Lesson planning with AI: teachers, keep your students out of it
A six-lesson unit, exercises at three levels, an assessment: AI can prepare all of it in one evening. What goes with it if you get it wrong, and how to get it right.
"I did my whole French Revolution unit with AI last night, and I even gave it the support arrangements for my Year 9s so it could adapt the exercises." The remark is made in the staff room, between two rounds of photocopying, and it draws envious looks: six lessons mapped out, worksheets at three levels, an assessment with its mark scheme, all in one evening.
The preparation is a success. What is less so is what went with it: the list of a class of minors and, for some of them, a learning difficulty, an individual support plan, a family problem mentioned in passing to explain late work. Here is the same preparation done twice: first the way it is often done, then the way it should be done.
What school rules require for student data
Information about your students does not belong to you: it is entrusted to you so that you can teach. It is personal data about minors, and when it concerns a learning difficulty, a disability or medical follow-up, it is health data, the most protected category of all. In the European Union, the GDPR applies, as does its UK counterpart in the United Kingdom, and teachers in public service are often bound by a duty of professional discretion; in Switzerland, the Federal Act on Data Protection (FADP) and, for state schools, cantonal laws play that role.
On top of that come the rules of your school or education authority: approved tools, guidance on the use of AI, an IT acceptable use policy. They come first. If you are not allowed to use AI to prepare your lessons, the right move is to raise it with your head teacher, not to work around the rule.
There is good news, though: building a unit of lessons requires almost no personal data. The curriculum, the topic, the class's year and level, and the time available are enough. The risk almost always comes from giving too much context.
Framing the unit of lessons before you ask for it
Before: the conversation that starts from scratch every evening
On the first evening, you paste in the curriculum extract, describe the class in detail and ask for six lessons. On the second evening, the previous conversation is lost or mixed up with the planning for another class, and you start again: the class's year and level, the length of a lesson, the textbook you use, the way you like instructions to be worded. To save time, you end up pasting the same full description of the class every time, names included, because it is already written somewhere.
After: one space per unit, instructions written once
In IA Confidential, the teacher creates a space called "Year 9 History – French Revolution". In the space's Assistant tab, she writes down once and for all what does not change: the class's year and level, the length of a lesson, the expected outcomes at the end of the unit, the format of her worksheets (one instruction per line, a worked example at the top), and her working rule: "Never ask me for students' names; think in terms of levels of need." These instructions apply to every conversation in the space and take priority over her general preferences.
The class is described there without anyone in it: "twenty-six students, a third read slowly, several need instructions broken into steps, two are working well ahead of the rest". That is all the AI needs to pitch the work at the right level, and it does not allow anyone to recognise a single child.
To make sure she is in line with the official text, she pastes the link to the curriculum page into her question: the assistant reads the page from the link. She still checks it against the version currently in force before building on it.
Differentiated exercises without student profiles
Before: three first names and their support arrangements
To adapt the exercises, the temptation is to say everything: "Léa is dyslexic and gets extra time, Yanis has an attention disorder, Maëlle has an individual support plan." That message has just passed the health data of three children to a provider that neither the families nor the school chose, and which often keeps the history. And it added nothing: the AI does not need to know who has a learning difficulty, only what kind of adaptation to produce.
After: learning needs, not diagnoses
Differentiation is described by what the exercise must offer, not by the person who will receive it: a shorter text, a more spacious font, instructions broken into steps, a visual aid, an extension version. It is the teacher who then hands each version to whoever needs it. Which student gets which arrangement stays in her planner, in the school's management system and in her head.
Here is the request she makes, to copy and adapt:
For lesson 3 of the unit (the storming of the Bastille and its aftermath), suggest a source analysis exercise in three versions. Version A: a ten-line text, closed questions, one instruction per line, a sample answer at the top. Version B: the full text, short open questions. Version C: two sources to compare and a synthesis question. For each version, give the answers and an estimated time. Make no assumptions about the students who will receive each version.
The last instruction matters: it stops the assistant from labelling a version "for students who struggle", a heading no student should ever read at the top of their worksheet.
