Module 4 · About 25 minutes

Responsible use

The first three modules answered "how do I use it". This module answers "where do I stop". You will clean the data leak out of a real-looking prompt with your own hands, hunt the bias in a model's output, rank how much scrutiny each kind of task deserves, and finish by generating a one-page AI policy you can take to your department.

You can also complete this module by listening

The audio summary of this module was produced with the Madlen Podcast tool.

0:00 0:00

Why this module

Knowing how to use the tool is not enough.

Module 2 showed you that a good prompt is the route to a good output. But a good output is not automatically a usable output. Three separate questions, three separate areas of responsibility.

Am I allowed to enter this data?

A pupil's name, a report, a diagnosis, a parent's phone number. Whatever you type into the prompt box is no longer on your computer.

Who does this output leave out?

A model reflects the world of the data it was trained on. That world does not represent every pupil in your class equally.

Who is accountable for accuracy?

The answer never changes: whoever distributes it. That is you. "The model cited a source" is not the same as "the source was verified".

Key concept

Human oversight

Almost every international framework rests on the same principle: AI should not have the final word in a chain of decisions that affects a pupil. UNESCO's AI competency frameworks for teachers and students, published in 2024, and the Department for Education's guidance on generative AI in education converge here. The UK has no AI Act, but the same principle is already law: Article 22 of the UK GDPR gives a person the right not to be subject to a decision based solely on automated processing where it has a significant effect on them, and DfE guidance is explicit that a teacher remains accountable for anything they hand to a pupil. Translated into classroom terms: AI drafts, a human decides.

Interactive · Find the leak

What leaked in this prompt?

The prompt below is the kind a real teacher would write. The intent is entirely good, the problem is serious. Click the phrases that carry risk, then check your answer. Every underlined fragment is clickable and some of them are completely harmless.

A teacher's prompt · click to mark

Write notes for a parent meeting for my Year 9 English class. The pupil's name is Emily Carter, pupil number 1247. She was diagnosed with dyslexia last term and her educational psychologist's report notes reading difficulties. Suggest three activities that can be done at home to build reading fluency. The meeting is on Tuesday during the last period. Her mother's number is 07700 900000, so please draft a message for her as well. Our school is Riverside Secondary, class 9B. Keep the language plain and respectful, and do not blame the parent.

Your marks Correctly caught Missed False alarm

0 marks

The same prompt can be rewritten and still do the job: "Write notes for a parent meeting about a Year 9 pupil who struggles with reading. Suggest three activities to build reading fluency that can be done at home. Keep the language plain and respectful." Without a single name, number or diagnosis, the model still does its work. The rule is simple: describe the task to the model, not the person.

The technical reality

Where does the data go?

The word "cloud" hides this question. What you type travels to a server, is processed there and is usually retained for some period. Which account you signed in with changes that entire chain. Click the tabs.

Default behaviour

On free and individual plans, most providers reserve the right to use conversations to improve their services and train their models. There is usually a setting to turn this off, but it is rarely off by default.

The riskPupil data you type may end up in a training set, deletion requests become technically difficult, and processing may happen in another jurisdiction.

Appropriate useWork that contains no personal data: lesson plans, sample questions, material ideas, anonymous parent letter drafts.

Contractual behaviour

On enterprise and education plans, providers typically commit not to use your data for model training, sign a data processing agreement and define a retention period. The difference is not marketing, it is legal: the school now has a lawful basis for processing.

Still a limitA contract does not make it acceptable to enter pupil data you do not need. Data minimisation applies regardless of the account type.

Appropriate useSchool-approved workflows, working with class data, generating feedback on pupil work.

The most closed option

Setups where the data never leaves the institution. Strongest on privacy, heaviest on cost and maintenance. Rarely realistic for a single school to run alone.

Practical limitModel quality usually lags a step behind, the update burden sits with the institution, and you still need a usage policy.

Appropriate useLarge institutions handling sensitive data, assessment centres, shared systems at district level.

Practical advice: if you do not know which account your school works under, make that the first question at your next department meeting. Working with pupil data without knowing the answer is like driving off without checking the brakes.

Same task, two environments

The practical consequence of this distinction: some of what you learned as "never put that into AI" actually means "never put that into a general-purpose assistant on your personal account". On a platform your school has a contract with, one that does not train on your data and has a defined retention period, the same task can become workable. The only thing that never changes is who signs the decision.

