Every serious survey tells the same asymmetric story: students use generative AI at supermajority rates, weekly for half of them, while the educators responsible for teaching, assessing and supervising them span the full range from fluent adopters to principled abstainers to the quietly unconfident, with policies varying by module because they vary by person. The asymmetry is the binding constraint on everything this series discusses: AI-literacy curricula, redesigned assessment, sensible policy and honest supervision all route through faculty capability, and a system that trains students while leaving educators to improvise has built the second storey first.
Why the bottleneck binds so tightly
Three transmission mechanisms. Assessment design: the capability-verification menu, supervised transfer, oral defence, reasoning traces, flawed-output critique, /blog/ai-proof-medical-assessments-wrong-goal, is faculty craft; unconfident educators default to prohibition-and-detection, which fails technically and corrodes culturally, and the failure gets attributed to students. Policy coherence: students navigating six modules meet six AI policies where faculty development is uneven, and the incoherence teaches exactly the wrong lesson, that the rules are personal weather rather than professional principle. And supervision modelling: students learn professional AI behaviour by watching seniors verify, disclose and challenge outputs, or watching them not; an educator who cannot appraise a generated differential cannot teach the appraisal, and cannot credibly assess it. The student-facing curriculum, in other words, is downstream of an educator curriculum most institutions have not built.
The five-component faculty curriculum
Deliberately parallel to the student version, because the capabilities rhyme, with the emphasis shifted to the educator's jobs. Output appraisal: hands-on practice reading AI-generated clinical and educational content critically, including the planted-error exercise run on themselves first, since appraisal confidence is built by catching errors, not by hearing they exist. Assessment redesign: the verification menu as workshop material, each educator converting one of their own assessments, which turns abstract policy into owned practice. Privacy and governance: the never-paste framework in educator form, where the stakes include student data, examination security and institutional material, plus the local escalation routes by name. Bias and equity: how model performance varies across accents, languages and demographics, and what that means for feedback, assessment and the students most exposed. And educational prompting and supervision: not prompt tricks but the pedagogy layer, how to use AI as the challenger in teaching sessions, how to supervise student AI use, and how to disclose their own, the modelling function made explicit. Faculty-facing infrastructure is emerging to support exactly this, ScholarRx's TAI direction being the visible institutional example, tools built for educators authoring and governing content rather than students consuming it, and the distinction between the two audiences is the whole design point.
Making it happen against real constraints
Faculty development competes with clinical pressure and curriculum overload, so the implementation has to be honest about cost. Three moves that fit reality: anchor it in existing structures, the education-day slot, the assessment-board cycle, rather than new committees; make it artefact-producing, each session ends with a converted assessment, a written module policy, a tested tool verdict, so development compounds into institutional assets; and pair confidence unevenly on purpose, the fluent early adopters running sessions for peers, which spends the existing capability where it exists. The interview layer this topic deserves, educators across the UK, US, Canada and Australia comparing what actually worked, is a follow-up worth building when those conversations can be gathered; the argument stands without it: the student AI story is written in faculty rooms, and the systems investing there first will find every downstream problem in this series smaller.
Frequently asked questions
Isn't it faster to centralise policy than train everyone?
Central policy without local capability produces compliant paperwork and unchanged classrooms; the curriculum exists so policy has hands, and the artefact-producing design keeps the training from being another document.
What about faculty who remain sceptics after training?
Trained scepticism is an asset: the appraisal component works identically for critics, and a faculty containing confident adopters and confident sceptics models exactly the professional pluralism students should see, versus the current mix of confidence and avoidance.
How should success be measured?
By the transmission mechanisms: policy coherence across modules, assessment portfolios shifting toward the verification menu, and student-reported consistency of supervision, all measurable annually with instruments schools already run.
Where should a school start if it can fund only one session?
Output appraisal with the planted-error exercise: it is the component every other one builds on, it converts scepticism and enthusiasm alike into calibrated confidence, and it produces the fastest visible change in how educators talk about AI in their own teaching.
Should students be involved in faculty development?
Deliberately: student co-teaching of real workflows grounds the sessions in actual behaviour, and the reversal, students as the fluent party, models the bidirectional professionalism the whole transition needs.