In 2017, the Royal College of Physicians and Surgeons of Canada launched Competence by Design, the most consequential reform of Canadian residency training in a generation.
The premise was simple: we can define all the elements that a graduating physician should know or do in practice, and we should only graduate them when a supervisor has actually watched them do the work competently and safely. The framework broke training into discrete clinical tasks, such as managing a patient with chest pain, that a trainee must be observed performing before a supervisor signs off.
Nearly a decade later, the architecture is in place. The observation is not.
I trained as a specialist in this country. In the entirety of my residency, consisting of thousands of patient encounters, day and night, in clinics, on wards and in emergency departments, fewer than 10 were formally observed by a supervising physician. I completed my training despite rarely being directly assessed on my performance. That remains true for many residents today. We built a competency framework on a foundation of observation that we never did.
This is not because supervising doctors are failing. The attending physician supervising a resident may also be running between consults, fielding pages from the operating room and covering overnight call. We ask supervisors to produce high-volume, structured, real-time observational assessments, while rushed physicians often only have time to review, not observe, their trainees. We designed a system that depends on observation, without making observation feasible.
AI may be the solution to address these gaps in medical training
At Scarborough Health Network (SHN), where I work, we have worked to close this observation gap. We are a community teaching hospital caring for 850,000 people in the Greater Toronto Area, with hundreds of learners rotating through our care settings each year.
In building and testing Artificial Intelligence tools for observation, feedback and simulation, our aim is to create a made-in-Scarborough solution for teaching hospitals across the country.
In Canadian primary care today, AI already is quietly being used to transcribe visits and produce draft notes during encounters between physicians and patients. With patient consent and clear governance, every student- or resident-patient interaction can be recorded and transcribed. Alongside a supervising physician, an AI system can review these encounters against competency frameworks and relay specific feedback directly to the medical learner. Instead of less than 10 observed encounters across a career, every encounter can become a teachable one.
AI can determine where our incoming physicians still need further training. For example, it can audit a resident’s case logs and reveal trends of deficiencies that may not be initially apparent. A resident may have seen hundreds of patients but may never have managed gastrointestinal bleeding or have never had a goals-of-care discussion with a critically ill patient. We have the data but unfortunately are not using it to ensure competence.
Virtual standardized patients offer another opportunity. These AI-driven programs, which we are validating at SHN, use video avatars and voice models to simulate clinical scenarios, similar to actors currently used in medical training. They allow trainees unlimited repetitions on the hardest parts of the job, such as breaking bad news, navigating goals-of-care discussions or explaining discharge instructions to frightened patients and families before they ever face a real patient in distress. As in other high-stakes professions, including aviation, simulation can be used to build mastery before errors carry real-world consequences.
SHN is also testing a performance assessment system that uses AI for medical training, with unobtrusive audio and video to evaluate student performance. The goal is consistency: creating a more objective view of how a trainee actually performs over time rather than relying on fragmented, memory-based evaluations. Together, these approaches move training from occasional observation to continuous, structured assessment: capturing what happened, identifying what is missing, and allowing trainees to practise until they improve.
None of this is risk-free, and the guardrails matter. Ambient capture demands explicit, granular patient consent and full compliance with provincial privacy law, such as the Personal Health Information Protection Act (PHIPA) in Ontario. Faculty must remain firmly human-in-the-loop, with AI augmenting – never replacing – professional judgment and teaching. Models must be audited for bias, particularly in how they assess trainees from underrepresented backgrounds.
The argument is not that AI replaces the preceptor. It is that AI finally gives the preceptor something to teach: a faithful record of what actually happened in the room, available when there is time to review it together. Imagine an encounter at 11 p.m. when a resident sees a patient who did not understand instructions for discharge from the emergency department. The next morning, the resident sits with the supervisor, and this time they review the patient-doctor conversation. The AI points to moments where confirmation of understanding did not occur. The supervisor and resident rehearse it. The following week, using a virtual patient simulator alongside real-life demonstrators, the resident practices discharge conversations repeatedly until those techniques become routine.
We promised Canadians we would watch their future doctors learn. For decades, we haven’t. AI won’t replace preceptors, but it may finally help us keep that promise.
