If you run IT for a school district, the 1:1 device program used to be your proudest line item. Every student with a laptop, every classroom connected. Then generative AI arrived, and overnight the same fleet became the largest unsupervised AI deployment in your county.
Nobody asked whether IT wanted to own academic integrity. The devices are yours, so now it's yours. This post is the playbook I wish someone had handed the IT leads we work with before their first board meeting on the subject.
Start with the question that actually matters
Skip "how do we block AI." You'll be asked for it; push back on it. The block-everything approach fails technically (the tools rebrand constantly), pedagogically (it kills legitimate study use), and politically (the first parent whose kid loses tutoring access will find your inbox).
The workable question is: how do we let the learning uses through while stopping the cheating uses, and how do we prove which is which? Every architecture decision falls out of that sentence.
Why the filter you already own doesn't do this
Most districts already run web filtering, and the vendor probably added an "AI category" checkbox last year. Two structural problems.
First, category filtering is binary per site. The same chatbot serves the kid asking for an explanation of covalent bonds and the kid ordering a book report. Site-level controls can't split that hair, and that hair is the entire problem.
Second, the filter lives on your network. Roughly none of homework happens on your network. A control that evaporates when the student goes home, or toggles to a hotspot, isn't a control. It's a gesture.
What's needed is enforcement that lives on the device itself and reads the content of the prompt, not the address of the website. On-device means it travels home with the laptop. Prompt-level means tutoring and cheating finally get different answers.
The rollout, in order
Week one: policy before software. Get teachers to write the classroom guidelines in plain English. Not a committee document, actual sentences: tutoring yes, essay-writing no, nothing during assessments. If the guidelines can't be written simply, no tool can enforce them, so this step is the real foundation.
Week two: pilot fleet. One grade level or one lab. Deploy the agent through whatever you already use for software distribution; anything that can push an MSI or a PKG can do this. The browser extension side deploys through your existing Chrome or Edge management. No new infrastructure is the bar a vendor has to clear here, and you should hold them to it.
Weeks three and four: watch, don't enforce. Run in alert-only mode. You'll learn more about actual AI use in your school in two weeks of passive visibility than in a year of surveys. Expect surprises in both directions: more use than anyone guessed, and more of it legitimate than anyone guessed.
Month two: turn on enforcement, loudly. Tell the students exactly what's enforced and why. The deterrent effect of a clearly announced system does more work than the blocks themselves. The goal was never to catch kids; it's for the dishonest prompt to never feel worth typing.
The three questions you'll get, with answers
"Is this spying on students?" No, and scope is your defense. The check happens only when a prompt is headed to an AI tool on a school-managed device. No message scanning, no document access, no browsing surveillance. Practice saying this sentence; you'll say it at a board meeting.
"What about their phones?" Out of scope, and that's fine. Your duty covers school-managed devices and school accounts. The kid with a personal phone on a data plan is a parenting conversation, not an IT deliverable. Don't let perfect become the enemy of the 95% you can govern.
"What happens when the AI tools change?" This is the question that separates real solutions from domain lists. Detection that works behaviorally, recognizing AI-style sites and AI-bound prompts rather than matching a fixed list, keeps covering tools that didn't exist at deployment. Ask any vendor exactly this question and watch whether they reach for the word "list."
The bar for success
Six months in, you want three things to be true: teachers reference the dashboard instead of their gut, students use AI for studying without sneaking, and your weekend domain-list spreadsheet is deleted.
That last one is personal. I've seen the spreadsheet. Nobody should live like that.