ChatGPT for teachers: where it helps and where it breaks privacy

ChatGPT and its equivalents are genuinely useful to teachers. They are also the fastest route to a student data incident that a school has ever had, and the two facts live about four inches apart in daily practice.

The line is not about the tool, the plan or the vendor's security page. It is about whose words are in the prompt box. Once you hold that distinction, almost every practical question answers itself.

The one rule

Your own words: fine. A student's words, work or identity: not without a school agreement covering exactly that.

Everything below is an elaboration of that sentence. It is deliberately simple, because a rule that requires a flowchart will not be followed at 4pm with forty books to mark.

What is fine, and why

These are your professional work product. No student is identifiable, no education record is involved, and the school has no obligation engaged.

• Drafting a lesson plan, a pacing guide or a knowledge organiser.
• Generating practice questions in the style of ones you already use.
• Rewriting your own worksheet at three reading levels.
• Turning your own learning objectives into a rubric.
• Drafting a parent letter about a trip, with no individual student in it.
• Asking for likely misconceptions on a topic before you teach it.
• Rewriting your own email so it sounds less irritated than you feel.
• Summarising a public policy document or a piece of published research.

This is a large fraction of the administrative load of teaching and there is no good reason to be nervous about it. A vague sense that "AI is risky" that stops teachers using it here, while doing nothing about the actual risk below, is the worst of both worlds.

What breaks the rule

Each of these involves a student's education record, their identity, or content they produced, going into a general-purpose consumer product on the teacher's personal account.

• Pasting a student's essay for feedback.
• Pasting a set of answers to be marked.
• Asking for help writing a report comment with the student's name and details in the prompt.
• Uploading a class list, a mark sheet or a spreadsheet with identifiers.
• Describing a safeguarding concern, a diagnosis, a behaviour incident or a family circumstance to get advice.
• Uploading a photograph of a class or of work with names visible.
• Asking it to draft an individual education plan from a student's real profile.

Almost every one of these is done with good intentions by an experienced teacher trying to do a better job in less time. That is exactly why "use your professional judgement" is not sufficient guidance: the judgement being exercised is about the quality of the feedback, not about the data.

Why it matters legally, in two paragraphs

Under FERPA, a district may share education records with a vendor under the school official exception, but that requires the vendor to be under the school's direct control with respect to those records and to use them only for the purpose the school specifies. A teacher's personal consumer account is not under the school's direct control. Nothing about the arrangement satisfies the exception, so the disclosure is simply a disclosure.

Under COPPA, where students are under 13, the school can consent on parents' behalf for educational purposes. That is only available where the operator limits use to those purposes. A general consumer product with a broad terms-of-service, on an individual adult's account, is not operating under that arrangement, and the school has not given consent for it.

There is a second-order problem worth naming. Consumer plans have historically defaulted to using conversation content to improve models, with a setting to opt out. A setting that individual teachers must find and toggle is not a control the school can rely on, and it is not one an auditor will accept.

The realistic risk, not the theoretical one

Nobody is likely to extract a specific student's essay from a model. The realistic damage is more ordinary and more common.

1. A parent asks. "Was my child's work put into ChatGPT?" If the honest answer is yes, the conversation is about your judgement, not about AI, and it is not a good conversation.
2. An audit or an information request. A school has to describe where education records go. "Individual staff members' personal accounts" is a sentence no leader wants to write.
3. A leak of a different kind. Screenshots circulate. A prompt containing a named student's difficulties, forwarded once, is a serious problem entirely independent of the model.
4. Institutional overreaction. One incident produces a blanket ban, which removes the legitimate uses in the first list and pushes the rest underground.

Making it usable: four practical patterns

1. Describe instead of paste

Rather than pasting a student's paragraph, describe the pattern: "A grade 8 student writing about a source keeps summarising instead of inferring. Give me three questions that push towards inference." You get the same useful output with no education record involved, and often better output, because you have done the diagnosis yourself.

2. Anonymize properly, or not at all

Removing the name is not anonymization. A description of a specific incident involving a specific child in a specific class is identifying to anyone who knows the school. Either genuinely generalise or do not use the tool for that task.

3. Keep planning and grading on different rails

Planning work: any tool, no constraints, go fast. Anything involving student work: only through a tool the school has an agreement with. Teachers hold that split easily because it maps onto something they already understand, which is the difference between their own materials and the children's records.

4. Give staff a sanctioned option for the student-facing half

The pasting happens because the need is real. A teacher with 140 students cannot write individual diagnostic feedback by hand, and telling them not to use the only tool that makes it possible without providing an alternative guarantees the rule is broken quietly. A school-provided tool that reads student work under an agreement, does not train on it and hands back a per-skill record removes the temptation entirely. That is what Coeducate is, and the commitments and the missing certifications are both listed on the student privacy page.

