如何打造敢於回答「我不知道」的 AI 醫學教育導師：Perceptors 的經驗

In most AI products a confident answer is a feature. In medical education it is a liability. A tutor that invents a dosing threshold or a guideline that does not exist teaches the wrong thing to someone who will act on it.

That constraint shaped how we built Perceptors, an applied AI platform for health education. Three design decisions did most of the work.

The source package comes first. Every course starts from an institution's approved material: guidelines, protocols, lectures, papers. The system maps each learning outcome to evidence in that package before anything is drafted. A claim with no source is flagged, not written around.
AI drafts, people approve. We split the work into specialist steps (evidence, curriculum, assessment, localization, quality) and put a named human reviewer at each stage gate. The system can move fast because it never publishes on its own. Reviewers edit drafts instead of writing from a blank page, which is where the time savings come from.
The learner guide is grounded and allowed to decline. The tutor answers only from the approved sources of the course, cites them, and declines or escalates questions outside them. Saying "this isn't covered in your course material, ask your faculty" turned out to be one of the most trusted behaviours in the product.

What we learned:

Retrieval quality matters more than model choice. Most bad answers trace back to a poorly chunked or outdated source, not the model.
Assessment items need rationales. Writing why each distractor is wrong catches more errors than reviewing the correct answer alone.
Translation is a governance problem, not only a language one. A translated module goes back through review like any other release.

Perceptors is now used by over 30 academic centres, including programmes in Saudi Arabia and the UAE. If you are building grounded AI for a regulated domain, I'd be glad to compare notes in the comments.

Fady Hannah-Shmouni, MD FRCPC, co-founder, Perceptors AI (perceptors.ai)