U.S. states — Education agencies are issuing divergent back-to-school policies on generative AI, from require-citation frameworks to stricter classroom device filters. Federal context and program pages live on ed.gov; statistical baselines on students, schools and technology access are published by NCES.

Teacher unions have asked for training time and clearer academic-integrity standards before enforcement, according to public statements from state affiliate leaders this summer. Ed-tech vendors ship dashboards that flag unattributed AI text — tools privacy advocates said need student-data guardrails and appeal processes before districts mandate them.

Districts that piloted AI tutors reported mixed outcomes, administrators said: gains in drafting support, uneven results for math concept mastery, and new questions about assessment design when take-home work can be machine-assisted. Looking ahead, mid-year board reviews will matter more than fall launch memos.

Parents split between enthusiasm for personalized help and concern about cheating and data collection. School boards will spend fall meetings reconciling those pressures with limited budgets for software and professional development.

Whatever policy a state chooses, implementation quality matters more than the press release. A citation rule without teacher training becomes theater; a ban without detection capacity becomes uneven and unfair.

NCES data on broadband and device access remain essential for equity analysis: AI classroom rules land differently in districts with one-to-one device programs versus shared lab carts.

State education agency PDFs should be linked whenever a specific state rule is discussed so readers can verify effective dates, scope (K–12 vs. higher ed) and any sunset clauses. Federal ed.gov and NCES pages provide the national program and statistics backbone.

District IT and curriculum teams also face procurement cycles: multi-year contracts for detection tools can lock in vendors before the evidence base is solid. Boards that demand opt-out, data-minimization and audit rights up front reduce later cleanup costs.

Pilot results need a full evaluation window — including peak demand, special events and failure modes — before operators treat early performance gains as durable enough for capital expansion or multi-year procurement language.

When a local agency evaluation PDF is not yet public, federal statistical hubs supply only macro context (trade, energy, manufacturing) and should not be misread as precision for a single corridor, fab site or school district pilot.

Integration risk — controllers, interconnect queues, cybersecurity, staff training — often dominates hardware cost in the second year of a deployment. Those operational constraints should appear in any honest before/after narrative.

The coming school year will generate the first large-scale evidence on what works. Until then, policy should stay revisable and data-aware, with mid-year reviews built into board calendars rather than waiting for a full annual cycle.