Task	Personal account, general-purpose assistant	Platform your school contracts with
Lesson plan, sample question, material idea	Fine. No personal data involved.	Fine
Feedback on a pupil's work, with their name attached	No. Name plus work identifies the pupil directly and the school has no legal basis for it.	Workable. Inside a school account, a defined workflow and a data processing basis; the teacher still edits the output.
Class-wide marks and progress analysis	No. That means carrying your class list outside the school.	Workable. If the data already sits on that platform, no new transfer is created.
Rubric-bound scoring and feedback drafts	Risky. With no rubric to anchor it, output is inconsistent and its reasoning cannot be audited.	Workable. Each score maps to a criterion and the draft comes back to the teacher for revision.
EHCPs, educational psychologists' reports, diagnoses, health data	No.	Not automatically cleared. Special category data needs its own consent and purpose assessment; a contract alone is not enough.
Fixed in both environments	Data minimisation still applies, output still gets verified, and a human still signs any decision that affects a pupil

Six questions to ask of any platform

Which column a tool belongs in is decided by the answers to these six questions, not by its marketing copy. If you cannot find an answer in writing, the tool sits in the first column.

Is the data used to train models?

What you want is a clear commitment that it is not. A setting you can switch off is not the same thing.

Where is the data physically processed?

Which country, which legal framework, is there a transfer abroad. Under the UK GDPR this is the decisive question, and transfers out of the UK need their own basis.

Who is the controller, is there a contract?

With no signed data processing agreement between school and provider, there is no legal basis either.

Who can see it?

Is the list of people with access to pupil work defined, and are the provider's own staff on that list.

How long is it kept, how is it deleted?

Is the retention period in writing, can a deletion request actually be executed, what happens when a pupil graduates.

Is there an independent certification?

An audited standard such as ISO 27001 is not the same as a self-declaration. Check its date and its scope.

Open note

How Madlen answers these questions

Madlen produces this course, so answering the six questions for our own platform is on us. Teacher and student data is not shared with third parties, not sold, and not used to train AI models; only teachers and school administrators can access restricted content. The platform meets GDPR requirements, and for international school partnerships we work to meet local regulation too, data is processed on Infercom infrastructure on servers in Germany within EU borders and is not transferred to the US, and the infrastructure is certified to ISO 27001:2022.

None of that means "use Madlen and the privacy question is settled". A contracted platform does not cancel out data minimisation, and a human still signs any decision that affects a pupil. Ask us those six questions, and ask them of every tool you use; if you cannot find the answers in writing, do not put pupil data into it.

The legal basics

Data protection in four sentences.

The UK GDPR and the Data Protection Act 2018 run on the same logic, and the ICO's guidance for schools spells out what that means in practice. You do not need to be a lawyer; there are four principles a teacher needs to hold.

1 Purpose limitation ›

You may use data only for the purpose you collected it for. Information gathered for attendance cannot be uploaded to an AI tool with "analyse my class for me". The question to ask: if the person who gave me this data knew I was using it for this, what would they say?

2 Data minimisation ›

Work with the least data needed to do the job. That is the whole point of the redaction exercise above. The model produces the same set of activities without knowing the pupil's name, so leaving the name out is not only safer, it is the legally correct behaviour.

3 Consent and special category data ›

Health status, diagnoses, ethnicity, beliefs and biometric data are special category data and carry the highest level of protection. They cannot be processed or passed to a third-party service without an explicit lawful basis. Copying a sentence from an educational psychologist's report into a prompt is exactly such a transfer.

4 Retention and deletion ›

Data cannot be kept indefinitely; once the purpose is gone, it goes. Chat histories are a storage location too. Make it a habit to delete conversations about a pupil once the work is done. That serves both the spirit of the law and plain common sense.

Interactive · Decision tree

Am I allowed to enter this?

Instead of memorising a rule list, ask three questions in order. The tree below builds a reflex you can use every day. Pick an option and move to the next question.

Bias

A model repeats the world it was trained on.

In Module 1 you saw the model predict the next word by probability. Probability comes from frequency in the training data. If engineers in that data are mostly men and nurses mostly women, the model repeats that as if it were a fact. There is no malice, only statistics. But the classroom effect is the same: some pupils never see themselves in the material you hand out.

Where you will see it

In the names it picks, the occupations it assigns, the families it describes, the life it treats as "normal", the skin tones and disability representation in generated images, and the economic assumptions inside word problems.

