The short answer is names, contact details, results, behaviour and health records, and anything that could identify a child, including photos, voices, and student work you have not cleared. If someone could trace an output back to one student, that input did not belong in the box.
This guide covers what counts as student data, the five categories to keep out of any chatbot prompt, how to rewrite a request so the tool is still useful, and the questions to ask your school before you use one at all. None of it requires a technical background.
What counts as student data in a chatbot prompt
Student data is any information that identifies a child or that describes a child in a way that would let you pick them out of a group. The second part is where most teachers get caught, because a single detail can be identifying even without a full name.
- Direct identifiers: Name, initials, date of birth, student number, photograph, email address, home address, bus number.
- Indirect identifiers: Class lists with any detail attached, seating plans, marked scripts with names in the header, group photos with captions, a named student described as the one who was absent, unwell, or disruptive.
- Records: Attendance, marks, behaviour logs, assessment data, medical or therapy notes, support plans, and anything held as an education record.
- Special category data: Health, disability or special needs, religion, and safeguarding concerns. This is the most sensitive material in a school.
The legal framing varies by country, and the basic duty is similar. In the United States, FERPA protects education records held by schools and districts, which sets the question of who a vendor is and what it may do with them. In the UK and the EU, most school data is personal data under the UK GDPR, with health, religion, and disability data treated more strictly. India's Digital Personal Data Protection Act 2023 places obligations on education providers handling children's data. Check your own school's policy for the detail that applies to you.
The five things never to paste
- Names and class lists. Remove them even when you think the tool is anonymous, because inputs can be stored, logged, or reviewed.
- Contact and account details. Guardian email addresses, phone numbers, home addresses, portal logins, and any staff credentials.
- Assessment and behaviour records. Grades, marks, attendance figures, behaviour notes, and exclusion records. These are the records most tightly governed.
- Health, support, and safeguarding information. Medical details, therapy notes, individual education plans, and any information covered by a safeguarding concern. Do not summarise these either.
- Identifiable media and work. Photos, video, voice recordings, handwriting samples, and student writing that carries a name, a unique error pattern, or a distinctive context.
A test that covers most cases: if the paste included a sentence about one child, delete the sentence, not just the name. If the sentence still points to a person, the sentence goes too.
Rewrite the request instead of deleting the name
Most prompts do not need student data to work. They need task data. Replace the individual work with your own example, a model answer, or a description of the problem, then ask the tool to build the thing you actually wanted.
I teach [subject and grade]. Here is a task description I wrote: [paste your own task, with no student details]. Students are meeting the objective [objective]. Give me: (1) three worked examples at increasing difficulty, (2) the misconception most likely in this topic and how to surface it, and (3) five exit-ticket questions matched to the objective. Write the examples yourself. Do not invent curriculum codes, exam board references, or student profiles, and do not ask me to paste student scripts.
Fill in these details
- Subject and grade: Sets the vocabulary and reading level of the output.
- Task description: Your own wording is safe input. A teacher's task sheet contains no personal data.
- Objective: The single learning goal the material must serve. Vague objectives produce generic material.
- What you replace: When you need a real error pattern, describe its shape, such as sign errors when moving terms across the equation, rather than pasting the script.
This form works for rubrics, exit tickets, differentiated versions, parent notes, and marking schemes. The tool receives structure, not people.
Worked example: an assessment request, de-identified
A common request looks like this: paste thirty marked scripts so the model can find the misconception. Here is the same request rewritten.
| What you wanted to paste | Why it is a problem | What to send instead |
|---|---|---|
| Thirty marked scripts with names | Names, marks, and identifiable error patterns for thirty children | Three samples you wrote yourself, or a described error pattern |
| A gradebook extract | Every mark and attendance record in one paste | The distribution you want to fix, described in words |
| A named student's behaviour note | Behaviour records and a named child | The pattern, such as instructions not followed on independent steps |
| A photograph of a worksheet | Identifiable image data | A typed transcription with the name removed |
| An individual support plan | Special category data, tightly governed | Nothing. Required supports come from your professional judgment |
The teacher still gets what the prompt was for. What changed is that the identifying layer was removed before the text left the browser, not after.
Six questions to ask before your school's first AI tool
- Which tools are approved? Some schools maintain a list of accounts with an agreement in place. Using a personal account for school data is usually the line to avoid.
- What happens to what I type? Ask whether inputs are used for model training, how long they are kept, and whether that can be switched off. Providers differ, and settings change, so confirm rather than assume.
- Is there a school agreement? Vendors often offer education terms with different retention and training defaults. Ask whether your school has signed one.
- Who can see the input history? Teachers with admin access on shared accounts may be able to review conversations.
- What does the policy say about pupil work? Some schools restrict uploading any pupil-produced material, even anonymised.
- Who do I ask when unsure? Usually the data protection officer, the head of department, or the IT lead. A written answer is easier to rely on than a corridor conversation.
Ask these once, get the answer in writing, and your prompts get much faster to write.
Common mistakes that put student data at risk
- Removing only the full name. Initials, a distinctive sentence, or a single marked detail can still identify a child.
- Assuming deletion makes it anonymous. A paraphrase of a real incident can be re-identifiable in a small school or class.
- Asking the tool to anonymise your paste. The unredacted text has already been sent by then.
- Using a personal account for school business. Consumer terms and school agreements are not the same thing.
- Trusting a summary. A long policy document summarised by a chatbot may drop the clause that mattered. Read the original.
Frequently asked questions
Is ChatGPT safe for teachers?
It depends on which version you use, what your school has agreed to, and what you paste. Many consumer plans allow providers to use conversations for model improvement unless you change a setting, and some offer business or education terms with different defaults. Ask your school before your first use.
Can I paste student work if I remove the names?
Only if your school permits it and the work stays unidentifiable. Even then, distinct error patterns can point to one child. Safer options are work you wrote yourself, published models, or a written description of the problem.
What about my own lesson plans and materials?
Your own material usually contains no student data, so it is lower risk. Check for names that slipped into an example, a photo of a whiteboard with students' work, or a pupil quotation you included.
Set one rule, then check every prompt
Write one rule for yourself: nothing leaves the browser that names a child or describes one in a way that could be traced back. Apply it before you paste, not after you read the output. For prompts that stay clear of student data, browse the AI prompts for teachers topic collection.