Why I Still Believe in Human Tutors (Even After Writing About AI)
I spent the best part of a week writing about AI and tuition, weighing up where it genuinely helps and where I think it falls short. Somewhere around the fourth draft I realised I hadn't actually written about the other half of that question: what human tutoring is actually for. So this is the companion piece nobody asked for, and probably the more important one.
A randomised trial run in UK classrooms not long ago found something that should give me pause, given the business I run. Students working with an AI tutor did very slightly better on some outcomes than students working with a human tutor alone, about five and a half percentage points more likely to solve a new problem they hadn't seen a worked example of before. If tuition were only about getting the next answer right, that result would be the whole story, and I'd be writing about redundancy rather than belief.
But it isn't the whole story, and the same body of research says so itself. A year long study of American seventh graders found that when human tutors and AI worked together, students corrected their own misconceptions and reached the right answer more than ninety per cent of the time. Take the person out and leave a static, pre-written response in their place, and that figure dropped to sixty five per cent. The AI wasn't the problem in either version. What was missing was someone who could tell the difference between a wrong answer and a wrong way of thinking.
That difference is most of what twenty two years of running Blue Tutors has taught me about this job. I think of Alex, who came to us in the first year of her English degree having the feeling that Shakespeare was something she would never feel comfortable with. Nothing in her marks that term pointed to the real issue, which her tutor picked up on inside the first two sessions: Alex simply didn't enjoy literature not written in a modern style. Within a few weeks they had gotten over the hump and Alex had started to enjoy Shakespeare and realise why we still study his plays today. No algorithm reads a face across a table and decides to stop the lesson and just talk for ten minutes. A good tutor does that instinctively, and it's often the ten minutes that matters most.
I'm wary of romanticising this, though. Plenty of the tutoring we've delivered over the years has been fairly mechanical: past papers, timed conditions, catching the same careless error for the fifth week running. AI is genuinely good at some of that, and I said as much in the last piece. It doesn't get bored repeating a concept for the third time in one evening, and it's available at eleven o'clock the night before a mock, when no tutor should reasonably be expected to answer a message. There's no shame in using it for that. We already do, in small ways, and I expect we'll do more of it.
What it doesn't do, and what I don't think it can do in any way that actually matters, is want the student to succeed. That sounds soft, possibly even naive, coming from someone who spends a fair amount of his working life on spreadsheets of grades and match rates. But it's the thing almost every parent who has rung us in a quiet panic before results day is really asking about, underneath the question about predicted grades. They want to know that somebody is paying attention to their child specifically, not to a category their child happens to belong to.
There's also something more ordinary going on in a good tutoring session that's easy to overlook precisely because it's so unglamorous. A tutor sitting opposite a student can see the half second hesitation before an answer, the flicker of confusion that the student themselves hasn't yet put into words, and adjust the next sentence accordingly. They know, because they set the last piece of homework and watched it come back, exactly which method this particular student reaches for under pressure and which one trips them up. Type a question into a chat window and the system responds to the words you've typed, not to the six months of watching how you actually think. That accumulated, slightly untidy knowledge of one specific person is not a feature anyone has managed to build. I'm not convinced it's the kind of thing that can be built at all, because it isn't really information in the way a model understands information. It's closer to familiarity, the sort that only comes from repeated contact with the same person over time.
I still think about Nikky, a history tutor who took one of her students to the British museum. It wasn't in the lesson plan. Nobody asked her to do it. She did it because she'd been sitting across from the same student for 25 weeks and had started to care whether things went well for them, which is a different thing entirely from being competent at explaining history.
There's a piece of research I keep coming back to, from outside education altogether: students who feel genuinely cared for by the adults teaching them engage more, take more academic risks, ask more questions and absorb more of what's in front of them. That isn't a tuition-specific finding. It's closer to something true about how people learn in general. We work harder for people we trust and believe are on our side. An AI tutor can be endlessly patient, and that's worth something, but patience isn't the same as being on someone's side.
UK students themselves seem to sense some of this tension already. A recent survey found that two thirds of them now say AI use in their schoolwork has actually harmed their own critical thinking, sharply up from barely more than half a year earlier. That's a striking thing to hear from the people using it most. It isn't an argument against AI so much as evidence that they can tell the difference between something that gives them an answer and something that helps them think, and that, often enough, they'd rather have the second.
A parent wrote to us a while back, not long after her son's physics exam, just to say that what had changed his year wasn't the grade itself, he ended up getting an 8, but his attitude towards science, and studying in general, was much more positive and he'd developed an interest. I've kept that message. I go back to it whenever I need reminding what we're actually selling, because it was never hours or worksheets. It's the feeling of somebody being properly in your corner.
I haven't changed my mind since writing about AI. I've mostly worked out what I actually think tuition is for. It isn't answer generation, however good the answer generator gets. It's a relationship with someone who notices, adapts, and genuinely minds how things turn out for the person sitting across the table from them. I suspect the tutors who do best over the next decade will be the ones who let AI take the repetitive weight off their shoulders, and spend the time that frees up doing more of that, not less. That's the bet I'm making with Blue Tutors, and it's one I'm glad to put in writing.