What you can do

Write diversity into the prompt explicitly. Before you hand an output out, ask "who is missing from this". Have your pupils ask the same question; a bias hunt is a strong critical thinking activity in its own right.

Bias hunt.

The text below is a realistic output for the prompt "write three word problems about fractions for Year 6 maths". Click the problematic assumptions.

The model's output · click to mark

1. Sarah cut a pizza into 8 equal slices and while her mum was cooking and her dad was still at work ate 3 of them. What fraction is left? 2. Tom, who flew to Paris with his family for the summer holidays, read 3/5 of his 240-page book on the way. How many pages are left? 3. James, who wants to be an engineer, and Chloe, who wants to be a teacher, shared a project. James did 2/5 of the work and Chloe did 1/4. How much of the work is done? All three problems are presented as text only.

Your marks Correctly caught Missed False alarm

0 marks

Write three word problems about fractions for Year 6 maths. Represent different family structures and economic circumstances, do not let the occupations repeat gender stereotypes, and write at least one problem in short sentences suitable for reading aloud. Keep the contexts to ordinary situations at school and in the neighbourhood.

Interactive · Accountability

How much scrutiny does each output need?

Checking every output with the same rigour is unrealistic; checking none is dangerous. The dividing line is the size of the risk: how directly does an error touch a pupil? Choose a level of scrutiny for each task below.

Brainstorming, a list of activity ideasRaw ideas produced for your own use

Correct. For ideas nobody else will see and that pass through your own filter, a quick skim is enough. You are going to discard the weak ones anyway.

A parent letter or announcement draftText leaving the school in its name

Correct. Tone, dates, times and names have to be read line by line. Whoever signs a letter owns it; a date the model invented becomes your mistake.

Dates, figures, formulas and quotations in a lesson textInformation the pupil will treat as true

Correct. Remember hallucination from Module 1: a model can produce wrong information in a confident tone. Every figure, date and quotation that reaches a pupil must be verified independently.

Grades, report comments, setting or placement suggestionsDecisions that shape a pupil's future

Correct, and this is the most critical row. This is exactly where UK GDPR Article 22 and DfE guidance draw the line: a decision with a significant effect on a pupil cannot rest on automated processing alone. The model may draft, the teacher forms the judgement, and the decision rests on the pupil's actual work.

Bibliographies, references and lists of studiesSources claimed to exist

Correct. Fabricated citations are the best-known failure mode of language models. The author and year can look entirely real while the paper itself never existed. Never use a reference you have not searched for yourself.

Instructions for a classroom game or warm-upUsed in class but carries no grade

Correct. Read it end to end for feasibility, timing and age suitability, but academic verification is not needed. An activity that collapses in class is a recoverable mistake.

Transparency

Copyright, attribution and open declaration.

The legal status of AI-generated material differs by country and is changing fast. The professional stance, however, has settled: if you used it, say so.

When you use it

Add one line under your material

Name the tool and the stage you used it at. This is not a confession, it is professional ethics. It also models the behaviour you want: pupils see an adult who does not hide their sources.

When a pupil uses it

Make the declaration part of the task

As Module 3 showed, detection tools are unreliable. The fix is not to punish concealment but to normalise declaration: which prompt did they write, how did they change the output.

A declaration template you can reuse

The first draft of this material was produced with [tool name], edited by the teacher against the learning objective, and the dates and figures in it were verified against [source]. Responsibility for the final version rests with the class teacher.

Age and consent

Who can use it, at what age, and how?

The answer to this does not start with age, it starts with the tool. For a pupil of any one age there are two entirely different situations: on one side a free-form conversation with a general-purpose assistant on their own account, on the other a school-provided education tool with its subject and content bounded, assigned by the teacher. The first column's answer changes as pupils get older; the second is already constrained by design, so it can open earlier.

The distinction

Raw model or bounded tool

General-purpose assistant (raw interaction): The conversation is open to any subject, nobody bounds the output in advance, the history accumulates in the pupil's own account, and the provider is the processor. This is where provider terms typically exclude under-13s entirely and require parental permission under 18.

School-provided education tool (bounded interaction): The task and subject are defined up front, output is produced inside the frame the teacher sets, the school opens the account, and the contract names the school as controller. Here what governs is less an age limit than the contract the school signed and the information given to parents.

The gap between the two is pedagogically real as well: a primary-age pupil holding a conversation with an unbounded assistant is not the same thing as that same pupil working on an exercise their teacher assigned. The table below keeps the two in separate columns.

