Personalising tasks with artificial intelligence (AI) is one of the most appealing promises made to teachers: type in a word problem, ask for a version about football or TikTok, and watch a bored class sit up. A new study from the United States puts real numbers on personalising tasks with AI for the first time, and the results are more useful than the sales pitch. Seven maths teachers rewrote 143 textbook problems around their pupils’ interests with ChatGPT. Each problem took about four minutes. Pupils gave the results lukewarm ratings, and what they wanted was something no teacher has time to give everyone.
The short version: AI makes it quick to rewrite one problem, but not quick enough to tailor every task to every pupil. The teacher’s real value was spotting unrealistic numbers, odd contexts and biased content. Use AI for a few well-chosen interest swaps, check every one, and let pupils help choose the contexts.

What the study did
The study, led by Candace Walkington at Southern Methodist University, was published open access in the journal Educational Technology Research and Development on 2 November 2026. Seven experienced middle school maths teachers, from six schools across the United States, took problems from their existing curriculum on proportional relationships, percentages, expressions and equations. They then used ChatGPT to write personalised versions linked to their pupils’ out-of-school interests, such as sports, food, music, gaming and social media.
The teachers had a two-hour training session first, covering how to find out what pupils are interested in and how to watch for bias, readability and cultural relevance. They usually wrote two versions of each problem, one about sport and one about food, for example, and pupils chose which to do. The 521 seventh-grade pupils (roughly Year 8 in England) whose parents did not opt out then completed the problems as classwork or homework, rated a sample of them and, where parents had consented, explained in writing what they thought.
Unusually, researchers recorded the teachers’ screens as they worked, so we can see all 558 prompts they typed and how long each problem took.
What personalising tasks with AI costs in teacher time
On average a personalised problem took 246 seconds, or about four minutes, and 3.9 prompts. Individual teachers averaged between two and five and a half minutes per problem. That sounds quick until you do the sums the authors did. A typical problem set has around ten questions, so one alternative version of each costs about 40 minutes per assignment.
The original plan was for teachers to create four versions of every problem, each tied to a different interest. Once the sessions started, the researchers found that was untenable.
AI made rewriting a single problem fast. It did not make it fast enough to tailor every task to every pupil, and the checking was the part that mattered.
There was also a hint, and it is only a correlation, that problems refined with three to five prompts were rated slightly better than first drafts. Endless tinkering did not help, though.
What pupils actually wanted
This is where the study gets honest. On a scale of one to five, pupils’ ratings of how interesting the problems were averaged around 2.5, and just over half of all ratings were a one or a two. In the written feedback, 163 responses suggested a problem was interesting and 243 suggested it was not.
The reasons were revealing. Pupils responded to specific references far more than broad topics. A problem about music in general left many cold. A problem mentioning Spotify delighted the pupil who uses it every day and irritated the one who does not. One pupil wrote that a problem was not interesting because they did not play Roblox. Another suggested swapping TikTok for Amazon.
That is the trap. The finer the personalisation, the more pupils like it, but a teacher working with AI can only realistically personalise for the class, not for each child. The authors note that individual-level personalisation is more feasible when the AI works directly with the pupil, as some tutoring tools already do. That takes the teacher out of the loop, and the rest of the study shows why that is a risk.
Why the teacher in the loop still matters
The most valuable prompts were not the clever ones. They were the corrections. Teachers asked ChatGPT to change items being bought from curtain rods to things a seventh grader would actually buy. One asked, in a problem rewritten from a plane to a car, whether the driver could obey the speed limit. Others fixed the maths, changed a variable letter, or simplified wording that had become harder to read.
Some corrections were about safety and fairness. Teachers removed problematic celebrities, noticed when ChatGPT quietly changed a problem about a boy and a girl into one about two boys, and rejected a scenario that tracked a person’s weight. Each of those would have reached pupils if nobody had read the output carefully.
Pupils, interestingly, rarely noticed realism problems. Only a handful pointed out that the numbers were off, for instance that headphones are not usually that expensive on TikTok Shop. If pupils will not catch unrealistic maths, the teacher has to. The authors argue that depth matters because we want pupils to see mathematics as a sensible way to model the world, not a set of arbitrary sums dressed in brand names. I agree. A problem with silly numbers teaches pupils that school maths has nothing to do with life.
The authors are careful to add that putting a teacher in the loop does not guarantee good checking. Teachers need the subject knowledge, the time and the permission to reject what the AI produces.
How to use this in your classroom
Personalised contexts can work. A 2024 meta-analysis cited in the study found that tailoring tasks to interests improves both interest and learning. The lesson from this study is to be selective rather than ambitious. Here is what I would try.
Personalise the hook, not the whole worksheet
Pick the one or two problems that open a topic or carry the key idea, and spend your eight minutes there. A percentage problem about a real discount at a shop your pupils use is worth more than ten generic sports swaps.
Ask pupils for the references
A quick exit ticket asking what pupils spend money on, play or watch gives you the specific details AI cannot guess. Feed those into your prompt yourself rather than asking the AI to pick what teenagers like.
Offer a choice of two
The teachers in the study usually gave pupils two versions and let them choose. That spreads the appeal without multiplying your workload.
Check every output against three questions
- Is the maths still correct and taught the way we teach it?
- Would this actually happen, at these prices and quantities?
- Could any pupil find this hurtful, excluding or inappropriate?
Also check readability. Rewritten problems can become wordier, which hits pupils with special educational needs and disabilities (SEND) and those learning English as an additional language hardest. Our earlier piece on AI and SEND covers this in more detail.
Let older pupils do the personalising
Ask pupils to rewrite a problem around their own interest, keeping the maths identical, then swap with a partner who checks whether it is realistic. That gives them ownership and practice in spotting nonsense numbers, which they did not do naturally in this study.
What school leaders should take from it
If a supplier promises personalised resources for every pupil at the click of a button, ask who checks them. This study suggests that the checking is where the quality comes from, and that it takes time. It fits a theme running through our coverage of the Aila lesson planning trial and AI and teacher workload: AI tends to shift teacher time from writing to checking, rather than removing it.
The study has clear limits. It involved seven paid volunteer teachers and one age group in the United States, the ratings came from a sample of problems, and the analysis of prompts and ratings is descriptive. But it is rare to see teachers’ real prompts and timings, and it is a far better guide than the brochures.
Can artificial intelligence (AI) personalise maths problems for pupils?
Yes. In a 2026 study, seven maths teachers used ChatGPT to rewrite 143 curriculum problems around their pupils’ interests, such as sport, food, music and social media. The AI did the rewriting quickly, but teachers still had to correct unrealistic numbers, odd contexts and occasional biased content.
How long does it take to personalise a task with AI?
About four minutes per problem in this study, using an average of 3.9 prompts. That adds up to roughly 40 minutes to make one alternative version of a ten-question assignment, which is why the researchers dropped their plan for four versions of every problem.
Do pupils find AI-personalised problems more interesting?
Only partly. Ratings averaged about 2.5 out of 5. Pupils responded best to specific references they actually use, such as a particular app or game, and were put off by references that did not match their own interests.
Should teachers check AI-personalised tasks before using them?
Always. Teachers in the study caught incorrect maths, unrealistic prices and quantities, harder wording and content that was inappropriate or biased. Pupils rarely spotted these problems themselves, so the checking has to come from the teacher.
Have you tried personalising problems with AI for your classes, and did your pupils notice? Tell us what worked, and what the AI got wrong, in the comments below.


