One of the questions I keep returning to in health professions education is:

If students are going to use AI in their future practice, should we allow them to use it during assessment?

The University of Sydney's approach caught my attention because, rather than treating AI as simply allowed or prohibited, it separates assessment into two different lanes.

Two lanes with different purposes

In the open lane, students can make appropriate use of contemporary tools, including generative AI. These assessments recognise that knowing how to work critically and responsibly with AI is increasingly becoming part of professional practice.

The secure lane, however, serves a different purpose. Here, students demonstrate what they can do independently under supervised conditions. This can include written examinations, oral assessments, practical examinations, OSCE-type activities and observed clinical performance.

I find this distinction particularly relevant to physiotherapy education.

Perhaps we do not need every assessment to prove that a student can work without AI. But somewhere in the programme, we still need to be confident that the student standing in front of a patient can reason, communicate, examine and make decisions independently.

AI may have an important role before assurance

At the same time, there is growing evidence that AI may have an important role before that point.

A 2026 randomised study involving 80 medical students compared conventional standardised-patient training with an approach combining standardised patients and AI virtual standardised patients. Students receiving the blended training demonstrated greater improvements in clinical reasoning, core competencies and OSCE performance in this study.

What I find most interesting is that the study does not necessarily make the case for replacing human simulation with AI.

It makes a case for blending them.

A possible educational pathway

AI virtual patients could provide students with repeated, accessible opportunities to practise clinical reasoning, make mistakes, explore alternative decisions and receive feedback. Human simulation and clinical encounters can then introduce the communication, unpredictability, emotional complexity and interpersonal judgement that are much harder to replicate.

Finally, secure assessment can answer the most important question: Can the student actually do this when the AI is no longer there?

Learn with AI → practise with AI → practise with people → demonstrate independently → enter clinical practice.

Start with what the assessment must prove

For me, this changes the AI assessment conversation.

Rather than asking whether AI should be permitted or banned, perhaps we should first ask what each assessment is supposed to prove.

If the purpose is learning, experimentation, feedback and development, AI may have considerable value.

If the purpose is to assure ourselves that a future healthcare professional can perform safely and independently, then there must still be moments when the student, not the AI, has to demonstrate that capability.

The challenge now is deciding where those moments should sit within the curriculum.

Perhaps that is where programme-level assessment design becomes much more important than trying to redesign every individual assignment in response to AI.

Related reading