AI doctor vs AI patient: which one makes you better?

What an AI doctor does

You describe a problem and it reasons for you. Symptom checkers, health chatbots and diagnostic assistants all work this way, and the good ones are useful: the answers in the demo follow NHS guidance on coughs.¹ For a patient, that’s the point. For a medical student, it’s a lookup tool. You read the reasoning; you don’t practice it.

The research versions are getting strong. In a randomized, double-blind study, Google’s AMIE took histories from patient actors over text chat across 159 scenarios and went up against 20 primary care physicians. It was more accurate on diagnosis, specialists rated it better on 30 of 32 axes, and the patient actors rated it better on 25 of 26.² The study ran on text chat, which isn’t how most consultations happen, so read it as a direction of travel, not a replacement for doctors.

Even the AI doctor practiced on AI patients

AMIE’s training didn’t stop at medical text. It included a simulated environment where AI agents played the patient, so the model could run consultation after consultation and get feedback on each.² Reps against a simulated patient, with feedback after each one, are the recipe that works for people too.

What an AI patient does

An AI patient flips the roles. It has a history you can’t see, a personality, and a reason for being in the room, and it only tells you what you ask for. Ask a closed question too early and you get a yes. Lecture it and it gets defensive. Afterwards you’re marked on what you did, not on what you know.

On My AI Patient you pick a case and a persona, talk by voice or text, and get feedback against a 48-point rubric, from introducing yourself to checking the patient understood the plan.

Does practicing with simulated patients work?

Simulation in general does. A 2011 meta-analysis of 609 studies and more than 35,000 trainees found that tech-based simulation training had large effects on knowledge, skills and behaviors compared with no training, and moderate effects on patient outcomes.³ Chatbot patients arrived a decade later, so none of them were in it.

Computer-based virtual patients, the scripted ancestors of today’s AI patients, have their own meta-analysis: large gains compared with no training, and small gains on average compared with other teaching.⁴ The honest read is that practice beats no practice by a lot and beats good teaching by a little. Studies of patients built on large language models are early, so treat them as extra reps on top of placements.

AI patient vs standardized patient

Standardized patients are trained actors who play a case and give you feedback, and they’re a fixture of objective structured clinical examinations (OSCEs). A good one is worth a lot. They also cost money and staff time to schedule, so sessions are rationed. Here’s how the two compare:

• Availability: an AI patient is there at 2 a.m. the night before your exam; an actor needs a booking.
• Repetition: you can retry the same case five times in an evening and compare the feedback.
• Difficulty: you pick the persona, from cooperative to openly hostile.
• Where actors still win: physical examination, body language, real human reactions and the feel of an exam room.

Use both. Actors for exam day and the physical exam, AI patients for the reps in between, and an AI doctor afterwards to check what you missed.

Practice before it counts

The patient in the demo is waiting in the room for real: same face, same chest pain, no script. He answers by voice or text, he still doesn’t trust doctors, and when you’re done you’ll see which of the 48 criteria you hit.