How pupils use artificial intelligence (AI) turns out to matter far more than how often they use it. That is the clearest message about how pupils use AI from a new year-long study of 4,062 secondary pupils in Germany, published as a preprint on 6 November 2026. Pupils who mostly used AI to get tasks done put in less effort later in the year. Pupils who used it to understand their work put in more. And the pupils leaning on AI hardest were the ones who were already struggling.

The short version: AI use did not raise or lower test scores over the year, but the purpose behind it was linked to effort, and lower-attaining pupils used it most. Schools should stop asking whether pupils use AI and start teaching them which jobs to hand over and which to keep.

What the study found

The study, led by Ronny Scherer with colleagues including René Kizilcec and Ulrich Trautwein, followed pupils in 136 secondary schools in Baden-Württemberg through the 2023/2024 school year. Pupils sat standardised tests at the start and end of the year in six subjects: mathematics, chemistry, biology, physics, German and English. They also reported how often they used generative AI for seven school activities.

The researchers split those activities into two groups. Task-completion use meant translating texts, doing homework, preparing presentations and essays, and solving tasks during class. Learning-oriented use meant understanding class material better, further learning and preparing for tests. Those are very different things, even when the same chatbot is open on the screen.

Use grew quickly. By the end of the year, 62% of pupils reported using AI for homework and 65% for exam preparation. Only about 5% said they never used it at either point. Remember that this was 2023/2024, so the figures in your school today are almost certainly higher.

Then came the findings that should interest every head of department:

• Pupils with lower test scores at the start of the year went on to use AI more, in every subject for task completion and in five of the six for learning.
• Neither kind of use predicted later test scores in any subject.
• A pattern of more task completion and less learning-oriented use predicted lower effort later in the year in four of the six subjects.
• Learning-oriented use was linked to higher effort in mathematics, chemistry and physics.

The authors are careful. This is a preprint, so it has not yet been through peer review. The use data come from pupils’ own reports, and the two kinds of use were closely related, so the effort findings rest on a fairly narrow slice of the data. But it is a large sample with real test scores, tracked over time, which is more than most AI claims can offer.

Why how pupils use AI matters for effort

The headline that AI made no difference to attainment could easily be misread as “nothing to worry about”. I read it the other way. Effort tends to move before results do. A pupil who quietly stops wrestling with chemistry problems in the spring may not show it in a test until the following year. The authors make a similar point: changes in effort may come before changes in achievement over a longer period.

The question for a school is no longer whether pupils use AI. It is which part of the thinking they hand over.

This fits what we already know about cognitive offloading. If the tool does the part of the task that was meant to make you think, you get the finished product without the learning. We covered this in our piece on cognitive offloading and AI, and in our look at research suggesting AI homework help can raise scores while reducing learning. The German data add something new: the pattern shows up over a whole school year, in ordinary schools, with no researcher designing the activity.

The pupils who need guidance most

The finding I keep coming back to is who uses AI most. It is not the high flyers. It is the pupils who started the year behind. The authors suggest these pupils may be turning to AI as a way of coping when work feels too hard.

That matches what teachers say. In a Reuters report published on 7 November, Rusty Shabe, a middle school English teacher in San Francisco, said pupils often turn to AI at home when they feel overwhelmed. His response was to bring more of the writing into lessons, spending a month on one essay with his class.

The researchers are clear that their findings give no grounds for telling lower-attaining pupils to use AI less. Banning it for them would remove a tool that some are using to understand things. The job is to help them use it in ways that keep them working.

The gap in teaching is wide. GoStudent’s Future of Education Report 2026, published on 2 November, surveyed children aged 8 to 16 across six European countries. In the United Kingdom (UK), 72% of children surveyed said they use AI for schoolwork, yet only 28% said they had had a structured lesson at school on how to use it. Among UK children who use AI for schoolwork, 73% said they had handed in AI-generated work as their own. That survey was commissioned by a tutoring company, so treat it as a temperature check, but the direction is hard to ignore.

What this looks like in a classroom

None of this needs a new platform. It needs teachers to name the difference between the two kinds of use, out loud, in their subject. Here is what that can look like.

Sort the jobs with your class

Give a Year 10 science group a list of things they might ask a chatbot: “explain why ionic compounds conduct when molten”, “write my conclusion”, “give me five practice questions on bonding”, “check my method for mistakes”, “do question 4”. In pairs, they sort each one into “helps me learn” or “does it for me”. The argument about the borderline cases is where the learning happens. Checking a method is fine after you have written one. It is not fine instead of writing one.

Make the thinking visible

Bobson Wong, a mathematics teacher in New York quoted in the same Reuters piece, now asks pupils which problem-solving method they prefer and why, since AI can show every step. That is a neat move. The answer is cheap to copy, but a choice between methods, explained in your own words, is not. In English, ask pupils to submit the question they asked the tool alongside the paragraph it helped them improve.

Target the pupils who are struggling

If lower-attaining pupils use AI most, they should get the most explicit coaching. A five-minute conversation during an intervention session about how they used AI on last night’s homework will tell you more than any detection tool. Ask to see the chat. Praise the “explain this again more simply” prompts and question the “write it for me” ones.

Teach a simple self-check

Before pupils hand a job to AI, they ask themselves one thing: “Is this the bit my teacher set this task to make me practise?” If the answer is yes, they do it first and use AI afterwards. It is the kind of metacognitive judgement the study’s authors say should become part of AI literacy, and it fits neatly into the AI literacy work many schools are already planning.

What school leaders should do next

For leaders, the practical step is to change the question in your AI policy. Most policies still read as a list of permitted and banned tools. Add a short section on purpose: which uses support learning in each subject and which replace it. Ask heads of department to write two or three examples for their own subject, because “use AI to understand” means something different in French translation than it does in physics.

Then check where your guidance actually reaches pupils. If structured lessons on AI happen once in a Personal, Social, Health and Economic (PSHE) education session, the pupils most likely to rely on AI will have forgotten it by the next piece of homework. The message needs to be repeated by subject teachers, at the moment the work is set.

Frequently asked questions

When your pupils use AI at home, do you know which jobs they are handing over? Tell us in the comments how you are teaching the difference in your subject.