AI, pupil data and GDPR: what a teacher may put into a chatbot

It is Sunday evening, the end of the quarter is three days away and 26 narrative assessments are waiting for you. You copy your notes about the children into ChatGPT on your private account and an hour later you are done. Did you just breach the GDPR? And who is responsible?

The answer: very probably yes, and the school is responsible. You may use AI for assessment. The law does not forbid it, and the Czech National Pedagogical Institute (NPI) even recommends it. What matters is where you send the data and what data it is.

In short:

• A private account in an AI tool is not a school environment. The school has no data processing agreement with the provider, so there is no basis for putting pupils' personal data into it.
• Free and personal plans may train on your conversations. With ChatGPT this is the default, with Claude and personal Gemini it depends on the user's choice.
• Schools that use Google Workspace for Education get Gemini with data protection at no extra cost. On a school account Google does not use conversations for training and the school's contract with Google covers them, even in the free Fundamentals edition. That does not automatically mean the data stays in the EU: only the paid editions offer EU data storage, and even they do not guarantee that the data is also processed in the EU.
• Renaming a pupil to "pupil A" is not anonymisation. In a class of 25 you can recognise a child by a diagnosis, a support measure or their family situation.
• The final assessment must always be the teacher's. AI may help with the wording, not decide.

Why it is a problem even with the best intentions

Pupils' personal data does not belong to the teacher but to the school. Under the GDPR the school is the controller: it decides what the data is used for and who it may be passed to. The teacher works with it as an employee, on the school's instructions.

When a school uses an external service, such as its pupil records system or Google Workspace, it signs a data processing agreement (DPA) with the provider. The agreement sets out what the provider may do with the data, how it protects it and when it deletes it. Your private ChatGPT, Claude or Gemini subscription has no such agreement with the school. It is a contract between you and the provider, with you as a consumer.

In practice this means: by putting notes about a named pupil into a private account, you pass personal data to a third party outside the arrangements the school has set up. This reading is ours, not a decision of a data protection authority. It does, however, follow the basic logic of the GDPR and the NPI recommendation cited below.

Two circumstances raise the risk further:

• They are children. The GDPR gives them special protection.
• School data often contains special categories of data. A diagnosis, a recommendation from a counselling centre, support measures or information about health are data concerning health. The GDPR treats them most strictly (Art. 9).

The other side needs saying too: AI at school is legal and welcome. In its FAQ from November 2025 the NPI states plainly that a chatbot can help a teacher word a narrative assessment. The condition is that the teacher sets the final assessment and uses information about pupils in anonymised form. So the problem is not AI but an uncontrolled flow of data.

Private account vs. school account

The difference is not which chatbot you use but which account you use it under. The same model can run in consumer mode, where it learns from the data, and in organisational mode under a contract, where it does not.

Turning off training does not solve the problem. The switch in the settings only stops the model from learning from your new conversations. It does not create an agreement between the school and the provider, and the data stays stored with the provider. Likewise, signing into a personal plan with your school email does not make it a school environment.

Providers change their terms often. The table reflects public information as of October 2026.

Gemini at school: what actually applies

If your school uses Google Workspace for Education, you most likely have AI with data protection available for free. You just may not know it.

What Google guarantees:

• Since 2025 the Gemini app and NotebookLM are a core service in all editions, including the free Education Fundamentals.
• Since June 2025 they are available to users of all ages, pupils included.
• On a school account, according to its FAQ, Google does not use conversations for training and human reviewers do not read them.
• Gemini is covered by the school's contract with Google, including the data processing addendum (Cloud Data Processing Addendum).

What does not apply automatically:

• "Data only in the EU". Google launched data regions for the Gemini app in June 2026. For schools they come only with the paid Education Standard and Plus editions, and they guarantee only that data is stored in the region, not that it is processed there. On the free Fundamentals edition the school cannot choose a region. Even there, though, data transfers are covered by the contract.
• It works only under the school account. If your browser is signed into a private Gmail account, consumer terms apply. Check in the corner of the screen which account you are working under.
• The administrator has to turn it on. If you do not see Gemini in your school account, ask your IT administrator.

May you put pupils' personal data into school Gemini? Google does not forbid it, and unlike with a private account there is a contractual framework for it. The decision, however, belongs to the school as the controller, not to Google. Some school systems take a stricter line: Wales's education platform Hwb instructs teachers not to enter personal or sensitive data into Gemini, even through a school account. The Czech NPI recommends always using made-up names and anonymised data in prompts.

A sensible default rule is therefore: school Gemini yes, but work without names there too. Do not enter health data or support measures until the school has explicitly assessed and allowed it.

Anonymisation is not renaming

Personal data is not just a name. It is any information by which a particular person can be identified, directly or in combination with other data.

Picture the sentence: "Pupil A from 4B, who has had a teaching assistant since September because of ADHD and moved here from Brno last year." There is no name, yet anyone at the school knows who it is. That is pseudonymisation, not anonymisation, and under the GDPR it is still personal data.

What gives a child away even without a name:

• a diagnosis, a support measure, a teaching assistant, an individual education plan
• the family situation (divorce, bereavement, moving house, social situation)
• nationality, mother tongue, religion
• one-off events (injuries, disciplinary measures, success in a competition)
• a combination of class, gender and a distinctive trait

Three rules that cover most situations:

1. Remove the details before you paste the text. An instruction to "anonymise this first" means you have already sent the original.
2. Describe observed behaviour and progress, not circumstances. "Solves two-step word problems independently" is all the AI needs. It does not need to know why the child struggled before.
3. Add the names on your own side. Work with a placeholder such as [NAME] and put the real name in only once the text is finished, in the school system.

