The Virtual Role-Play That Excluded Half the Cohort

Four issues are entangled: privacy and consent, disability accommodation, equity of access, and validity of the AI feedback. Privacy comes first because recordings of role-plays may contain simulated client disclosure and because students were not told about 24-month retention or model training. If the institution is subject to FERPA, recordings used for grading are education records, and the vendor is a school official only if the contract controls use and re-disclosure; the current agreement may not. The Deaf student's request is a civil-rights issue, not a preference: a text-based alternative must provide an equivalent opportunity to demonstrate the interviewing competency, and an audio-dependent AI score cannot be the sole basis for grading. Rural bandwidth is an equity issue that also distorts the data faculty are using to judge the tool. Faculty workload relief does not outweigh these concerns. Priority order: stop nonconsensual retention and model training, provide an equivalent accommodated pathway, reduce bandwidth dependence, then test whether the AI feedback is valid. The program should not renew on the current terms simply because the pilot is convenient.

Current term: pause new required uploads until students receive a plain-language notice and a consent choice; allow local recording with phone or offline upload, or a synchronous low-bandwidth option. For the Deaf student, provide a live text-based role-play with a standardized client and human scoring on the same rubric, not a lesser task. Re-score any AI-dependent grades with human review. Ask the vendor to disable model training and shorten retention, or move recordings to institutional storage; if the vendor refuses, do not renew. Renewal decision: run a four-week options appraisal comparing a renegotiated vendor, an institutional video tool plus human scoring, and a hybrid low-bandwidth model. Require accessibility conformance documentation and a data-protection review. Cost the faculty time for human scoring, because that is the hidden subsidy. Pilot with the two rural counties and the Deaf student's pathway, and collect completion, bandwidth, accommodation, and learning data disaggregated by group. Faculty who value time savings should see the trade-off explicitly. The final decision should be conditional and documented.

Evaluate learning, not engagement. Pre-register outcomes: rubric scores on interviewing skills by human raters blind to condition, transfer to a live role-play, and student self-efficacy. Compare AI feedback against human feedback on the same recordings. Check whether AI scores correlate with human scores and whether they are biased by accent, dialect, disability, camera quality, or cultural display rules, since eye-contact norms vary. Disaggregate by rurality, disability, and language. Look for score inflation and gaming. A tool that speeds grading but narrows the construct to speech rate and eye contact may teach students to perform for the algorithm rather than attune to clients. Use a small randomized or quasi-experimental design if feasible; otherwise compare matched samples. Report negative findings, too. Renew only if there is evidence of equivalent or better learning with no disparate impact, and if the privacy and accessibility conditions are met.