Can AI Virtual Patients Teach Dentists Empathy?

It’s 11 p.m., and a fourth-year dental student is explaining a root canal to a patient who never sighs, never checks the clock, and never remembers being asked the same clumsy question three times in a row. The patient isn’t real. It’s software — an AI virtual patient playing the person in the chair so the student can rehearse the hardest part of dentistry: talking to a frightened human being.

It sounds like a gimmick — teaching bedside manner with a bot. But a 2026 randomized controlled trial in Medical Teacher by Yuanmei Zhang and colleagues put the idea to a real test. Sixty fourth-year dental students were randomly split in two: half did eight weeks of self-directed practice with an AI virtual-patient system, half did the usual peer role-play. Blinded examiners then scored everyone in an Objective Structured Clinical Examination (OSCE). The guiding question is one dentistry keeps circling back to as AI arrives — and one we’ve asked before about how students actually use AI in their training: can a machine genuinely build a human skill, or does it just teach you to pass the test?

The study, in one glance

This was a prospective, single-blind, mixed-methods RCT at the First Affiliated Hospital of Zhengzhou University. The intervention group (n=30) spent eight weeks running self-directed patient interviews with an AI virtual patient that gave immediate feedback; the control group (n=30) practised the same skills through conventional peer role-play. Outcomes were measured at baseline, immediately after, and at a three-month follow-up using OSCE scores for inquiry completeness and communication, the Jefferson Scale of Empathy (JSE), and self-reported motivation and confidence. Twelve students were also interviewed about the experience.

AI virtual patients: a real trial, not a demo

What makes this study stand out isn’t the headline — it’s the design. Most AI-in-dentistry papers are retrospective or run on archived images. This one randomized real students, blinded the assessors, and followed them for months. And the effects weren’t trivial: a nearly 13-point communication gap with a large effect size, inquiry completeness up more than 12 points, and — crucially — an advantage still measurable at three months rather than evaporating once the novelty wore off. In interviews, students described practising deliberately, failing privately, and trying again without the social cost of fumbling in front of a classmate; several said it lowered their anxiety about hard conversations. A tireless, non-judgemental partner you can rehearse with at midnight is genuinely useful.

But did it teach empathy — or just how to pass the test?

Here’s where a skeptic has to slow down. Every outcome in this trial was a proxy for the thing we actually care about. The OSCE is a standardized performance exam, not a real clinic. The Jefferson Scale of Empathy is a questionnaire measuring a student’s self-reported attitude toward empathy — not warmth observed by a patient who is actually scared. Motivation and confidence were self-reported too. It is entirely possible to raise all of these numbers by teaching students to produce fluent, well-structured, rubric-friendly conversation — which is exactly what an AI trained on a scoring rubric would reward.

An AI virtual patient is a flight simulator for bedside manner: it builds real reflexes in a room where no one can be hurt — but it has never once been the frightened person in the chair.

The sample is also small and narrow: 60 students at a single university hospital, compared over a short three-month window, with no external validation in another school or country. And the comparison group matters enormously. The AI beat peer role-play — students practising on each other — not expert-led standardized-patient coaching, which is the real gold standard in communication training. Beating your classmates taking turns is a lower bar than beating a trained actor and a skilled supervisor.

Who checks the AI’s advice?

A virtual patient doesn’t just present symptoms; it judges the student’s performance and feeds back. But the trial measured what happened to the students, not whether the AI’s feedback was itself correct, balanced, or safe. If the system quietly rewards confident, polished phrasing over genuine listening — or nudges every interview toward the same template — students could score better while learning subtly worse habits. The rubric the AI optimizes toward is invisible to the learner, and the job of checking that feedback can’t be handed to the software; it stays with the human educators.

Who gets a tireless practice partner — and who doesn’t?

There’s a real equity story on both sides. Standardized-patient actors are expensive and scarce, so a scalable AI that lets any student rehearse hard conversations hundreds of times could be a gift to under-resourced schools — the authors rightly flag this. But it cuts the other way too: these systems need devices, licences, and connectivity, which can widen the gap between well-funded and struggling programs. And communication is deeply cultural: a virtual patient built and validated in one country, one language, and one set of social norms may quietly misread what counts as warmth or respect elsewhere. It’s a theme across the human side of dental AI — a tool tuned to one population rarely transfers cleanly to another.

Frequently asked questions