The question has moved from whether to how, because the alternative to provision is not abstinence, it is the current reality: near-universal private use on consumer terms, unevenly distributed by budget, ungoverned by anyone, and invisible to the institution responsible for the learning environment. Provision is the lever that converts that reality into something equitable and governable, and it carries one serious risk worth engineering against from the start, the accuracy halo, the inference students reliably draw that approved means always right.
The case for provision, component by component
Four problems provision solves that policy alone cannot. Equity: frontier-model access increasingly tracks subscription tiers, and a cohort split between paid and free tiers is a cohort with unequal tutoring, which institutional licensing flattens at negotiated rates. Privacy and governance: enterprise agreements bring data-processing terms, retention controls and audit that consumer accounts never will, moving the amber band of the what-not-to-paste rules meaningfully: /blog/what-not-to-paste-into-ai-medical-school. Training surface: a provided platform is a teachable platform, the verification workflow, disclosure norms and appraisal exercises can be taught against a common tool rather than nineteen private setups. And visibility: institutions cannot govern what they cannot see, and provision creates the legitimate observation point that policy-writing currently lacks. Against these sits cost, real and recurring, and the halo risk, which deserves its own treatment because it is the failure mode that turns good governance into worse epistemics.
Engineering against the accuracy halo
The halo mechanism is simple: institutional endorsement transfers authority, and a tool the school pays for reads as a tool the school vouches for, claim by claim. Three design choices puncture it. Frame provision as environment, not oracle, explicitly and repeatedly: the account is provided the way the library is provided, as governed access, with accuracy remaining the user's verification problem, and the framing belongs in the rollout communications, not the terms nobody reads. Teach against the provided tool: the planted-error exercise, critiquing a deliberately flawed output from the school's own platform, is the single fastest halo-removal intervention, and it doubles as assessable appraisal skill per /blog/ai-proof-medical-assessments-wrong-goal. And publish the tool's known limits: jurisdiction coverage, content boundaries, the incident route for wrong outputs, because an institution that documents its provided tool's failure modes has modelled exactly the professional relationship with AI its graduates will need.
The portfolio model, and the decision questions
Single-tool provision buys simplicity and teaches a distortion, that one system covers medicine's tool landscape; the more honest configuration is a small portfolio mapped to functions: a general model for dialogue and drafting, a grounded evidence tool for clinical questions, a simulation environment for consultation rehearsal, and validated assessment resources for calibration, which mirrors the stack logic students should learn anyway, /blog/best-ai-stack-every-year-medical-school, and lets procurement apply the twelve-item rubric per function rather than seeking one impossible generalist: /blog/how-medical-schools-evaluate-ai-education-vendor. The decision questions for any faculty weighing it: which functions does our cohort already use privately, provision should meet demonstrated demand first; what does our governance actually require, data terms, audit, accessibility testing across accents and disabilities; who owns training and the halo-puncturing curriculum, because provision without teaching is subsidy, not education; and what is the exit plan, contract terms, data portability, the lock-in questions asked at institutional scale. Provision done this way is infrastructure; done as a headline procurement with a login email, it is the halo with a budget line.
Frequently asked questions
Should provision be mandatory to use?
No, and the distinction matters: provided means available and governed, not compulsory; students who prefer other tools within policy keep that freedom, and the abstainers' rights from the adoption debate apply here too.
What about specialist medical platforms, should schools buy those too?
Where a function justifies it, by the rubric: simulation and assessment resources are the commonest institutional buys; individual examination-preparation subscriptions sit more naturally with students, with equity handled through bursaries rather than blanket licensing.
How should schools handle tools students bring anyway?
With the same three moves minus the contract: taught verification, published guidance on categories rather than brands, and an incident route that accepts reports about any tool, because governance that only covers the provided platform governs a fraction of the behaviour.
Who should pay where provision is not affordable?
Then honesty beats theatre: publish the free-layer stack, negotiate student rates where possible, target bursaries at the genuinely differentiating paid components, and teach against free tools, which preserves the training surface even without the enterprise contract.
How should provision handle model updates and behaviour drift?
As a governed change, not ambient weather: version notes to students, re-run appraisal exercises after major updates, and the incident route open for behaviour regressions, the same update discipline the procurement rubric demands of vendors.