Steal my AI policy. Here is the thinking behind it.

Here is the policy I run in my engineering classes. Take it. Change the parts that do not fit your subject. I will tell you which parts those are.

Why most school AI policies fail

Almost every policy I have read is binary. AI is banned, or AI is allowed. Both fail for the same reason.

The ban fails because it is unenforceable and everyone knows it. Detectors do not work well enough to act on, they false-flag hardest against students who write in a second language, and the moment you accuse an honest kid on a detector score, you have lost the room for the year.

The permission fails because "you can use AI" does not tell a fifteen year old whether pasting the prompt into the assignment counts. They are not being sneaky. They genuinely do not know, and neither do you until you sit down and decide.

The real problem is not AI. It is ambiguity. Nearly every AI conflict I have seen came down to a student and a teacher holding two different unstated assumptions about the same assignment.

So the policy's actual job is to remove ambiguity, assignment by assignment.

The spine: defensibility

One rule sits above everything else.

This matters more than it sounds, because it relocates enforcement from detection to demonstration. You are no longer in the business of proving what tool a student used, which you cannot do reliably. You are in the business of checking whether they understand what they turned in, which you can do in ninety seconds and have been doing your entire career.

Every other part of the policy exists to support this one.

Three stamps, applied per assignment

Per-assignment, not per-course. This is the load-bearing design decision, and it is the part most policies get wrong.

Then you put the stamp on the assignment. Header, Canvas description, top of the handout, wherever your students actually look. Default when a stamp is missing is ASSIST, and you say that once, out loud, in week one.

This takes about four seconds per assignment after the first week, because you already know which category each task belongs to. You just have never had to say it before.

Disclosure has to be nearly free

When students work at COLLABORATE, they log the tool, the methods they used with it, the questions they asked, any research agents they ran and for how long, and what they did with what came back.

Keep it short enough that it actually happens. Not a form, not a reflection paragraph, not a separate document. Mine runs under a minute. If disclosure costs much more than that, students stop doing it, and then you have a policy that exists only on paper.

Frame it as professional practice rather than confession. Engineers document their tools. Researchers cite their instruments. This is the same move, and students accept it immediately when you present it that way.

Adapting it to your subject

The three stamps are portable. What changes is where you draw the NO ASSIST line, and that depends entirely on what your course is actually assessing.

English. The bulk of the writing process moves to ASSIST, not COLLABORATE, because the composing is the skill. Brainstorming and outlining can go COLLABORATE. Final drafts almost never should. Watch for the trap of letting AI "polish," which quietly removes the exact voice you are trying to grow.

Math. Heavily NO ASSIST for procedural fluency, and genuinely useful at ASSIST for the student stuck at 9pm with nobody to ask. That second case is the strongest argument for AI in a math classroom and it rarely gets made.

History and social studies. COLLABORATE works well for source gathering and gets dangerous for interpretation. Add one subject-specific rule: any factual claim sourced from a model must be traced to a real, citable document. Fabricated citations are the single most common failure mode here, and catching one is a better lesson than any lecture you could give about it.

Science and CTE. Verification is the whole game. Any figure that comes from a model and goes into a calculation, a build, or a lab writeup gets checked against a datasheet, a standard, or a hand calculation. Models produce confident, plausible, wrong numbers, and a student who learns to catch that has learned something that transfers directly into the workplace.

Elementary. Mostly NO ASSIST, with AI used by the teacher and shown on the projector rather than handed to students. The literacy goal here is watching an adult question a machine's answer out loud.

The three lines that do not move

Whatever subject you teach, these carry over unchanged.

1. No student PII in public AI tools. No names, no grades, no IEP details, no behavioral notes. This one is not a classroom-culture preference, it is a FERPA exposure, and it applies to your use as much as to theirs.
2. Nothing gets submitted unread. Trivially enforced by the defensibility rule, and worth saying explicitly anyway.
3. No AI-written self-reflection. A model has no access to the student's thinking, so the output is fiction by construction.

What to do on the first violation

Talk, then redo. First offense is a teaching moment, not a referral.

I say this in the policy itself, and I say it deliberately, because students who believe the first mistake will end them will hide the mistake. A stated amnesty on the first crossing buys you honest conversations for the rest of the year, which is worth vastly more than the one grade you gave back.

Repeat violations go through your existing academic integrity process. You do not need a new one. AI did not create a new category of dishonesty; it made an old one cheaper.

Take the one-pager

I built the whole thing as a single printable page. Classroom wall, syllabus packet, back-to-school night handout, department meeting. Editable, no attribution required, use it however it is useful.

If you want to see this running live in an actual classroom, the student-facing version is posted for my own families over at the STEMAGEDDON program site. Same policy, written for fifteen year olds and their parents instead of for you.