Teachers Are Using AI. Why Don’t They Want It Tutoring the Pupils?

There are 2 numbers that policymakers working on AI-in-Education should probably place beside each other - 76% & 14%

In the National Education Union’s 2026 survey, 76% of teacher respondents in English state schools said they were already using AI tools in their day-to-day work and only 14% supported the government’s plan to introduce AI tutors for disadvantaged pupils.

If those numbers are treated as a contradiction, we miss the most important part of the story - Teachers are not rejecting AI. They are already using it.

What they appear much less willing to do is automatically turn usefulness into authority. And that distinction matters now because this is no longer an abstract technology debate. The UK government intends to move AI tutoring into real schools, for real children, at potentially significant scale.

Policy stakes are larger than another classroom experiment

In January 2026, the UK government announced plans for AI-powered tutoring tools that it said could eventually support up to 450,000 disadvantaged pupils. The tools are intended to provide personalized learning support, with development beginning alongside teachers and technology companies and school availability targeted by the end of 2027.

By April, the programme had moved further - The government invited EdTech companies and AI labs to help build the systems, saying that up to eight companies would begin testing tools in schools from summer 2026 under teacher supervision. This is no more a debate about whether a few experimental teachers should try ChatGPT. The policy ambition is to use AI as part of the response to one of education’s hardest problems: unequal access to individual academic support.

The attraction is obvious. The government points to evidence that high-quality one-to-one tutoring can accelerate learning by around five months, while access to private tutoring remains unequal. If AI could reproduce even part of that individualization at dramatically lower cost, the upside would be substantial.

But that is precisely why the 14% number matters. The teachers being asked to help deliver the policy are not yet convinced by the premise.

This is not a profession resisting technology

The NEU survey included 9,408 teacher members in English state schools in its headline AI findings. AI use among respondents had increased sharply, from 53% the previous year to 76% in 2026. And the use cases tell us something important- 61% reported using AI for resource creation, 41% for lesson planning, 38% for administrative work and only 7% used it for marking.

This is not indiscriminate enthusiasm. The highest adoption sits in work where AI can extend teacher capacity without independently deciding a student’s educational outcome. Usage falls sharply as the task becomes more consequential. That is a signal policymakers should take seriously.

Teachers appear to be drawing a line between: AI helping the professional and AI assuming part of the professional relationship with the learner.

That is a far more useful distinction than “pro-AI vs. anti-AI.”

Why disadvantaged pupils make this a higher-stakes decision

The government’s equity argument is compelling. Children from families able to purchase private tuition receive access to individual attention that many others do not. Technology may offer a way to narrow part of that gap.

But there is another question hiding underneath the policy: Should the children with least access to human educational support become the first population for whom technology is expected to approximate it?

That requires a higher evidence standard, not a lower one. If AI tutoring proves effective, scalable and safe, it could expand access to support in a meaningful way. If it does not, there is a risk that a technology intervention becomes a cheaper substitute for the human attention disadvantaged pupils were missing in the first place.

That is why this programme should not be judged primarily by: number of licences, sessions, student engagement or how convincing the product feels.

The relevant outcome is: Did these pupils actually learn more than they would have under the realistic alternative?

Everything else is secondary.

Teachers are also signaling a governance problem

The NEU survey found that 49% of respondents said their school had no AI policy covering either staff or students and 66% said there was no student-specific AI policy.

That produces an uncomfortable sequence - AI adoption is accelerating, government-backed tutoring pilots are moving ahead. Yet in many schools, the basic rules governing everyday AI use are still not settled. That is not a reason to stop experimentation, it is a reason to recognize the order of operations.

A school needs to answer basic questions before AI becomes infrastructure: What student information can enter the system? What does the provider retain? What decisions require teacher review? How are hallucinations handled? What happens when the AI explanation conflicts with the curriculum? How is inappropriate or unsafe output escalated? How do teachers know whether the student is learning rather than simply completing more work? Who is accountable when something fails?

Those are not administrative details. They determine whether a useful tool becomes dependable educational infrastructure.

The NEU also reports that 66% of secondary teacher respondents believed pupils’ critical thinking had declined because of AI use, compared with 28% of primary respondents.

This does not prove AI caused a measurable decline in critical thinking - It is teacher perception data, the distinction matters. But it should not be dismissed either.

If a large share of teachers report that they are observing changes in how students think through problems, there is a question worth investigating. Especially because generative AI changes the cost of avoiding cognitive effort.

A student who once had to formulate an answer can now generate one.
A student who once had to read a difficult source can request a summary.
A student who once had to struggle with the opening paragraph can begin with something fluent.

Those capabilities are useful. They also make it easier for task completion to become detached from learning.

That is precisely the problem any AI tutor needs to solve.

Early adopters suggest the right attitude is neither panic nor blind rollout

Its 2025 research examined 21 early-adopter schools and further-education colleges.

Leaders were using AI, building internal expertise and experimenting with approaches to governance. They also reported useful applications around lesson planning, resource creation and administration.

But Ofsted was careful not to turn experience into proof. The research involved a small, purposively selected group of early adopters, and Ofsted explicitly identified continuing gaps in knowledge about how AI affects educational outcomes.

That is exactly the posture the next stage requires: move, but measure.

The next phase of AI in education is about responsibility, not adoption

The argument over whether schools will use AI is largely over. They already are.

The more consequential argument begins now. Which responsibilities should be handed to these systems?
A teacher using AI to prepare a worksheet is one thing, an AI system tutoring a child who is behind in mathematics is another, an AI system grading that child is another again.

Same technology category, different levels of responsibility, different consequences when it fails and different standards of evidence should apply.

That is why the gap between 76% adoption and 14% support is worth paying attention to.

It is not inconsistency. It may be evidence that teachers are already learning the distinction policymakers now need to make: a technology can be useful long before it is trustworthy enough for every role we can imagine giving it.

If AI tutoring is going to reach hundreds of thousands of pupils, proving where that line sits should come before scale - not after it.