AI for Teachers

The school had forty-one teachers and a computer room with two working machines. Every teacher had a phone. Perhaps half had data that lasted to the end of the month. When the principal asked Sunita Verma, the science head, to recommend one tool for the staff, the question looked like a comparison of models and turned out not to be.

There were two realistic candidates. The first was the assistant inside WhatsApp. It needed no new account, no new app and no email address, it ran on the data plan people already paid for, and the staff group where timetables were shared was already open on every phone. The second was ChatGPT or Gemini in the phone browser: better on long, careful work, better at holding a pasted syllabus, and requiring each teacher to create an account and then walk through a settings menu to find the control that stops conversations being used to train the model.

Mrs Verma tested both for a fortnight, on the work she actually had to do. Most of what she found was reassuring. Both explained photosynthesis and the water cycle well enough to use. Both, asked what Class 8 students get wrong about the water cycle and what they are thinking when they get it wrong, produced eight answers, two of which she had not met in nineteen years. That single request turned out to be the most valuable thing either tool did all fortnight, and it cost nothing but a sentence.

Both also produced a Class 8 worksheet on the district’s part in the freedom movement, fluently, with a founding year for the town’s oldest school, and the two years did not match. She checked the school’s own foundation stone. Neither was right. She noticed, separately, that the answers she got at nine in the evening were thinner than the answers she got at six in the morning, which she assumed was her own tiredness until a colleague reported the same thing.

All of that was useful and none of it was the decision. The decision was the setting.

In the browser tool she had found the control in four minutes and switched it off. In the WhatsApp assistant she looked for twenty minutes, could not find an equivalent, and could not get an answer from anyone at the district office about whether one existed. She knew what teachers in that staff group already did, because she was in it: they forwarded photographs of student scripts to each other and asked for help marking them.

The two costs

Recommending WhatsApp meant something close to full adoption. The eight teachers with the oldest handsets could use it. The five who had never had an email address could use it. It would be in the staffroom by Monday, and the material generated with it (worksheets, explanations, question lists) would be real and useful, because the tool that gets used is the tool people already have open.

It also meant she would be recommending, to forty-one colleagues, a tool she could not tell them was safe to paste a child’s work into. She could write a rule. She could not enforce one.

Recommending the browser tool meant a better position on data and a worse one on everything else. Her estimate, from the two afternoons she had spent helping colleagues make accounts, was that perhaps twelve of the forty-one would complete the setup and keep using it. The rest would carry on as before, and the tool would become one more thing that the better-connected half of the staffroom had and the other half did not.

There was a third option she considered and dropped. A laptop in the office could run a small open model with nothing leaving the machine, which answers the data question completely. It also meant one machine, in a room that is locked at four o’clock, running a model noticeably weaker than either candidate, for a staff of forty-one who plan their lessons at night. It is the right answer for a school that cannot send any data out and the wrong answer for a teacher with a phone, and this was a school of teachers with phones.

She had one staff meeting, forty minutes, and a principal who wanted a single answer.

Adaeze Okonkwo had been teaching biology for eleven years and had used a chatbot twice, both times badly. The third time she asked for “a revision sheet on the circulatory system for Grade 11” and got a sheet about the cardiac cycle, ECG traces and blood pressure regulation, with the word “sophomore” in the introduction and none of it inside the WAEC syllabus.

What fixed it took twenty seconds. She pasted the six lines of the WAEC topic list for the circulatory system, one past question with its marking guide, and a line that said SS2, Nigerian secondary school, British spelling, and use only the topics listed. The next sheet stayed inside the six lines, phrased its questions the way WAEC phrases them (state three functions of, describe the passage of blood from) and allocated marks the way the guide did.

She kept the prompt, the pasted material and the output in a text file. Not the prompt alone: she had learned that the hard way in September, when a prompt that had produced a superb set of distractors in the morning produced ordinary ones in the evening and she had no copy of the good version to show it. By half term the file had nineteen blocks in it, each with a date, the tool, and a line saying what had needed fixing.

Three of the blocks were worth more than the rest. An examiner block, which named the board and produced questions with real command words and a mark scheme with one line per mark. A confused-student block (you are an SS2 student who has just learned this and holds one common misunderstanding) which handed her three realistic wrong answers a week before she met them in a pile of scripts. And a levelled-passage block for the eleven students in her class reading two years below the rest, which produced the same content at two reading levels with every fact held constant.

The head of department, told about the file, wanted it in the shared drive by Monday. Three new teachers were starting in January and none had taught a WAEC class.

