Local Milestone 0No authentication, database writes, or Voice control

Plain-language overview

Intended access: Anyone

How Machine Interview works

A practice patient that never gets tired, written and controlled by the teaching psychiatrist. This page explains who is involved, how a session comes together, and the rules that keep it safe.

The idea

Rehearse the hard interviews without a real patient

Learning to interview a psychiatric patient takes repetition, and real patients are not the place to practice the hard parts: asking someone guarded about thoughts of suicide, staying with silence, handling a patient who plays symptoms down. Actors are costly and scarce.

So the program director writes a practice case: a complete, invented patient with a history, a mental status exam, a way of speaking, and a list of what they will and will not reveal. The system turns it into a patient script: precise instructions that make an AI voice play that patient faithfully for one interview. Residents take turns interviewing it out loud in class while the group watches. Then everyone debriefs against the case's learning objective.

People and parts

Who is involved

Moderator

The program director, or a psychiatrist they name. Writes and approves practice cases, runs the session, and holds the only account that can save anything. One moderator per program.

Residents

Interview the simulated patient in turns. The system records nothing about them: no names, no scores, no performance notes.

The AI patient

ChatGPT's voice mode, given the patient script by hand. It has no connection to this website or its database. It only ever sees the invented patient.

This website

Where cases are written, checked, saved, and turned into scripts. It never talks to the AI during a live session.

In order

How a session comes together

  1. Write the practice case

    An eight-step form: the visit and the patient's story, the full mental status exam, how the patient talks, which facts the patient shares freely or only after trust is earned, how the patient reacts to empathy or interruption, and what the resident is meant to learn. The moderator confirms that nothing in it comes from a real person.

    Moderator · this website
  2. Save it

    The case, its learning objective, and the exact script it produces are stored together as one numbered version. A saved version never changes. Edits create the next version, so the record of what was used in class is permanent.

    Moderator · this website
  3. Run a pre-flight check

    Before a case meets residents, the system can quiz the AI in text about the patient and grade whether it kept to the facts. The check uses a question set written for that case. Question sets exist for the two built-in cases today; writing them for new cases is the next feature.

    Moderator · this website and the AI, text only
  4. Create and paste the patient script

    The moderator opens the saved case, copies the script, opens a fresh private chat in ChatGPT with memory off, pastes the script, and starts voice mode. The copy is done by hand on purpose: the website is never wired to the AI, so nothing from the classroom can flow back into it.

    Moderator · ChatGPT Voice
  5. Interview

    A resident interviews the patient for a set time, for example five minutes, while the group observes. The moderator ends the segment with a fixed spoken phrase the AI is told to obey. In a relay, several residents take consecutive turns with the same patient.

    Resident · moderator
  6. Reveal and debrief

    The website shows the full case to the group: what the patient was hiding, what the objective was, and what a good interview would have surfaced. Discussion follows. The moderator asks the AI for its short structured report and copies it out before closing the chat.

    Moderator · the group

Boundaries

The rules that make it safe

  • No real patients, ever.Cases must be invented or thoroughly composite. No names, initials, record numbers, exact dates, copied notes, recordings, or images. Every case requires the author's signed statement to that effect before it can be created.
  • No records about residents.The system stores no learner names, ratings, or feedback. That is a deliberate design boundary, not an oversight. Changing it needs an explicit program decision.
  • The AI is never connected to the classroom.The live session runs on a hand-pasted script in an isolated chat. The only automated contact with the AI is the text-only pre-flight check, run against an invented case with no resident present.
  • One person can publish.The moderator is the sole authority to save and approve cases. A database rule enforces that, not only the website.

Vocabulary

Words you will see

Practice case
The complete invented patient, written by the moderator.
Patient script
The instructions pasted into ChatGPT so it plays that patient exactly.
Learning objective
The one skill the interview is meant to practice, written in the moderator's own words and shown in the script and the reveal.
Mental status exam (MSE)
The 18 standard observations about a patient's appearance, speech, mood, thinking, and cognition. Each one is specified so the AI can perform it.
Sharing rule
A fact the patient volunteers, withholds, minimizes, denies, or reveals only after the interviewer earns it.
Pre-flight check
A text-only test that quizzes the AI about a case and reports whether it kept to the facts.
Question set
The written probes a pre-flight check uses for one case.
Reveal
Showing the group the full case after the interview.

Today

Where the project stands

Working now (alpha): sign in with an emailed link, write a case from blank, save it, reopen it, and produce its patient script. Two built-in cases, depression and subtle psychosis, also work in the pre-flight check.

Next: question sets for moderator-written cases so the pre-flight check works on them; limiting the moderator pages to signed-in users before anyone besides the moderator is given the address; then a structured debrief with feedback for the resident, which needs a program decision about what, if anything, may be kept about learners.

Ready to start? Sign in and write a new practice case.

Machine Interview is a private, clinician-governed tool for one residency program. It is not approved for use with real patients or real patient information.