ChatGPT for students: writing a classroom policy that holds

Most classroom AI policies fail for the same reason: they are written about intentions rather than observable behaviour. "Students must use AI responsibly and with integrity" cannot be enforced, cannot be appealed against fairly, and cannot be explained to a fourteen-year-old in a way that changes what they do at 10pm.

This is a template built the other way around. Every clause describes something a teacher can see, a student can check before they act, and a principal can defend to a parent. It is written for ChatGPT and equivalent general chat tools, because that is what students actually have.

Before you write anything: three decisions

Decision 1: default permitted or default prohibited

Pick one and state it in the first line. Schools that leave this ambiguous get inconsistent enforcement between classrooms, which students detect within a week and which destroys the credibility of the whole document.

Our recommendation is default permitted, with per-assignment restriction. Default prohibited requires you to detect violations, and detection does not work. Default permitted moves the burden to assignment design, which does work, and it puts the policy in the teacher's hands where it belongs.

Decision 2: what gets disclosed

Disclosure is the cheapest enforcement mechanism available, but only if what you ask for is trivial to provide. A reflective paragraph on the ethics of AI use is not trivial and will be fabricated. One line naming the tool and the task is trivial and will mostly be honest.

Decision 3: what happens when someone breaks it

Write the consequence into the policy before an incident, and make it educational rather than punitive for a first occurrence. A policy whose only sanction is an academic-misconduct process will not be used by teachers, because the process costs them more than the violation.

The three-label system

The core mechanism. Every assignment carries one of three labels, printed on the task itself, not held in a policy document nobody rereads.

Label	Means	Typical use
AI-free	No generative AI at any stage. Work is produced in conditions the teacher can observe.	In-class writing, assessed labs, anything used for a formal grade or a placement decision.
AI-assisted	Permitted for specified stages. The student discloses use in one line. The submitted reasoning must be the student's own.	Most homework, research, studying, drafting and practice.
AI-required	Using the tool is part of the task. Often the task is critiquing what it produced.	Source evaluation, prompt design, comparing an AI answer with a textbook, spotting a confident error.

Two practical notes. First, the AI-free category should be small. If half your assignments are AI-free you have written an unenforceable ban with extra steps. Second, teach the labels once, explicitly, in the first week. Every minute spent on this saves an hour of individual argument later.

The policy text

Adapt the wording; keep the structure. Every clause below is observable.

Section 1: Scope

This policy covers generative AI tools that produce text, images, code or audio, including chat assistants, writing assistants, solvers and summarizers. It applies to all work submitted for this course, whether produced in school or at home.

Section 2: Default

You may use generative AI for schoolwork unless an assignment is labeled AI-free. Every assignment carries a label: AI-free, AI-assisted or AI-required. If an assignment has no label, treat it as AI-assisted and disclose your use.

Section 3: What you must always do

Whatever the label, these apply:

1. The reasoning you submit must be reasoning you can explain out loud, unaided, if asked.
2. You must be able to answer questions about any work you submit, including how you got to it.
3. You must not submit generated text as though you wrote it. Quote it, cite it or rewrite it in your own words.
4. You must not enter another person's private information, including a classmate's work, into an AI tool.
5. You must not use AI to produce material that would harass, mislead or impersonate anyone.

Section 4: Disclosure

On AI-assisted work, add one line at the end: the tool you used and what you used it for. For example: "Used ChatGPT to check my structure and to explain the second source." That is the whole requirement. Honest disclosure is never itself a penalty.

Section 5: AI-free work

AI-free assignments are done in conditions we can observe: in class, on paper, or with version history visible. We will tell you in advance which assignments these are and why. We will not use AI detection software to accuse anyone, because it is not reliable enough to base a decision on.

Section 6: If something goes wrong

If a teacher believes work does not reflect your own reasoning, the first step is a conversation, not a sanction. You will be asked to explain your work. If you can, that is the end of it.

Where a first occurrence is confirmed, the outcome is redoing the work under observed conditions and a note home. Repeated occurrences follow the school's existing academic integrity process.

Section 7: Your data

Anything you type into a public AI tool leaves the school. Do not enter your full name, your address, anything about your family, or anything about another student. Where the school provides a tool, we will tell you what happens to your work in that tool.

Why detection is deliberately excluded

Every version of this policy we have seen work has a clause ruling detection software out, and it is worth being explicit about why.

• False positives are not rare enough to build a disciplinary process on, and they cluster on students writing in a second language and on students with an unusual but genuine style.
• An accusation you cannot substantiate costs enormous trust with the student and the family, and one bad case poisons the policy for everyone.
• Detection creates an arms race that the school loses, because the tools improve faster than the detectors.
• It substitutes for assignment design, which is the thing that actually works.

