AI in the classroom: what it should do, and what it shouldn't
Why the teacher leads, what the evidence says, and the limits teache keeps.
Francisca Akoto · October 27, 2026 · 10 min read
AI has reached Ghana's staff rooms. Teachers ask chatbots for lesson ideas on their phones, and apps write a week's lesson notes in minutes. The question now is not whether teachers will use AI. It is what AI should be trusted to do in a school, and what it should be kept away from.
We build an AI tool for Ghanaian teachers, and our answer is this: AI is a tool the teacher uses, and the teacher leads the class. AI can take on much of the paperwork around teaching. It cannot run a lesson, and a tool that claims it can leaves the teacher to answer for the gap.
In short
- AI prepares the documents. The teacher decides what goes in them, and teaches the class.
- teache leaves the remarks on how a lesson went to you, writes activity lines rather than a script, and doesn't pretend a written paper can test everything.
- It says plainly that it can get things wrong, and you are never charged for a document it fails to make.
- Pupils don't talk to it, your data is not sold, and the AI company we use does not train its models on your work.
AI can't see the classroom
A lesson note is written days before the lesson. When teache drafts one, it knows your subject, your class, the curriculum for the week and whatever you tell it. That is enough to write a good plan. It is not what decides how the lesson goes.
An example
Take a Basic 5 Mathematics lesson on comparing fractions. The note is good and your headteacher has signed it. Ten minutes in, you ask the class which is bigger, a half or a quarter, and most of them say a quarter, because four is bigger than two. The fraction cards you planned for the main activity would now build on a mistake. Most teachers would put them away, cut an orange in front of the class, and spend the rest of the period on that one idea. No plan could have made that call on Sunday, because nobody knew until a pupil answered.
A lesson is full of calls like that: whether the class has understood or needs one more example, who to ask next, when to stop and go back. They rest on things only the teacher knows. You know who was absent last week, which pupil goes silent when she is put on the spot, who came to school without eating, and that the back rows can't hear you when it rains on the roof. None of that is written in a curriculum or a planning page, so none of it reaches the AI.
Better AI won't change that
The obstacles have little to do with how capable the tool is.
- It isn't in the room. Even a far better planning tool would know only what you typed days before. Everything that decides the lesson happens later, in front of the class.
- A class follows a person. Pupils settle down, have a go and risk a wrong answer because they trust the adult in front of them. When you change course halfway through a lesson, they come with you because it is you asking. A screen telling a class what to do next does not earn that kind of trust.
- Someone has to answer for the lesson. To the headteacher, to parents, and to the child who still doesn't understand fractions. That person is the teacher. If you answer for the lesson, the decisions inside it have to be yours.
AI does better with a teacher in charge
In the 2023/2024 school year, researchers from the University of Pennsylvania gave nearly a thousand high school maths students in Turkey access to GPT-4, the AI behind ChatGPT at the time, during practice sessions, and published the results in the journal PNAS in 2025. One group used a plain chatbot. Another used a version built with the school's teachers, which held the teachers' notes on common mistakes and gave hints instead of answers. A third group had no AI at all.
| Group | Practice, with the AI | Exam, without the AI |
|---|---|---|
| No AI | The comparison group | The comparison group |
| Plain chatbot | 48% better | 17% worse |
| Teachers' version | 127% better | About the same |
The researchers' explanation for the drop is that students used the plain chatbot as a "crutch": it did the thinking, so they learned less. The teachers' version largely avoided that drop, although it did no better on the exam than having no AI at all. The plain chatbot was not reliable either: in the researchers' working paper, it gave the correct answer to the practice problems only 51% of the time.
A study in Nigeria found real gains, and the teacher was at the centre of it. In 2024, a World Bank team studied a six-week after-school English programme in Edo State in which secondary students worked with an AI tutor. In the World Bank's own account, "each session began with a teacher introducing a topic", and teachers, trained beforehand, kept students on task. The World Bank describes the students' gains as "roughly equivalent to 1.5 to 2 years of typical learning progress in similar settings". The authors are careful about what that proves: "We cannot isolate the effect of the chatbot alone." The result belongs to the whole programme, the teachers included, and it was only six weeks long.
Neither study left the AI to run a class. The evidence is still early, but it points one way: AI working under a teacher, not in place of one.
What UNESCO, and Ghana, already say
International guidance says the same. In 2024 UNESCO, the United Nations agency for education, published an AI competency framework for teachers. It is blunt:
"Teachers should not and cannot be replaced by technology."
UNESCO, AI competency framework for teachers, 2024
It adds that "teachers should remain accountable for pedagogical decisions", and that AI tools "should never be designed to replace the legitimate accountability of teachers in education." UNESCO's earlier guidance on generative AI in education warns that the technology "should not be allowed to usurp human thinking", and recommends 13 as the minimum age for children talking to AI on their own.
