Clinical Reasoning in the AI era: A focused scoping review and narrative synthesis of conceptual, pedagogical, and assessment frameworks in health professions education

The Big Picture: The "GPS" Problem

Imagine you are teaching someone how to drive a car. For years, you've had a clear map of what "good driving" looks like: knowing the rules, reading the road, handling a sudden storm, and knowing when to trust your gut versus the dashboard.

Now, imagine a new Super-GPS (Artificial Intelligence) has been installed in every car. It can tell you exactly where to turn, predict traffic, and even write your driving log for you.

The problem this paper identifies is this: We are using the GPS to grade the driver, but we aren't sure if the GPS is actually teaching them how to drive.

The author, Naomi Campbell-Woods, looked at hundreds of studies to see how medical schools are using this "Super-GPS" to teach students how to think like doctors (a skill called Clinical Reasoning). She found that while the technology is impressive, we are mostly testing the parts of driving that are easy for the GPS to measure, while ignoring the messy, human parts that make a doctor truly safe.

---

Part 1: The Confusing Map (The "What")

Before we even talk about AI, the paper points out that medical education has been struggling to agree on what "Clinical Reasoning" actually is. It's like trying to teach someone to be a chef without agreeing on what a "good meal" is.

The paper says there are three main ways experts currently describe "good thinking":

1. The Library Card System (Script Theory): Doctors organize their knowledge like a library. When they see a patient, they pull out a "script" (a mental file) about that illness.
2. The Fast vs. Slow Brain (Dual-Process): Good doctors use two modes: a fast "gut feeling" mode and a slow, careful "check the facts" mode.
3. The Team Sport (Situated Cognition): Thinking doesn't happen in a vacuum. It happens in a messy hospital room with a scared patient, a tired nurse, and time pressure. You can't separate the thinking from the environment.

The Issue: For a long time, schools tried to test this thinking with a single, final exam. The paper says this is like judging a whole season of a sports team based on one play. It doesn't work.

---

Part 2: The AI Shortcut (The "How")

Now, enter the AI. The paper reviewed 158 studies where AI was used to teach or test medical students.

What the AI is doing well:
The AI is great at grading things that are neat, written down, and structured.

• Analogy: Imagine the AI is a very strict librarian. It can instantly check if you wrote down the right symptoms, if your diagnosis matches the textbook, and if your notes are formatted correctly.
• The Result: Studies show students get better at these specific tasks when they use AI. They gather data faster and write better notes.

What the AI is missing:
The AI is terrible at measuring the messy, human, "in-the-moment" parts of thinking.

• Analogy: The librarian can check your spelling, but they can't tell if you were brave enough to ask a difficult question to a patient, or if you knew when to ignore the GPS because the road was flooded.
• The Gap: The paper found that AI studies rarely measure:
  • How a student handles uncertainty (when the answer isn't clear).
  • How they adapt when the situation changes.
  • Whether they are just blindly trusting the AI or actually thinking for themselves.

---

Part 3: The Danger of "De-skilling"

The paper warns of a specific risk called "De-skilling."

• Analogy: If you let the GPS drive the car for you every day, you might forget how to read a map or steer in a snowstorm.
• The Risk: If students rely on AI to do the hard thinking (like figuring out a diagnosis) too early, they might never learn to do it themselves. They might become "never-skilled" (never learning the skill) or "mis-skilled" (learning to do it the wrong way because they copied the AI).

The paper notes that when students moved from a computer simulation (where AI helped them) to real patients, the AI's help often disappeared. The students didn't perform better in the real world, suggesting the AI was just a crutch, not a teacher.

---

Part 4: The Proposed Solution (The New Framework)

The author argues we need a new "Rulebook" for the AI era. We can't just keep using old tests with new tools.

She proposes a Six-Layer Cake for teaching and testing reasoning in the age of AI:

1. Knowledge Organization: Do they have their mental library files in order?
2. Process Enactment: Do they know how to move from a clue to a conclusion?
3. Uncertainty Handling: Can they stay calm when the answer isn't clear?
4. Contextual Performance: Can they think well in a noisy, busy hospital, not just in a quiet test room?
5. Sociotechnical Judgment: This is the new one. Can they use the AI as a tool without letting it take over? Can they say, "The AI says X, but I think Y because..."?
6. Programmatic Architecture: Instead of one big test, we need a long-term system that gathers many small pieces of evidence (like a portfolio) to judge a student fairly.

The Bottom Line

The paper concludes that AI hasn't broken medical education, but it has exposed a crack in the foundation.

We have great tools to measure the "easy" parts of thinking (like writing a note or picking a diagnosis), but we are failing to measure the "hard" parts (like handling uncertainty or knowing when to ignore the computer).

The takeaway: We shouldn't just build more AI trials. We need to build a new framework that teaches students how to be the captain of the ship, using the AI as a helpful navigator, rather than letting the AI steer the boat while they sleep. If we don't fix this, we might end up with doctors who are great at using computers but terrible at thinking for themselves.