Your teaching resources: what leaves and what stays
Before: the textbook chapter and the shared test pasted as they are
To make the exercises match the class textbook, you photograph or copy two pages of the chapter. To prepare the students, you also paste in the common assessment written with the other teachers of the same year, which students will not sit for another three weeks. This is not personal data, but these are not texts to pass around either: a textbook is a copyrighted work, a shared test does not belong to one teacher alone, and once they have been sent, you no longer know what becomes of them.
After: the teaching folder stays on the computer
On a computer, with Chrome, Edge or Opera, the teacher adds to the space the folder where she keeps last year's lessons, worksheets and assessments, ticking the "Confidential models" option: the folder is never sent to an external model. Once she is signed in to her account, the assistant finds the relevant passages there when she asks it, for example, to reuse last year's timeline or to match the vocabulary of her worksheets. The search index stays in the folder, on her computer. She chooses the permissions for each folder: search and read for her lessons, never write access if she does not want it.
On a tablet or in another browser, she simply attaches the document to her message: it is read in the browser. The confidential models, open models installed on servers in Switzerland, are more than enough for this kind of formatting work and pass nothing on to third parties. As for the shared test, it does not need to leave its folder: she asks for practice exercises "of the same type", describing the skills being assessed.
The end-of-unit assessment and student scripts
Before: the most powerful model, the scanned scripts
For the detailed answers to the assessment, you pick an external model reputed to be stronger, and feed it a dozen scanned scripts to "check that the mark scheme holds up". The scripts carry the student's name, the class, sometimes a comment in red ink about behaviour. Everything goes as it is.
After: the Confidentiality filter, and scripts that do not need to leave
To test a mark scheme, a few model answers are enough, written or copied out without names: a complete answer, a partial answer, a common misunderstanding. The confidential model handles that very well.
If she chooses an external model all the same, the Confidentiality filter steps in before anything is sent: it detects sensitive data in the message and offers to answer with a confidential model, message untouched; to send an anonymised version that she can reread and correct; or to send it as it is, knowing the risk. For an attached document, only pseudonymised extracts leave: names are replaced by placeholders, then restored in the answer she reads.
The assessment, the worksheets and the answers are written by the assistant as documents: she rereads them, edits them, prints them to PDF or exports them to Word or OpenDocument to finish them in her word processor. The layout is rebuilt from the text, not taken from her usual worksheet template: she adds the school header and logo herself.
What AI does not do in the teacher's place
A unit produced by an AI looks finished. It is not. The assistant can get things wrong and make things up with confidence: a date that is off, a quotation attributed to the wrong revolutionary, a "historical" source that does not exist, maths answers that go wrong on the third line, an unrealistic lesson timing. Do every exercise yourself before you photocopy it, check every source cited, and read the instructions through the eyes of a thirteen-year-old.
Nor does it know your class. The real level, the atmosphere on a Thursday afternoon, the student who switches off as soon as an instruction runs to more than two lines: your judgement makes the adjustments. The sequence of learning, the choice of sources and the assessment remain your professional responsibility, and carry your name.
What stays confidential
With a confidential model, your questions, your attached documents and the answers are processed on servers located in Switzerland, and no file is kept on our servers. A folder added with the "Confidential models" option is never sent to an external model.
Your conversation history lives in your browser, not with us: we could not produce it if asked to, because we do not have it. There is a trade-off worth knowing: there is no synchronisation between your computer at home and the one at school. To find your units again from one year to the next or from one computer to another, back up your history to a file, which you can restore without duplicates.
With an external model, nothing leaves without going through the filter, and you decide what goes.
One precaution specific to schools: in the staff room, the library or the computer room, the history can be read by anyone using the same browser session. On a shared computer, use a personal session and close it when you leave. And if you have followed the rule of needs rather than first names, whatever is shared does not allow any student to be recognised in any case.
For the habits that apply in every profession, read our list of data you should never paste into an AI and our guide to using AI at work safely. To find out exactly where your exchanges go: what happens to your conversations.
To plan your next unit for an imaginary class: open the chat.