What was in the file

One teacher, Emeka, read it and objected to two things, and both objections were reasonable.

The first was that the file contained a colleague’s own questions. Adaeze had pasted the best mock paper the department had, written by a teacher who had since retired, as an example of the shape she wanted. It was in the file, in full, under PASTED.

The second was harder. That mock paper was still in use. The department set it every February. Pasting it into a free consumer chatbot meant it had been sent to a company’s servers, and although Adaeze had switched off the training setting on her own account in October, eleven of the nineteen blocks predated October.

Adaeze’s argument was that the induction value was real and immediate. A new teacher who pasted the department’s calibration on her first day inherited eleven years of knowing what WAEC asks for, instead of spending two terms discovering it, and the students taught by that teacher in her first term are the ones who benefit. Emeka’s argument was that a department cannot ask students to be careful about what they paste and then paste its own live assessment, and that the department was three months from teaching a lesson on exactly this.

The head of department made a third point that neither of them wanted. Removing the example would weaken the file, and a weakened file is one nobody uses, and a department that stops sharing its prompts goes back to five people generating the same eight questions every February. He was not defending the paste. He was saying that the safe version had to still be good enough to use, or the safety was theoretical.

He had a staff meeting on Thursday and three new teachers arriving in six weeks.

Rekha Bisht taught 46 children in Class 6 and had been differentiating by hand for nine years, which meant in practice that she did not. There was one worksheet and there were three groups of children, and the bottom group copied.

In November she spent a Sunday learning to name the axis. Not “make it easier” (which had given her, the first time, the same eight percentage questions with smaller numbers and the same barrier still standing) but the three that the course had named. A support version with the same 800-rupee shirt and the same fifteen per cent, and four numbered blanks that walked through ten per cent, then half of it, then the discount, then the price. A representation version with a bar model instead of a paragraph. An extension version where the shopkeeper had paid 600 rupees, wanted to keep 120 rupees of profit, and needed the largest discount he could offer.

All three were about the same shirt. That was the point, and when she read them back she could see her class staying in one conversation for the first time rather than splitting into a group doing real work and a group doing smaller sums.

She had also stopped asking for four options and started asking for the errors. Not “write four multiple choice options” but: the correct answer, plus three wrong answers produced by taking fifteen per cent of the wrong number, by subtracting fifteen instead of fifteen per cent, and by finding the discount and stopping there. Her tally of wrong answers on Monday morning would now tell her what each child had done, which is a teaching plan rather than a mark.

Generating all three versions took eleven minutes. She had the prompt, she had the axis in it, she had told it the shop was a shop and the money was rupees, and the model did what she asked.

The rule she could not afford

The course had been unambiguous about the next step: work the sheet yourself, before you print forty copies. Fifteen minutes. AI worksheets fail confidently and consistently within themselves (word problems whose numbers do not resolve, matching exercises with two items that fit the same answer, an answer key that contradicts its own question) and none of it is visible from reading.

It was twenty past eleven. Her own daughter’s fever had taken the evening. The photocopier at school was reliable until about half past seven in the morning and then queued. She had time to work one of the three versions properly.

The obvious choice was the extension version, because it was the one with the interesting arithmetic and the one most likely to be wrong. Four children would get it, all four of whom would notice an error, tell her, and be fine.

The support version had the simplest numbers, which felt like the safest sheet in the pile. It was also going to eleven children who had spent two years being handed work they could not do and had learned to stop when something did not come out, without telling anyone. If step two of that scaffold pointed at the wrong number, they would not report it. They would decide, again, that they were the sort of person for whom maths does not work.

There was a version of this decision she had already got wrong once. In September, in a hurry, she had printed a generated matching exercise without doing it, and two of the six items fitted the same answer. The class found it in four minutes and the lesson was spent on the sheet rather than on the mathematics. That failure had cost her a period and nothing else, because a matching exercise fails loudly. A wrong step in a scaffold does not; it fails quietly, in the book of a child who does not tell anyone.

She had one fifteen-minute block and three sheets.

The arithmetic had never worked. Sixty-two essays at twelve minutes each is over twelve hours, which meant essays came back in the third week, by which point the class had moved on and the comments were read for the grade and not for the comment. Joseph Mwangi had run that cycle for six years and knew exactly what it was worth.

So when he built an observable rubric (quotes at least two specific phrases and explains one effect of each; every claim followed within two sentences by a reference; each paragraph opening with the sentence that states its point) and found that a model applying it could produce two specific, quoted, actionable points on a draft in about four seconds, the prize was obvious. Not marks. Feedback on Tuesday for a redraft on Thursday, which the research says is worth several times the same words delivered after the grade.

