Real actors, real voices: what actually made our AI patients convincing

Most conversational AI avatars have a tell. The face is a little too smooth, the voice is a little too eager, and within about ten seconds you know you're talking to software. We built three of them for a client, an international medical devices company, for a congress stand in Auckland this November, and the one thing we were determined to avoid was that tell.

The brief was patient history-taking practice for surgeons: three AI patients, built from real patient profiles, sat in private consult rooms on the stand and waited to be asked about their lives. The twist wasn't the AI. It was what we built the AI out of.

We cast and filmed real actors rather than generating synthetic faces. A patient who looks like someone you'd actually see walk into clinic reads completely differently to an obviously rendered avatar, no uncanny valley, no stiffness, just a person. And we used the actors' own voices, recorded properly, in a real Australian accent, not a text-to-speech voice defaulting to the overly warm American tone most people now associate with AI. It's a small production decision on paper. In the room, it was the difference between a patient who sounded like someone from down the road and one who sounded like a demo.

It worked. Across three show days, surgeons ran 43 real consults with a median length of five minutes thirty-eight seconds, well past the point anyone sits through out of politeness. More than one surgeon, unprompted, said some version of “I thought this was real.” One asked if the hospital could use it for training.

There's a second layer to this worth a mention: the brief also carried a strict medical-privacy standard, nothing recorded, nothing kept beyond room, patient and length. Meeting a bar that strict usually costs you realism, you scale back to make the privacy problem go away. This one didn't have to.

The full story is in our case study: Meet the Patient.