The AI-in-education question has moved from 'how do we ban it' to 'how do we teach with it'. The working answers: redesign assessment, set explicit AI-use policies per assignment, and protect student data like the regulated record it is.
Why: Blanket bans are unenforceable and blanket permission is meaningless; assignment-level clarity is the only version students perceive as fair — and follow.
Why: Generic prose is now free. Assessment that observes thinking (process, defense, application to local context) stays valid regardless of what tools exist.
Why: Detector false-positive rates are material and skew against non-native English writers — institutions have already faced formal complaints over detector-only discipline.
Why: Student records carry stronger protections than most commercial data, and the shadow-tool pattern (teacher pastes essays into a free tool for grading help) is the most common breach vector.
Why: Tutoring AIs can be genuinely effective, but they hallucinate in exactly the confident register novices can't detect. Transcript sampling is cheap; misconceptions at scale aren't.
Three years into generative AI, the institutions doing best made the same pivot: they stopped treating AI as a cheating problem to detect and started treating it as a literacy to teach and an assessment-design problem to solve. Students will use these tools for the rest of their careers; the educational question is whether they learn to use them well — verifying outputs, understanding failure modes, knowing when the tool undermines the learning goal itself.
The operational half is less philosophical: student data enjoys some of the strongest privacy protection anywhere, and casual AI tooling in schools keeps colliding with it. Procurement discipline and clear staff guidance prevent most of it.