Every other setting asks “how do I use this tool well?” Clinical education asks a harder question: how does someone develop clinical judgment in the presence of a machine that simulates its products? A student can now produce a fluent SOAP note, a plausible goal, a polished reflection — without the reasoning those artifacts are supposed to evidence. Policing that is a losing game. The winning move is making AI output an object of instruction: something students learn to read critically, because reading LLM writing is now a clinical skill they’ll need for the rest of their careers.

The Practice Realities

• Supervision minimums still apply. ASHA certification standards require direct observation of at least 25% of each student’s client contact. No AI-assisted efficiency changes what must be directly observed and formatively assessed.
• The artifact is no longer evidence of the reasoning. Before LLMs, a well-structured eval report implied the student could structure clinical thinking. That inference is dead. Assessment has to move to where the reasoning is visible: live sessions, oral defense of documentation choices, “walk me through why this goal has this criterion.”
• CFs are building a clinical voice. A Clinical Fellow who drafts everything through a model risks fossilizing at “competent-sounding generic.” The mentoring conversation isn’t “did you use AI” but “show me what you changed and why” — the edit is the clinical thinking.
• Your syllabus and clinic policy need an actual position. “Don’t use AI” is unenforceable and teaches nothing. A workable policy names permitted uses (organizing, structure), prohibited uses (generating clinical interpretations, fabricating data), and a disclosure habit. The clinical education policy template is adaptable.
• You’re modeling, whether you mean to or not. Supervisors who use LLMs for feedback drafts and disclose it — showing their edits — teach responsible use better than any policy document.

Worked Example: The Critique Assignment

Have the model generate the artifact, then grade the student’s reading of it:

“Generate a treatment goal for a 7-year-old with a phonological disorder in a school setting.”

Then the student evaluates the output against the four questions: What are the conditions? Is the behavior observable? Is the criterion defensible or a round number? How would you actually measure it? What did the model not know about this child that a real goal requires? The deliverable is the critique, not the goal. Students who can dismantle a plausible-but-hollow goal have demonstrated exactly the judgment the artifact used to imply — and they’ve practiced the reading skill on a target that doesn’t mind being wrong.

Red Flags Specific to This Setting

• Grading artifacts that no longer evidence reasoning, and not noticing the inference broke
• Students using AI to skip the struggle that builds judgment — fluent goals they can’t defend orally
• CFs whose documentation voice is indistinguishable from a model’s default register
• Supervisors privately using AI for feedback while publicly banning it — the hidden-use model students actually learn from

Related

• Workflow: CFY Supervision Feedback
• Prompt: Ethical Reflection
• Reading: the antipatterns gallery doubles as a ready-made critique curriculum

See it in practice: Case study — grad student supervision.