I have been tidying up my website again. After redesigning parts of the homepage and giving my About Me section more character, I started thinking about what I actually want this space to become.
This time, I wanted to write something more serious: my changing thoughts about artificial intelligence and education.
A few years ago, around 2024, much of the AI conversation centred on one question:
How do we use it?
Everyone seemed to be learning how to prompt. We talked about prompt engineering, generating lesson plans, creating resources and finding ways to introduce AI into education, business, design and almost every other industry.
I was part of that conversation too (see below).
But in 2026, I think the conversation has changed.
Two people can now use the same AI platform—even the same prompt—and experience very different results. AI increasingly works within individual contexts, conversations, preferences and workflows. Someone who has used AI extensively may know how to refine an idea, question an output, provide context and recognise mistakes. A new user using exactly the same prompt may simply wonder why they are not achieving the same result.
So I have become less interested in teaching people how to prompt AI.
Instead, I have become more interested in another question:
How can AI actually support learning?
Starting With the Learning Gap
Over the past six months, I have been experimenting with this in my own teaching.
Within a British international curriculum, we already have a clear destination. Exam boards publish specifications describing what students should know, understand and be able to do.
If we know the destination, we can work backwards.
What should students understand by the end of the year? What skills should already be secure? What comes next?
Giving AI these curriculum objectives is relatively straightforward.
The more interesting part is understanding where students currently are in relation to them.
I began combining curriculum expectations with evidence from formative assessments, summative assessments, examinations and classroom work. Rather than asking AI to generate another worksheet, I started asking:
What is the gap between what this learner currently demonstrates and what they need to do next?
In Computer Science, one student might understand basic input and output but struggle with selection. Another might understand selection but need more practice with iteration.
AI can help organise these patterns and make feedback more specific.
Instead of saying:
"You need to improve your programming."
We might say:
"Your input and output skills are secure. Your next area of development is using selection confidently within a larger program."
That gives the learner something tangible to work towards.
But identifying the gap creates another question:
What should the student learn next—and how?
This is where AI still becomes surprisingly generic.
It can generate practice activities, explanations and learning pathways, but deciding whether those materials are appropriate for the curriculum level, prerequisite knowledge, cognitive demand and individual learner still requires professional judgement.
That judgement is pedagogical.
What the Research Says
This led me to the systematic literature review Artificial Intelligence-Based Personalised Learning in Education by Farhood, Nyden, Beheshti and Muller.
The researchers reviewed 125 studies published between 2015 and 2025. Their findings strongly connected with what I had been experimenting with.
Among the studies reviewed:
- 41.6% focused on feedback and assessment
- 20.8% focused on tutoring and support
- 20.8% focused on content recommendation
- 16.8% used hybrid approaches
This suggests that one of AI's most useful educational roles may not be replacing teaching, but analysing learning and responding to what students demonstrate.
The review also reports encouraging findings from individual studies. One intelligent tutoring system reported a 22.95% improvement in student performance. Another found learners completed tutoring conversations around 27% faster using dynamic student modelling. Other studies reported improvements in learner performance, satisfaction and assessment scores.
These numbers are encouraging, but they require caution.
They come from different technologies, learners and contexts. They do not mean that introducing ChatGPT into a classroom automatically produces a 20% improvement in results.
What they suggest is that carefully designed personalised interventions can produce measurable benefits under particular conditions.
The Problem of Data
I cannot yet claim that my own approach improves student attainment by a particular percentage.
I do not have that evidence.
Collecting it properly would require appropriate research design, permissions, safeguards and ethical procedures.
And this raises one of the biggest tensions in AI-supported personalised learning:
The more personalised we want AI to become, the more it potentially needs to know about the learner.
Not every piece of student information belongs inside an AI system.
We need to consider what information is genuinely necessary, whether it can be anonymised, how it is processed and whether it should be used at all.
More data does not automatically create better learning.
We need the right data, sufficient data and high-quality data.
Because even an extremely sophisticated AI can produce a sophisticated wrong answer.
And in education, that wrong answer concerns a child and their learning.
Technology Is Not Pedagogy
The literature also identifies continuing challenges: privacy, algorithmic bias, scalability, teacher training, time constraints and pedagogical alignment.
That last point particularly interests me.
Something can be technologically impressive while being educationally poor.
AI might generate a beautiful learning pathway in seconds—but does it understand prerequisite knowledge? Cognitive load? Misconceptions? Productive struggle? Examination objectives?
And does it understand that the teacher only has fifty minutes before the lesson ends?
Nor do I see it as teachers becoming increasingly sophisticated prompt engineers.
What interests me is somewhere between the two:
AI that can understand enough about curriculum, assessment and learner needs to identify where support may be required—combined with a teacher who understands pedagogy, context and the actual human being sitting in front of them.
Perhaps the real question for education is no longer how well we can use AI.
It is how well these two forms of intelligence can learn to work together.
References
Farhood, H., Nyden, M., Beheshti, A., & Muller, S. (2025). Artificial intelligence-based personalised learning in education: A systematic literature review. Discover Artificial Intelligence, 5, Article 331. https://doi.org/10.1007/s44163-025-00598-x