Every new teaching technology inherits two conversations: “is it safe?” and “is it fair?” AI role play raises both, and the honest answers are better than the anxious ones — provided you handle the design work this chapter covers.
Disclosure: role versus self
The oldest safety advice in role-play pedagogy is also the most relevant here: separate role from personal belief. The NIU guide, Harvard’s facilitation notes, and Duke Kunshan’s ground rules all converge on the same move — before any role play, discuss the difference between what the student says in character and what they believe, and establish that the classroom (or the AI session) is a space where students can inhabit unfamiliar positions without being held to them.
For AI role play, add one explicit disclosure: the counterpart is an AI. This sounds obvious, and it is, but it needs to be in the syllabus, not assumed. Students should know they are talking to a system, that the session is recorded and transcribed, who has access to those recordings, and for what purposes (feedback, grading, program improvement, research). Write it as a paragraph in the syllabus — the exercise at the end of this chapter produces it.
Do not bury the disclosure in a terms-of-service link. Students who discover mid-semester that their therapy-practice sessions were stored somewhere tend to feel surveilled, and that feeling is a trust collapse you do not recover from.
The stress question
The NYU oral-exam data is worth sitting with: 83% of students found the AI viva more stressful than written exams. And 83% had never taken any oral examination. The stress is real, but most of it is unfamiliarity, and unfamiliarity is the thing this playbook is built to fix.
The mitigation stack, in order of effectiveness:
Publish the structure. When the exam format, the rubric, and the persona are all available in advance, anxiety drops from “what will happen?” to “can I do the thing?”, which is the productive kind of stress.
Provide unlimited practice access. The same examiner or persona available as a practice partner for weeks before the graded attempt. A student who has talked to the persona ten times arrives at the assessment having already survived the worst it can do. This is the single largest stress reduction available, and it costs nothing beyond the scenario existing.
Use the stakes ladder. Chapter 15’s rung 1-2-3 progression exists for this reason: students encounter the format low-stakes before they encounter it high-stakes. The first graded role play should never be the first role play.
Normalize the difficulty. Tell students the persona is designed to push back, that being pushed back on is the point, and that everyone gets the same pushback. When students know difficulty is by design, they experience it as challenge rather than personal attack.
Offer the exit. For scenarios touching emotional content — grief, abuse, trauma, crisis — a student should be able to stop the session at any time without academic penalty. The formative window especially should be genuinely low-pressure: a student who quits a practice attempt and returns tomorrow has done exactly what the program wants.
Accessibility and accommodation
AI role play has accommodation properties traditional role play lacks, and limitations it inherits:
What improves. Scheduling flexibility (students with health conditions, caregiving responsibilities, or access needs practice when their condition permits, not when the facilitator is available). Repeatability (a student with a processing disorder can run the same scenario at their own pace without consuming more human resource). Written transcripts as a complement to the spoken encounter.
What needs design. Students with speech or hearing impairments need text-mode alternatives. Students with anxiety disorders may need extended time limits, lower difficulty settings, or the option to submit a text-based walkthrough instead of a voice session. Accommodation decisions should live with disability services, not with the platform default — which means the instructor must know the available configuration and communicate it.
One operational note: if the platform produces per-student analytics visible to the instructor, those analytics include attempt counts. A student with a disability who requires twenty attempts to reach the score a neurotypical student reaches in five should never have that attempt count held against them; the design explicitly says unlimited practice is the feature, not a signal.
FERPA-shaped caution
Universities in the United States operate under FERPA, and role-play recordings are education records. The practical implications: recordings should be stored on a platform with an appropriate data-processing agreement; access should be limited to the instructor, authorized TAs, and the student; anonymized excerpts used for the classroom debrief (Chapter 13) should be genuinely anonymized (remove names from transcripts, not just from video); research use requires IRB review and consent. Programs outside the US have analogous frameworks — GDPR in Europe, institutional data policies elsewhere — and the principle is the same: treat recordings as you would exam scripts, because that is what they are.
The equity upside
The safety conversation usually stops at risk mitigation, but the equity case for AI role play is strong enough to state affirmatively.
Consistent quality. Every student gets the same quality counterpart. In actor-based programs, the 9 a.m. actor and the 4 p.m. actor are not the same person having the same day. In peer role play, the student paired with the most committed partner and the student paired with the least get different educations. AI deletes both sources of variance.
No cast-by-accent effects. A standardized-patient program that relies on a small actor pool unavoidably gives students a skewed demographic sample; the bias is structural, not intentional. AI personas can be configured for any demographic presentation without a casting constraint.
Access parity for off-campus students. Commuter students, online learners, and students at satellite campuses have historically received less practice time because they are not in the building when the facilitator is. Asynchronous AI practice is the first modality where geography doesn’t determine dosage.
Financial parity. Programs that charge lab fees for simulation sessions or require travel to a simulation center impose costs that correlate with socioeconomic status. A browser-based practice scenario imposes the cost of the laptop the student already has.
None of this eliminates the need for the risk-mitigation work above. But it does mean the status quo — peer role play with uneven partners, occasional access to actors for some students, no recorded practice for most — is not the equity baseline. It is the equity problem AI role play can fix.
Exercise
Write the syllabus paragraph that discloses and frames the AI role play in your course. It should cover: (1) the counterpart is AI, (2) sessions are recorded and transcribed, (3) who has access and for what purposes, (4) the format is practice-first — unlimited ungraded attempts before any graded one, (5) the persona pushes back by design and this is normal, (6) how to request accommodations. Read it aloud and notice whether it sounds like a warning or an invitation. Rewrite until it sounds like the second.