Key Stage	General-purpose assistant, pupil's own account	School-provided bounded tool	Parent communication
Primary, KS1-2	No. The terms do not allow it and pedagogically there is no need.	From the teacher's screen, one screen at the front; where content is assigned, under teacher supervision and in class	Information is enough
Secondary, KS3	Usually no. Even where terms open up above 13, individual use should be tied to school policy.	On a school account, group work under supervision; outputs reviewed together in class	Written information and permission
Secondary, KS4-5	Yes, within stated rules. Use is declared and task design is explained with an AIAS level.	Individual use can open up; the teacher defines the task and the assessment criteria up front	Written information
Fixed at every Key Stage	Whichever the tool: no personal data entered, outputs verified, a human signs any decision that affects a pupil		Policy shared

Work out for yourself which category your tool falls into: is the conversation open to any subject, whose account holds the history, who does the contract name as controller? Terms change often, so always confirm the current rule on the tool's own page; the table above is a pedagogical recommendation.

Interactive · Department discussion set

Five situations without a clean answer.

Some of the questions in this module have no single right answer; it depends on your school's context. Pick a side in each of five situations and see that side's strength and its weakness. The aim is not to talk you into a position, it is to prepare the discussion you will open with your department.

Common mistakes

Five common ethical mistakes.

1 "I only used the initial, so it is anonymous" ›

Why it failsAn initial plus a class, a gender and a diagnosis leaves exactly one person in a class of thirty. That is indirect identification.

The fixRemove the person entirely. "A lower secondary pupil who struggles with reading" does the job.

2 "The model gave a source, so it must be right" ›

Why it failsThe shape of a citation is generated too. A plausible author, year and journal can belong to a study that never existed.

The fixSearch for every source by name. If you cannot find it, treat it as non-existent and cut it.

3 "I ran the essay through a detector and it flagged AI" ›

Why it failsRemember the figures in Module 3: detectors mislabel authentic writing and regulators have warned against relying on their scores. On top of that, uploading a pupil's essay transfers their data to a third-party service.

The fixNever accuse on the basis of a detection score. Ask about the process instead: drafts, prompts, reasoning.

4 "I left the final grade to the AI, that is more objective" ›

Why it failsThe failure is not AI taking part in scoring, it is the decision going unsigned. Hand a raw model an essay with no rubric and ask it to grade, and it will look consistent while consistency is not accuracy: it can score the same text differently at different times and invent a rationale afterwards. Assessment that affects a pupil is also exactly where UK GDPR Article 22 and DfE guidance draw the line, so a human signature has to sit behind the decision.

The fixReverse the order. Define the criteria first, then use AI for scoring and feedback drafts bound to those criteria. The same distinction applies here: a tool that works against your rubric, maps each score to a criterion and leaves the revision to the teacher is not the same thing as asking a free-form chat "what would you give this?". Whichever tool it is, the last step is fixed: check the score against the pupil's actual work, then award the grade and the reasoning yourself.

5 "Nobody at school said anything, so it must be allowed" ›

Why it failsThe absence of a policy is not permission, it is uncertainty. When something goes wrong, the teacher who acted pays the price of that uncertainty.

The fixTake the draft you generated above to your department. A written, shared rule protects both the pupil and you.

Check yourself

Fill in the blanks.

Working with the least data needed to do the job is called . Health and diagnostic information counts as personal data and never belongs in a prompt. In decisions that affect a pupil's grade or placement, the final word belongs to . A bibliography produced by a model must always be .

You want to build an individual support plan using the wording of an educational psychologist's report a parent sent you. In a general-purpose assistant you signed into with your personal account, what is the right thing to do?

An educational psychologist's report contains special category personal data, and even with the name removed the content still belongs to the pupil. On a personal account there is no data processing agreement behind you and you are not the one deciding where the data goes or how long it is kept, so the right route is to never enter the report at all: describe the need in anonymous, general terms and shape the draft with your own professional judgement. The answer on a platform your school contracts with can be different. Special category data is not automatically permissible there either, but the decision moves from a flat "never enter it" reflex to an assessment of the contract's scope, parental consent and the retention period. The next section opens up that distinction.

Module 4 complete.

You now know which data never goes in, how much scrutiny each output deserves and which rule you will defend at your school. One module left: we have drawn the boundaries, now we bring the pupil to the table.

Module 5: AI in the classroom → All modules