Removing detection from the policy forces the productive question: what would I have to collect for this assignment to be resistant in the first place? Usually the answer is the process rather than the artifact, and usually it makes the assignment better independently of AI.

Assignment design that does the enforcement for you

1. Collect the plan. A hand-written plan produced in class anchors the essay written at home.
2. Require line references. Generated answers about a text are confident and detail-free. Asking for specific lines closes most of the gap.
3. Use local and personal context. A task grounded in this week's lesson, this town or the student's own data cannot be answered generically.
4. Assess in stages. Plan, draft, studying, final. Each stage is small, and each one is evidence.
5. Add a two-minute oral. "Talk me through your second paragraph." Nothing survives that if the student did not write it, and it takes less time than grading would.
6. Occasionally make the AI answer the object of study. Ask students to find the error in a generated response. It teaches evaluation and it demystifies the tool.

Rolling it out

• Week 1: teach the three labels to students in ten minutes. Give one example of each from your own subject.
• Week 1: send the policy home in one page. Parents mostly want to know whether their child is allowed to use it and whether the school has an opinion.
• Weeks 2 to 6: label every assignment. This is the part that gets skipped and it is the part that makes the policy real.
• Half term: review with staff. The usual finding is that the AI-free category was too large and can shrink.
• Ongoing: one short device-free checkpoint per week per class, as calibration rather than surveillance.

Where a school-provided tool changes the calculation

A policy about public chat tools can only ever be defensive. It restricts, it discloses, it designs around. It cannot make the tool report anything back, because those products are built for the student as the customer.

When the school provides a tool whose default is a hint rather than an answer, and which hands the teacher a per-skill record of what the session showed, several clauses above become unnecessary. Homework starts producing better evidence than it did before, rather than worse. That mechanism is described on how it works, and the practical version for homework specifically is in AI homework help that does not do the homework.

Whatever you choose, get the data answer in writing before you recommend anything to students. What happens to a minor's work is a question you will be asked by a parent eventually, and "we did not check" is not an answer. Ours is on the student privacy page, including the certifications we do not hold.

A worked example: the clause that gets tested in week three

Policies fail at a specific sentence, not in general. Here is the sentence that gets tested first, and what happened when it was.

The clause: "You may use AI to help you understand a problem. You may not use it to produce work you submit as your own." Reasonable, widely copied, and ambiguous at exactly the point where a fourteen-year-old will push on it.

The test case

A grade 8 student is set five percent-change problems for homework. He gets stuck on the second one, a $40 shirt up 20% then down 20%, asks a chat product, and it returns a full worked solution ending at $38.40. He then writes that solution out in his own handwriting, in his own phrasing, and submits it.

Has he broken the clause? He would say no: he used it to understand, and the words on the page are his. The teacher would say yes, because he submitted work he did not produce. Both readings fit the sentence as written, which means the sentence is not a rule, it is a disagreement waiting to happen. And it is the teacher who has to have that argument, one student at a time, with no evidence either way.

The clause that survives the test

The version that holds names the artifact rather than the intent: "If a tool produced a step you did not work out yourself, say which step and which tool, in one line at the bottom. Unattributed steps are the problem, not the tool."

That is enforceable because it describes something observable, and it is honest because it does not pretend the tool is avoidable. In practice a majority of students will comply with a disclosure rule they consider fair, and the minority who do not are now doing something unambiguous rather than something arguable.

What the clause still cannot do

Disclosure tells you a tool was used. It does not tell you what the student did not understand, which is the only thing that changes tomorrow's lesson. On the homework above the teacher learns that a solver was used on question two; she does not learn that the student thinks two 20s cancel, or that nineteen others in the room think so too.

That gap is not a policy problem and no wording closes it. It closes when the sanctioned tool hands back a per-skill record of what the session actually showed, the misconception named and the student's own words quoted as the evidence, so the homework produces better diagnostic information than it did before AI existed, rather than worse.

Two pages worth reading before you circulate a draft

If the subject in question is math, the wording of the record matters more than the wording of the policy. "Struggled with percentages" cannot be reteaching; "applied both percentages to the same base" is a five-minute activity. How the skill vocabulary stays specific enough to act on, across fractions, ratio, algebra and measure, is on the AI math tutor page.

And before a student-facing policy names any tool, get the data answer in writing. What happens to a minor's work, who controls the record, whether it trains a model, and what happens at the end of the year are the questions a parent will eventually ask. Ours are answered on the student data privacy and COPPA compliance page, including the certifications we do not hold.