When you are not sure whether a child could be recognised, ask: Would anyone at our school know who this is? If so, the text is not anonymous yet.

Three typical situations

Narrative assessment

This is where AI saves teachers the most time and where they take the biggest risk. Notes about a child tend to be personal and detailed.

• Do not enter: the name, the class, a diagnosis, support measures, family circumstances, an export from the pupil records system.
• Do enter: specific observations without context ("reads a half-page text with understanding"), and the tone, length and structure you want.
• A better approach: have the AI draft wording templates for typical levels of progress, then match them to the children yourself.

The NPI adds an important point: an assessment has legal effects. Under Art. 22 of the GDPR it therefore must not result solely from automated decision-making. You have to read the finished text, edit it and be able to stand behind it.

Bulk tasks

Uploading a spreadsheet with the names and marks of a whole class into a chatbot to sort it, calculate averages or write letters to parents is the riskiest option. One paste hands over the data of dozens of children at once.

• A spreadsheet or the school system can calculate averages and statistics without AI.
• If you do need AI, replace the names column with numbers and keep the key to yourself.
• Have bulk letters generated as a template with placeholder fields and fill in the names with mail merge in Word or Google Docs.

Marking tests

A photo or scan of a test usually carries the child's name, class and handwriting. The pupil's answers are, moreover, their personal data, as the Court of Justice of the EU confirmed in 2017 (Nowak, C-434/16).

• Cover the name before taking the photo, and number the tests.
• Ideally have the AI prepare only the marking criteria and model answers. Do the marking yourself, or in a school tool covered by an agreement.
• Always check the points the AI suggests. The final mark is your decision.

A traffic light for teachers

A simple split into three colours is enough for most situations. We took inspiration from Patrick Zandl's traffic light for companies and adapted it for schools.

Three questions before every paste:

1. Does the AI really need this information, or only help with the wording?
2. Am I working under the school account or a private one?
3. Would anyone at the school know which child this is?

What about the AI Act?

The EU regulation on artificial intelligence classifies AI systems intended to evaluate learning outcomes as high-risk (Annex III). For schools, though, it is not pressing yet.

• The obligations for high-risk systems were to apply from August 2026. The so-called Digital Omnibus (Regulation 2026/1744) postponed them to 2 December 2027. It has been in force since 27 July 2026.
• The category itself has not changed. AI for assessment in education remains high-risk.
• Whether a teacher who has an ordinary chatbot suggest a narrative assessment is "deploying a high-risk system" is not clear-cut. The classification depends on the purpose the provider intended the system for. No Czech methodological guidance we could find settles the question yet.

The practical upshot: the GDPR applies now, and in full. In 2027 the AI Act will add further obligations, mainly for providers and for schools that use specialised assessment tools. The principle that "the teacher has the final word" is worth following today.

What to do

As a teacher

• Find out whether your school has Google Workspace for Education or Microsoft 365 Education with AI, and work there under your school account.
• On a private account, use AI only for the green tasks from the traffic light. Turn off training on your data there.
• Work without names and without circumstances that give a child away. Add the names on your own side.
• If you have already put personal data into a private tool, delete the conversations and tell the school leadership or the data protection officer. The officer will assess whether it is a personal data breach that has to be reported to the supervisory authority (usually within 72 hours of becoming aware of it).
• Read and edit every text from the AI. You are responsible for the assessment.

As school leadership

• Approve one school AI tool covered by an agreement and announce it to teachers. Banning AI makes no sense: teachers will use it anyway, only less visibly.
• Issue a short policy: which account, what may be entered (the traffic light), who handles an incident.
• Keep a record of the AI tools in use and inform parents. The NPI lists both obligations.
• Discuss with the data protection officer whether processing pupils' data in AI requires a data protection impact assessment (DPIA).

How to recognise a safe tool for pedagogical records

When a school is choosing a specialised tool for assessing or observing pupils, ask the supplier five questions:

1. Will you sign a data processing agreement with the school?
2. Where is the data stored and processed, and which subprocessors have access to it?
3. Do you use the data to train AI models?
4. Does the tool suggest the assessment or decide it? (The right answer: it suggests, the teacher confirms.)
5. How is the data deleted when the school ends the contract?

Sources

• NPI and aidoskol.cz: Artificial intelligence and the law at school, FAQ (November 2025, in Czech)
• NPI: Recommendations for using artificial intelligence in primary and secondary schools (in Czech)
• Google for Education: Privacy & Security FAQ
• Google Workspace: Generative AI in Google Workspace Privacy Hub
• Google Workspace Updates: Gemini and NotebookLM as core services for all Education editions (April 2025)
• Google Workspace Updates: Gemini for users of all ages (June 2025)
• Google Workspace Updates: Data regions for the Gemini app (June 2026)
• Hwb (Wales): Google Gemini in Hwb
• OpenAI: Turning off training in ChatGPT
• Anthropic: Changes to Claude's consumer terms
• Patrick Zandl: Smlouvy, AI, GDPR (Marigold.cz, October 2026, in Czech)
• AI Acto: The Digital Omnibus and the AI Act postponement

As of October 2026. This article is not legal advice. The specific setup at your school is for the school leadership to assess with its data protection officer.