A new case study from Australian National University and partners tested whether AI-generated “process maps” could capture clinical reasoning as reliably as human experts. The results were promising, offering a glimpse of how feedback in simulation-based learning could become faster and richer. 

When a nursing or optometry student works through a simulated clinical scenario, the toughest part to assess isn’t what they decide, it’s how they got there. Clinical reasoning happens largely inside someone’s head, which makes it extremely hard for tutors to observe, and harder still to give useful feedback on. 

Process mapping offers one solution. Developed by Gerry Corrigan in 2001, it turns interview transcripts into visual maps of a student’s reasoning pathway, making the invisible visible. The trouble is that building these maps by hand takes serious time and expertise, which limits how often it gets used in everyday teaching. 

A team from the Australian National University, Charles Sturt University, Deakin University and Stanford University set out to see whether artificial intelligence could shoulder some of that burden. Their case study, part of our 2026 Assessment and Feedback compendium, asked two questions: could an AI reliably spot the components of clinical reasoning in a transcript, and could it produce process maps that stood up against those made by human experts? 

To find out, the researchers generated three synthetic transcripts using GPT-4o, based on a simulated eye-health assessment. Both human experts and the AI then annotated the transcripts using an established codebook, before the AI converted its annotations into Mermaid.js diagrams, rendered as visual maps of each student’s reasoning. 

The results were encouraging, if uneven. Agreement between the AI’s repeated attempts, and between the AI and human coders, was strong for two of the three transcripts, with kappa scores in the “substantial” range. The third, notably the shortest and vaguest of the transcripts, fared far worse, dropping to only “fair” agreement. In other words, the AI did best with fuller, clearer material, and struggled when a student’s reasoning was less developed or more uncertain. 

Visually, the maps told a similar story. Side-by-side comparisons for the strongest transcript showed the AI capturing much the same structure and reasoning pathway as the human-generated version, even if the finer details didn’t always match exactly. 

The team is careful not to oversell the findings. This was a small, controlled study using synthetic data, designed to test feasibility rather than deliver a ready-made tool. They’re clear that implementing AI-generated process mapping still demands careful transcript preparation, prompt refinement and expert oversight, it’s not a case of switching on the AI and walking away. 

Even so, the implications are worth taking seriously. If AI can reliably support process mapping, it could open up a labour-intensive but valuable feedback method to far more students, more often, embedded into the debriefing that follows simulation-based learning. The authors are careful to frame this as augmentation rather than replacement: a way of helping educators see more, not a substitute for their judgement. 

As with so much AI-in-education research, the honest answer is ‘promising, with caveats’, but for anyone wrestling with how to give timely, meaningful feedback on reasoning that’s normally locked away in a student’s head, that’s still worth getting excited about.