He was careful about two other things first. The rubric went to the students in the same words it went to the model, printed and stuck in the front of the exercise book, because a criterion that tells a marker what to look for tells a writer what to do. And no drafts left the room with a name on them: each was coded, the codes were on a sheet in his drawer, and the coding took four minutes for sixty-two.

Then he ran the check the course insists on. Five scripts he had already marked, marks hidden, same rubric.

It ranked four of the five the way he had. The fifth was Amina’s.

What the fifth script showed

Amina wrote in her third language. Her essay was 340 words, plain, with two quotations, each connected to a claim, each with the effect on the reader explained in a short sentence. Against the rubric it was a clear top band and he had given it one. The model put it two levels below a 900-word script from a fluent writer that had one quotation and a great deal of confident paraphrase.

He swapped in a short strong script and a long weak one from the previous term and ran it again. The ranking did not survive. Whatever the model was reading, it was not the criterion.

That left him with a decision he did not want. Abandoning the tool meant going back to the three-week cycle, which he knew, from six years of evidence, delivered almost nothing to anybody, and it would fall hardest on the students who most needed a second draft, including Amina. Using it meant putting a marker with a documented bias against second-language writers in front of a class that was almost entirely second-language writers.

His head of department, who wanted the tool used across the department by January and had a workload paper to write, thought he was overreacting to one script out of five. Four out of five, she pointed out, is better agreement than two of her markers manage on a moderation day.

His deputy head took the opposite view and it was not a foolish one: a marker with a documented bias against second-language writers, in a school where nearly every student is a second-language writer, is not a tool with a flaw, it is a tool aimed at the wrong class. He wanted it dropped and the three-week cycle defended as honest.

Joseph thought both of them were answering a question he had not asked. Neither had said what should happen to the sixty-two students in the meantime.

The essay was on Silas Marner and it was good. It was better than anything Fathima had submitted that year and it argued a line (that Eppie’s arrival is written as an economic event before it is written as an emotional one) that had not come up in class.

A colleague, unprompted, had run it through a free detector and forwarded the result to the head of department and to Priya Nair, who taught the class: ninety-two per cent AI-generated. By the time Priya read the message, four members of staff had seen the number, and Fathima’s father had emailed the school asking why his daughter had been called out of a lesson.

Nobody had called her out of a lesson. She had gone to the toilet. But the email had arrived, and it made the case urgent in a way the evidence did not.

What the evidence was

The class policy, written with the students over one period in August and signed by all of them, said the first draft of any assessed essay was Level 0 and named exactly what the consequences were: the work is redone under supervision, there is no mark penalty on the redone work, and a second occasion brings a note home. The students had been stricter than Priya expected about which tasks counted as the exercise, and more precise than she expected about which tools they actually used.

It also contained a line Priya had insisted on and now had to live up to: I will not use AI detectors.

What she had was a version history, because the essay had been written in the school’s shared documents. It showed the essay arriving in four large blocks over two evenings, which is what a paste looks like and also what a student who drafts on her phone and pastes looks like. It showed no small edits at all.

She also had an in-class paragraph. Every assessed piece in her class began with twenty minutes of writing in the room, and Fathima’s opening paragraph, in her own hand, contained the economic reading of Eppie’s arrival in almost the words the essay used.

The head of department wanted a decision by Friday and thought the ninety-two per cent settled it. He was not being unreasonable; he had a deadline, a parent, and a number that looked like evidence.

Priya’s position was that the number was not evidence, that the tool it came from is known to flag careful second-language writing, that Fathima writes in her third language, and that the version history was ambiguous in both directions. She also knew a piece of arithmetic that nobody in the staffroom had done: a detector with even a one per cent false positive rate, run over every piece of homework from thirty students, produces a wrong flag against an innocent child about one week in three.

And there was the part of the case that had nothing to do with evidence. Four members of staff had seen the number before the teacher of the class had. Whatever the finding turned out to be, a fifteen-year-old’s name had already travelled, and a decision that took a fortnight would let it travel further. Speed was not only the head of department’s interest. It was Fathima’s too.

What Priya did not have was certainty, or a colleague who agreed with her.

The offer was serious and the vendor was not a fraud. Four hundred licences for Classes 8 to 10, free for a year, an app that worked on a low-end Android phone, and a dashboard for teachers showing who had practised and for how long. The chart in the deck showed a 63 per cent improvement in practice scores against a control group, over eight weeks.

