A quarter of teenagers now use ChatGPT for schoolwork, and that share doubled in just two years — from 13% in 2023 to 26% in 2025, according to Pew Research Center. The first instinct in most schools was to ban it. That instinct mostly failed. Districts that blocked the tools quietly reversed course, and some, like Los Angeles Unified, went on to build AI assistants of their own.

So the question worth your time is no longer whether AI belongs in the classroom. It is where. Most guidance you will find splits into two camps: lock it out to stop cheating, or let it in to stay current. Real classrooms need something more useful than a side to pick — a map of where AI genuinely helps, where it is harmless, and where it has to stop.

Start with what students already believe

Teachers get quoted constantly in the AI debate. Students rarely do. That is a problem, because students are not asking for a free-for-all — they already draw lines, and those lines are sharper than you might expect.

In the same Pew data, 54% of teens said using AI to research a new topic is acceptable, but only 18% said the same about using it to write an essay. Read that again: most students distinguish between AI as a study aid and AI as a stand-in for their own work. They have an instinct for where help ends and dishonesty begins. When you co-write classroom norms with students instead of handing them a rulebook, you are building on a sense of fairness that is already there. Imposed rules invite workarounds. Shared rules invite buy-in.

Manage the everyday, protect the high-stakes

The dividing line that actually works is not "AI good" versus "AI bad." It is stakes.

Low-stakes, formative work — brainstorming, practice problems, study guides, feedback on a rough draft — is exactly where AI earns its keep, and where policing it burns energy you do not have. High-stakes, credential-bearing work — the final exam, the certification test, the graded essay that stands alone as proof of what a student can do — is where unmanaged AI quietly empties the result of meaning. A grade that no longer signals competence helps no one: not the student, not the next teacher, not the employer reading the transcript.

Sort your assessments by that logic and the path forward gets clearer:

This is the same instinct behind every tactic the cheating-prevention playbook recommends — timed tests, in-class writing, project work submitted in stages. Those tactics work because they raise the cost of outsourcing where the result has to be trusted, while leaving room for AI everywhere else.

Where to stop: three lines worth holding

"Manage everyday use" is the easy half. The harder half is knowing where to plant a firm boundary. Three lines are worth holding.

Stop treating AI detectors as proof. Detection tools feel like a tidy answer, but they flag honest student work as AI-generated often enough to wreck a student over nothing — some tools are wrong close to half the time. A false accusation costs a student more than a missed catch costs you. Use a detector to start a conversation, never to deliver a verdict.

Stop letting AI do the thinking that builds the skill. This is the cost the current coverage skips. The damage from offloaded thinking does not show up on the due date — it shows up a year later, when the reasoning, the writing, or the problem-solving that should have formed simply never did. Skills are built through productive struggle. When a student hands the struggle to a chatbot during the exact task meant to develop it, the grade looks fine and the learning is gone. Protect the difficulty that does the teaching.

Stop pretending access is equal. The strongest models sit behind paywalls, and home setups vary wildly. A blanket "use AI freely" policy quietly rewards students who can pay for the better tool and the quiet room. A blanket ban lands hardest on students who rely on a shared school device and have nowhere to work at home. Either extreme widens the gap you are trying to close, so write policies that name a specific tool and offer it equally, or assume nothing about what students can reach.

Write a policy students can actually follow

A forty-page acceptable-use document that nobody opens is not a policy. A clear label on the assignment is. The most workable frameworks tier AI use per task rather than ruling on it for the whole course:

• Green — AI encouraged. Spell out how it helps: generate practice questions, explain a tricky concept, check grammar.
• Yellow — AI allowed for specific steps, with disclosure. Students may use it to research or outline, then note where and how they did.
• Red — no AI. Work is produced under supervision so the result can be trusted.

Put the colour on the assignment itself, not buried in a handbook. Revisit the tiers each term as the tools change — and they will. A policy this concrete does what a sweeping ban never could: it removes the guesswork that pushes honest students toward accidental violations.

Measure learning, not just the output

Here is the uncomfortable shift underneath all of this. A finished essay or a clean problem set is no longer reliable evidence that learning happened. So move some weight onto the things AI cannot fake at scale: in-class work, a short oral defense of an idea, drafts that show thinking over time, the explanation behind an answer rather than the answer alone.

And for the handful of results that must be trusted by someone outside your classroom — a board, a recruiter, the next institution — supervised conditions remain the cleanest signal you have. That is not about distrust. It is about making sure a credential still means what it says.

How AI proctoring actually works for a teacher

"Set supervised conditions" sounds heavy when you picture invigilating every online test by hand. You don't have to. AI proctoring is what makes the red tier practical at the scale a real class runs at — and used well, it does the opposite of watching students like a hawk.

Here is the shape of it. You attach proctoring to a test you already use — a Google Form, a Typeform quiz, or your own platform — and students take it in their browser. While they work, the system handles what one teacher cannot watch across thirty open laptops at once: it confirms identity, watches for tab-switching, a second device, or extra faces on camera, and notes when something looks off. AI proctoring is built for exactly this — monitoring many test-takers at once so you are not refreshing video thumbnails for an hour.

The part that matters most is what happens with a flag. A good tool hands you a report and the captured evidence, then leaves the judgment to you. AutoProctor is deliberate about this: it dropped eyeball-gaze tracking after finding it flagged honest students who simply looked away to think, and it treats its trust score as a prompt to review the footage, not a verdict to act on blindly. That is the earlier principle in practice — the software flags, the human decides. Automated proctoring earns its place when it narrows your attention to the few moments worth a second look, not when it swaps your judgment for a number.

The discipline is restraint. Proctor the certification test, the final, the exam whose result someone outside your room has to trust. Leave the practice quiz, the rough draft, and the green-tier work alone. Supervised conditions exist to protect what a grade proves — not to become the default setting for everything a student touches.

What to do this term

You do not need a district task force to start. You need three moves:

• Audit your assignments into green, yellow, and red. Most will be green or yellow, which is the point — it shows you how little actually needs locking down.
• Co-write the norms with your students. Their instinct for fair use is already sharper than the debate gives them credit for.
• Reserve proctored, supervised conditions for the few assessments that genuinely certify competence — and let AI breathe everywhere else.

The goal was never to win a war against AI. It is to decide, on purpose, where AI helps your students learn and where it has to step aside. Manage the everyday. Hold the line on what a grade is supposed to prove. That is a boundary students can respect — because you drew it with them, and you drew it for a reason.