Ghana's own planning formats already leave the teacher in charge. The Weekly Learning Planner in the SHS teacher training handbooks has a box for differentiation: what you will do for learners approaching proficiency, for those who are proficient, and for those who are highly proficient. A plan can describe all three groups, but only the teacher in the room can see which pupil belongs in which group today. The same planner ends with Reflection and Remarks, a box that can only be filled in once the lesson has been taught.
How this shapes teache
If the teacher leads, then AI belongs in the work around the lesson, not in the lesson itself. Much of a teacher's week is spent away from the class on work that is necessary but is not teaching: counting the weeks in the term, finding the right indicators and their codes, copying them into the right boxes, laying everything out the way the headteacher expects, writing a fresh test, filling in the record of work. That is the work teache takes on, and in each document it leaves the decisions with you.
| Document | What teache does | What stays with you |
|---|---|---|
| Lesson note, KG to SHS | Drafts the activities from the curriculum for your class | What your class actually does, and the teaching |
| Scheme of work | Paces the term from the curriculum, so the counting is done | The order you teach in, and the extra week when your class needs it |
| Practice or class test | Sets questions and an answer key from the indicators you tick | Reading every question, checking the answers, and the marks |
| Record of work | Starts the table from your lesson notes | Changing it to show what you really covered |
A test can't know which indicators you actually reached, how you explained them or which words your class has met, so the marks are your judgement, not the software's. A record of work is evidence of what happened, not a copy of the plan. And the order of a scheme is yours to set, as we explain in do you need a scheme of work.
Our commitments
A tool that writes documents teachers sign, teach from and put in front of children has to hold itself to clear limits. These are ours. Some are built into how teache works and the rest are written into our public pages, so you can check each one rather than take our word for it.
We won't write a lesson that hasn't happened
On every note, the Remarks box is left empty for you to fill in after teaching. For KG and SHS notes, teache's writing rules go further and say it directly:
Never write how a lesson went: the teacher writes reflection and remarks after teaching.
teache also looks out for lines such as "most learners were able to..." in its KG and SHS drafts. A signed note is a record, and a tool should never put you in the position of signing a record of something that has not happened. If a draft ever reports a lesson before you have taught it, take the line out.
We won't write your words for you
teache writes short activity lines: what learners do, and what you set up and check. It is not built to write a script for you to read aloud. You already know how to talk to your class, and a script written for an imagined class would only get in your way.
We won't pretend a paper can test everything
Some of the curriculum is about doing things, and a written paper can't test it. When you choose what a test covers, teache marks indicators like these "Class activity only" rather than inventing a written question that only looks as if it tests them. You assess those parts the way you always have, by watching your pupils do them.
We won't pretend it is never wrong
The 51% figure above came from an older model than the ones in use today, but no AI is right every time. It can get a fact wrong, pitch a task too high or put a wrong answer in a marking key. We say so on how our AI works and in our guides, and every document stays yours to read and change. We would rather you trust teache a little less and check it than trust it completely and be caught out in front of your class.
We won't charge you for something you didn't get
You see the exact price before anything is written, and if teache can't make what you asked for, it tells you and you are not charged.
We won't put pupils in front of a chatbot
teache is built for teachers, who must be adults to use it. Pupils don't sign in and don't talk to the AI. What reaches them is a test or an activity their teacher has read and chosen to use.
We won't sell your data, or let our AI provider train on your work
What you write goes to the AI only to produce your document. Our privacy policy names the company that runs the AI, says that company does not use your work to train its models, and says what it may keep and for how long. We don't sell your data to anyone.
We won't sell the idea that AI replaces teachers
teache is not built to argue that schools need fewer teachers, or that anyone with an app can stand in front of a class. It is built to take paperwork off your evenings so that more of you is left for the lesson. A tool for teachers should make the teacher's position stronger, not quietly argue it away.
What we ask of teachers
Our limits only work if teachers keep their side too. We ask five things of everyone who uses teache:
- Read everything before you sign it, print it or hand it out.
- Check every answer in a marking key before you mark with it.
- Teach the class in front of you, not the draft.
- Write your remarks honestly, including when a lesson didn't go to plan.
- Keep pupils' names out of what you type. Children can't give their own consent to having their names put into software, and initials do the job.
Treat a teache draft the way you would treat a draft from a capable colleague: useful, often right, and still yours to check before your name goes on it.
The teacher leads, the tool follows
The teacher sees the class, makes the calls and writes down what happened: the remarks, the record of work, what the class still needs. The tool takes that and does the paperwork, so next week's draft starts from where your class really is, not from where the curriculum assumed it would be.
That is how we think AI in the classroom should work. If you ever find an AI tool, ours included, trying to work the other way round, ask why.
Sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26). Earlier working paper on SSRN.
- World Bank (25 June 2025). Addressing the learning crisis with generative AI: lessons from Edo State in Nigeria.
- UNESCO (2024). AI competency framework for teachers.
- UNESCO (2023). Guidance for generative AI in education and research.
- T-TEL (2023). Introduction to the SHS/SHTS/STEM Curriculum, Professional Learning Community handbook, Session 12 and Appendix 12.