The correspondent for the school, Vinod Rao, taught physics and had read the two studies that exist. He asked the five questions.

Was assignment random? Students had opted in. Was the outcome measured without the tool? No: the 63 per cent was practice scores inside the app. How long after? Eight weeks, same-term. Who paid? The vendor. Against what comparison? Students who had no extra practice at all.

The deck did not survive the second question, and the vendor’s representative, to her credit, said so when asked and offered to arrange a call with the researcher who had run the pilot. That call happened and was useful. The researcher was straightforward: the pilot measured engagement and in-app performance, it was never designed to measure learning, and she would not have put the chart in a sales deck herself.

Vinod took the trouble to say to his colleagues that this did not make the app bad. It made the evidence for it thin, which is a different finding and a common one.

The decision the school actually faced

The app had two modes. In the default mode a student stuck on a problem could ask for the answer and get it, worked, immediately. In a mode the school could switch on centrally, the tutor gave hints, asked what the student had tried, and did not produce the answer.

The Turkish study said what happens. Practice scores rose 48 per cent with the unrestricted version and 127 per cent with the hint version. Then everyone sat an exam without the tool, and the unrestricted group came out 17 per cent below students who had never had AI at all. The hint group matched them: no harm, no measurable gain.

So the hint mode was the defensible choice and it had two costs the head teacher named immediately. Usage would be lower, because a hint is less satisfying than an answer at ten at night, and the vendor’s dashboard would show it, and the vendor was providing four hundred licences partly for the usage figures. And the visible short-term result would be worse. Parents in that school pay fees they can barely afford and look at the practice scores in the app, which is what the dashboard shows them. A mode that produces a 127 per cent practice rise instead of 48 sounds better, until you notice that neither number is the thing anyone is buying the app for.

There was a further complication that the maths staff raised and Vinod thought was the strongest argument on the other side. About a third of the school’s students have nobody at home who can explain a quadratic equation, and for those students the alternative to an answer at eleven at night is not effortful practice, it is stopping. Withholding the answer is the right design for the child with a tutor and a parent who studied to Class 12. It is a harder thing to defend for the child who has neither, and the studies do not settle it, because they were not run on that question.

The board wanted a decision in a fortnight, and one member had already told a parents’ group that the school was getting AI tutors.

The pack arrived as a PDF on the headmaster’s phone: forty pages of Class 5 environmental science, in Santali, in the Ol Chiki script, produced by the district under a programme to supply mother-tongue material. It was the first home-language teaching material anyone at that school had ever been sent.

Subodh Hansda taught the class. He speaks Santali at home, teaches in a mixture of Santali and Hindi, and had been translating the Hindi textbook aloud, page by page, for four years. He read the first page of the pack and stopped.

It was not wrong in an obvious way. The sentences looked like Santali. The problem was that several of them were not sentences anybody would say, one term for a common tree was a word he did not recognise, and a question about the water cycle had a verb in a form that changed what it was asking.

He ran the check he had been taught. He typed three of the Santali sentences into a chatbot in a fresh conversation (fresh, so that nothing about the original was in front of it) and asked for English. What came back did not match the Hindi original in two of the three. The second sentence, which in Hindi asked children to name three uses of a river, came back in English as a statement about rivers being useful. A question had become a sentence.

He then did what a teacher does before accusing a document: he tried to make it work. He asked the same tool to translate the Hindi page himself, with a short glossary of the terms his textbook uses, and an instruction to match the register of a passage he typed in from a Santali reader. The output was no better and in one place worse. This was not a prompting problem, and half an hour spent rephrasing instructions taught him that faster than any explanation would have.

The two ways to be wrong

Using the pack meant putting material in front of thirty-four children in their own language, printed, official, and endorsed by the district: with errors in it that the children would not be able to identify as errors, because it was in their language and it came from the government and they were ten. Children who cannot tell a broken sentence from a real one in their home language learn that their language is the one that comes out strange.

Refusing the pack meant going back to what he had: a Hindi textbook and his own voice. That is not nothing (he is good at it, and the children learn), but it also meant that the only home-language material the school had ever received went unused, that the district would record the school as non-compliant, and that the headmaster, who had been pleased, would be told his teacher had rejected the government’s Santali books over three sentences.

There was a third possibility and it cost something too. Two mothers in the village write Santali well. One teaches at the anganwadi and has an hour on Thursdays. Asking her to read forty pages was asking for unpaid work from a woman who already had a job, in order to fix a district’s document.