Human-in-the-loop learning: Where AI belongs in education (and where it doesn’t)

Pretty much every school and ed-tech company has an AI strategy. They have answered the question: “should we use AI?”

But, far fewer have a clear answer to the harder question underneath: “where, specifically, does AI help, and where might it quietly make things worse?”

That distinction matters more in education than almost anywhere else, because the thing being optimised isn’t just output, it’s a person’s ability to think. Get AI wrong and there’s a risk you produce a student who finishes more homework, but understands less of it.

Where AI earns its place in education

The strongest evidence for AI in education comes where it sits behind the teacher, not in front of the student.

Stanford’s Dora Demszky built M-Powering Teachers, a tool that gives teachers AI-generated feedback on their own classroom discussions rather than putting AI in direct conversation with students. In a randomised controlled trial, it improved teachers’ uptake of student contributions by 13% – an impressive jump in a skill that’s tricky to coach at scale. Stanford’s Vice Provost for digital education, Matthew Rascoff, frames this as “many-to-one” AI: a single teacher gets support drawn from many data points, but the teacher still makes every judgement call in the room.

This is a pattern worth noting. Marking speed, diagnostic reports on where a class is struggling, adaptive practice sets, instant feedback on a set of practice questions, are the jobs AI is good at – and the jobs that used to eat into a teacher’s evening instead of their teaching time.

Used this way, AI doesn’t replace judgement, but can buy back the hours a good teacher would otherwise spend on marking, freeing them up to spend time on the aspects of teaching a machine can’t do.

Where AI and education don’t mix

The US Institute of Education Sciences (IES) reviewed the research on AI chatbots used by students and found a less positive picture: students working with general-purpose AI chatbots showed lower exam performance, shallower learning processes, and reduced brain activity compared with peers using traditional methods – even when the students themselves rated the AI tool as helpful. It turns out that ‘feeling helped’ and actually ‘being helped’ are two different things.

A 2025 Frontiers in Education study comparing ChatGPT with human tutors found the same split. Human tutors scored meaningfully higher on encouraging critical thinking (4.5/5 vs 4.2/5), personalised feedback (4.6/5 vs 3.8/5), and helping students grasp complex concepts (4.8/5 vs 4.0/5).

Where ChatGPT scored highly was in how accessible students found it: reporting that they were far more comfortable asking it questions (4.9/5 vs 2.8/5 for human tutors) and preferred it for self-paced study.

Rascoff makes the sharpest version of this point about writing specifically: for most students, the ability to write and the ability to think are intertwined, so handing the drafting over to AI doesn’t just save time, it can quietly remove the thinking the task was supposed to build. The IES review flags something similar with AI’s tendency toward “sycophancy” – agreeing with a student’s reasoning rather than challenging it – which is close to the opposite of what a good teacher does when a student is confidently wrong.

Human judgment matters

Across the research, the dividing line isn’t really “AI vs no AI” – it’s whether a human stays in the loop in the situations where judgement, challenge, and struggle all matter.

A useful way to think about this is: if the task involves marking, retrieving, drilling, or surfacing where a student is stuck, AI can take work off a teacher’s plate. But, if the task involves working out what a student needs, having a probing conversation with a student who’s confidently wrong, or working through the moments of struggle that build understanding, the teacher is key.

It boils down to “personalised learning” where personalisation isn’t a chatbot that never runs out of patience – but a teacher who knows a specific student well enough to know exactly where to push and where to hold back, using AI-generated information to do it quicker.

It’s why MyEdSpace has deliberately built its GCSE and A-level teaching around live teachers rather than outsourcing the teaching parts to AI tutors: technology can speed up practice, feedback, and preparation, but the teaching itself – the bit where a student gets properly stuck and a real-life person notices and helps them through it – stays human. So, students studying everything from their 11+ exams, all the way up to complex topics like A-level Maths and Further Maths with MyEdSpace still learn from a named teacher, in live, online lessons.

AI in education isn’t a ‘yes-or-no’ question. It’s a placement question. Put it behind the teacher and it can work wonders. Put it in front of the student instead of the teacher, and you risk optimising for the feeling of learning rather than actual knowledge acquisition.