A medical student can know the right diagnosis and still be a bad interviewer.

They might ask the wrong questions, miss an important symptom, interrupt a patient, or fail to build enough trust for the patient to actually share what’s going on. Those skills are difficult to teach from a textbook.

Traditionally, medical schools have addressed this through the use of “simulated patients”, usually actors who are trained to mimic specific patient conditions. Though this practice has long been in place since it was introduced in 1963 at the University of Southern California ¹, it has several shortcomings:

1. Students have limited access to standard patients - a university may book actors for a short window of time, so practices must be done synchronously and typically are in-person.

2. Feedback is limited - typically, feedback is not given by the patients but by a supervising attendee; however, the supervisor’s time is limited and expensive, so they are only able to supervise one student at a time and for only a short duration, which limits feedback.

3. Not all medical students can access this training - especially in a global context, not every university has the financial or technical resources to hire and train actors to portray these conditions in a “standarized way”, or likewise cannot afford to dedicate a specialist professor’s time to supervising and giving feedback to these mock encounters.

Through the DALI Lab at Dartmouth College, I was placed on a project that sought to address these issues through technology. The self-explanatory “AI Patient Actor” project, developed in collaboration with the NILE Lab at the Geisel School of Medicine, seeks to use artificial intelligence to reliably automate the “Simulated Patient” at scale.

Led by Dr. Thomas Thesen, the project originated started in 2023, when he brought the original idea to the Research Computing department at Dartmouth Library, which built the “v1” of the app, building a wrapper and prompt engineering a ChatGPT model to simulate various patient conditions. ²

I joined the project during the winter of 2026, as the construction of “v2” of the app Through the platform, a student could interview the patient, ask questions, request diagnostic tests, work toward a diagnosis, and then receive feedback on the encounter. My work on v2 was full-stack, spanning both the interview/feedback frontend and the backend logic connecting the two.

Here’s a short demo of the platform:

The project worked. As of May 2026, AI Patient Actor has achieved remarkable stats:

3,000+Students enrolled	250+Educators
9,000+Clinical encounters	200+Active institutions
45+Countries represented	20+Languages used in practice
4,000+AI feedback reports delivered	300,000+Conversation messages exchanged

The Technical Implementation

I don’t own the intellectual property behind the platform, so I cannot get into specifics of the implementation - the case-generation prompts, the models used, or the internal evaluation logic all belong to the project. At a high level, though, my work on v2 touched the full stack:

• on the frontend, building out the interview and feedback flows students interact with directly
• on the backend, the logic that routes a student’s questions to the simulated patient, keeps a case consistent across a conversation, and turns a finished encounter into structured feedback.

I also worked on infrastructure pieces like credential management, building proper role-based access so students, educators, and administrators across institutions could be provisioned securely, and on early work toward prosody analysis.

Prosody, broadly, is the pattern of rhythm, stress, and intonation in speech - the parts of talking that live outside the literal words: pace, pitch, pauses, loudness, hesitation. In a clinical encounter, a lot of what a good doctor picks up on isn’t in the transcript. A patient who answers in a flat m onotone may be depressed. A long pause before answering a question about pain could mean they’re downplaying it, or unsure how to describe it. Rising pitch and speeding up can signal anxiety a patient hasn’t put into words.

A text-only evaluation of a student’s interview misses all of that, since it can only judge the questions they asked and the answers they got, not how the exchange actually felt. Voice-based prosody analysis works with audio directly rather than a transcript, which opens the door to assessing things like whether a student picked up on a patient’s distress, whether their own tone came across as rushed or dismissive, and whether they gave a patient the physical and conversational space to answer fully instead of talking over pauses that mattered.

Key Takeaways

Historically, Dartmouth’s connection to AI goes back to the beginning of the field itself - the term “artificial intelligence” was coined here in 1956, at the Dartmouth Summer Research Project on Artificial Intelligence.³ Getting to work on a project like this through DALI Lab, nearly seventy years later and in the same place that gave the field its name, was something I didn’t take for granted.

I’m also incredibly excited that the work has also gotten recognition outside of Dartmouth. Thesen and the partners published a paper in Medical Science Educator, ⁴ and the project was featured by OpenAI as part of its Professors Reshaping Pedagogy project. ⁵ Given the reach the platform already has - thousands of students, hundreds of institutions, dozens of countries - it’s a project with real potential to change how medical training happens at scale, not just at Dartmouth.

There aren’t many chances, this early on, to work on something that touches this many people’s lives in such a direct way. Somewhere out there, a student who once practiced a hard conversation with an AI patient is now having that same conversation with a real one - and doing it a little better because of work I got to be part of. That’s something I’m incredibly proud of, and that motivates me to keep chasing work that actually matters.

---

Footnotes

1. In 1963, USC professor Howard Barrows “started to systematically use healthy actors to simulate patient’s signs and symptoms, in order to teach and assess his students. The standardized patient was born, an umbrella term for situations where a person is trained to simulate a clinical case or an actual patient is trained to present his or her illness in a standardized way.” https://journal.ppcr.org/index.php/ppcrjournal/article/view/12 ↩

2. https://home.dartmouth.edu/news/2024/01/geisel-professor-harnesses-ai-act-patient ↩

3. https://ai.dartmouth.edu/our-story ↩

4. https://link.springer.com/article/10.1007/s40670-024-02250-2 ↩

5. https://academy.openai.com/public/videos/thomas-thesen-medicine-2025-08-21 ↩