A one-page staff briefing you can copy

Use it freely for: your planning, your resources, your rubrics, your emails, your rewriting of your own material.

Do not use it for: anything containing a student's work, name, results, behaviour, family circumstances or special education information, unless it is a tool the school has an agreement with.

Instead of pasting student work: describe the pattern you are seeing and ask for what you need.

If you are unsure: ask before you paste, not after. Nobody is in trouble for asking.

School-approved tools for student work: [list them here, with a link to what each does with the data].

Questions to ask before adding a tool to that last line

• Is student work used to train models? Get it in writing.
• Who is the controller of the records, and is that stated in the agreement?
• Can we delete a student's records on request, and does that include derived data?
• Which subprocessors touch the content, and under what terms?
• What certifications do you actually hold, and which do you not? A vendor that cannot name a gap is not being straight with you.
• What happens to our data if we leave?

A worked example: one Tuesday afternoon, one paragraph, four decisions

The abstract rule is easy. Here is what it looks like at 4pm on a Tuesday, when a grade 8 math teacher has 28 percent-change answers in front of her and a data meeting on Thursday. Four moments, four decisions, and the third one is where most schools actually lose control of the policy.

Decision 1: the thing she wants to paste

Marcus wrote: "the shirt is $40 and it goes up 20% so thats $48 then down 20% so the 20s cancel out and its back to $40." She wants to know why nineteen students in the class made a version of that mistake and what to do about it before Thursday. The fastest route is to paste all 28 answers into a general chat product and ask it to group them.

Every element of what she wants to do is reasonable. The need is real, the tool is competent at it, and it would take four minutes. This is the moment the policy either works or is quietly ignored, and a rule that pretends the need does not exist will lose here every single time.

Decision 2: what actually goes in the box

Under the split above, that paste is not allowed: it is student work, it is identifiable by handwriting and by position in a class set, and the general product has no agreement behind it. What is allowed is the pattern without the people. She types instead:

"Grade 8, successive percent change. About two thirds of the class are treating a 20% rise then a 20% fall as cancelling, and getting back to the original price. Give me four diagnostic questions that separate 'the 20s cancel' from 'I applied both percentages to the original number', and one five-minute board activity."

That prompt contains no student, no work and no record. It gets her 90% of what she wanted, and it is inside the rule rather than outside it. This is the move that has to be taught explicitly in the briefing, because it is not obvious: describe the pattern, not the people.

Decision 3: the part the rule cannot reach

What she still does not have is per-student evidence. She knows roughly two thirds got it wrong. She does not know which of the four possible faults each student has, and she cannot say in Thursday's meeting that Marcus has "successive percent change: not yet" while Priya has "arithmetic slip, method sound". Getting that would require the actual work, which is exactly what the rule forbids her to paste.

So on Thursday she reports a class-level impression and reteaches percentages to everyone, including the nine students who did not need it. The rule held, and it cost her the diagnostic. This is the honest failure mode of a paste-prohibition with no sanctioned alternative behind it: it does not remove the need, it just moves the cost onto the teacher and the class. In most schools, a term of that is how the rule ends up being broken quietly.

Decision 4: the sanctioned route, and what it changes

With a school-agreement tool that reads student work, the same Tuesday ends differently. Marcus's answer goes in as itself. What comes back is not a grade: it is the named misconception ("applies both percentages to the same base"), the evidence line quoting his own words, four skill rows with states, and a five-minute reteach. The class view aggregates the 28 into three distinct groups instead of one blur, and Thursday's meeting gets a per-skill record rather than an impression.

The difference is not that the AI is better. It is that the data is somewhere it is allowed to be, under an agreement that says what happens to it. The teacher stops having to choose between doing her job and following the policy, which is the only version of a policy that survives a full academic year.

The two pages to read before you write your own version

If your subject is math, the vocabulary problem in Decision 3 is the whole game. "Struggled with percentages" is not a diagnosis and cannot be reteaching; "chose the wrong base for the second percentage" is both. What a per-skill record looks like when the skill names are mathematical rather than generic, and the specific fault lines it separates across fractions, ratio, algebra and measure, is set out on the AI math tutor page.

And before any of it reaches a classroom, somebody will ask the question in Decision 4: where does minors' work go, does it train a model, who controls the record, and what happens at the end of the year. Our answers, including the parts where the answer is "no, we do not hold that", are on the student data privacy and COPPA compliance page. A vendor that cannot name a gap in its own posture is not being straight with you, so that page names ours.

Related reading: writing a student-facing AI policy that holds, and deploying AI in schools without losing student privacy.