit length or business hours.
The rise of the shadow record
As generative AI moves from the periphery of care delivery into its communicative core, clinical conversations are occurring entirely outside the view of the health system. Without intentional integration, doctors like me risk missing opportunities to improve access, efficiency, and early detection. In principle, bidirectional information flow between patient-facing AI interactions and the clinical record could resolve routine clarification questions, reduce the number of messages that require a doctor’s response, and surface concerning symptom narratives that merit expedited evaluation. Early deployments of AI for clinical documentation and message drafting already demonstrate that large-scale integration is technically feasible, even as the move from documentation support to patient-facing advice raises new liability and governance challenges.
From a systems perspective, the greatest risk may be relational and informational rather than technical. Most health systems lack structured pathways for patients to share AI-generated explanations, upload conversation logs, or flag discrepancies between chatbot advice and clinical recommendations. As a result, a growing volume of symptom narratives and decision-influencing information now exists outside the electronic health record, forming a parallel “shadow record” disconnected from clinical workflows, quality oversight, and accountability.
Strategic integration and the path forward
The next phase of generative AI in health care will hinge less on model performance than on integration strategy. The question is no longer whether patients will consult AI. They already do. The question is how health systems can incorporate these tools in ways that enhance, rather than erode, clinical relationships. Doing so will require:
- Governance models that distinguish fluency from responsibility
- Workflows that allow clinicians to contextualize and correct AI-derived information without stigma
- Infrastructure that captures patient-reported narratives wherever they originate
If patients are already benchmarking clinicians against generative AI, health systems cannot afford to remain passive observers. The task now is not to debate whether AI belongs in clinical care, but to design guardrails for how it is integrated. Done thoughtfully, AI can reduce communication friction and surface early warning signals between visits. Done haphazardly, it risks widening mistrust and fragmenting clinical responsibility. The choice is no longer whether to engage, but whether to lead.
Cybil Sierra Stingl is a plastic and reconstructive surgery resident. Robert M. Kaplan is an emeritus professor of health services and medicine.
