The Metacognitive Trap: The Receipts

The 60-second Briefing

• The NEU's State of Education: AI 2026 surveyed 9,408 teacher members – 66% of secondary teachers report a first-hand decline in pupils' critical thinking due to AI usage.
• Teacher AI usage has jumped from 53% to 76% in a single year, but 49% of schools have no AI policy whatsoever for either staff or pupils.
• Wensleydale School in the Yorkshire Dales spent £600 on AI marking for GCSE mocks – workload temporarily went up, not down, because teachers double-marked alongside the system.
• Just 14% of teachers support the government's AI tutoring plan for disadvantaged pupils. Nearly half (49%) actively oppose it.
• The procurement principle from earlier this week now has classroom data behind it.

Earlier this week I argued that if a piece of software simply removes the cognitive friction required for learning, IT departments should refuse to buy it. That was the principle from the boardroom. The data from the staffroom landed almost in parallel, and it tells the same story from a different angle.

The National Education Union's State of Education: AI 2026 report surveyed 9,408 teacher members in English state schools, ahead of their April Annual Conference in Brighton. The headline finding is one we should note: 66% of secondary teachers reported they had observed a first-hand decline in pupils' critical thinking as a result of AI usage. A third of those agreed strongly. In primary, the figure was 28% – still significant, but the secondary number is the one that should be making heads of department uncomfortable.

That is not a vendor pitch, a think-tank report, or a politician's keynote. That is what teachers see when they look at a class of teenagers handing in work.

The same survey shows that teacher use of AI tools has climbed from 53% to 76% in a single year, mostly for resource creation, lesson planning, and admin – with just 7% using it for marking. Yet 49% of schools have no AI policy at all, for staff or pupils, and 66% have nothing specific to pupil use. We have a profession adopting a powerful technology faster than the institutions around it can write the rules. Adoption without governance is just a polite term for chaos.

There is also a sharp piece of feedback in the NEU numbers worth flagging. Just 14% of teachers expressed support for the government's plan to roll out AI tutors to disadvantaged pupils, and 49% are actively opposed. Whatever the merits of the Pioneer Group tender, the profession that is being asked to host the trials is, at best, sceptical. That is worth knowing before any independent school SLT decides to mirror the state sector approach.

For an actual data point on what happens when a UK school tries the efficiency case in good faith, look at Wensleydale School in the Yorkshire Dales. They became one of the first state schools in the country to trial AI marking on GCSE mock papers, using a platform called TopMarksAI. They marked English language and English literature mocks for the entire Year 11 cohort, plus selected history papers. Cost: £600 for 1,250 credits, which works out at 45p per extended answer.

The result, in headteacher Julia Polley's own words on the BBC's coverage, was that workload went up rather than down. Teachers continued to mark alongside the AI – partly because they wanted to verify the output, partly because they professionally wanted to know where their pupils were at. Detailed feedback to pupils improved. Hours saved did not materialise.

Now, this is not a story about AI marking being bad. The school did the right thing: they piloted, they double-marked, they evaluated, and they were honest about the results. That is exactly the kind of disciplined trial we should all be running. But it does land a clean punch on the standard sales pitch. The promise is "AI marking saves teacher time." The reality, when you actually run the experiment, is more nuanced. You either get faster, more detailed feedback at additional cost and equivalent workload, or you get the time saving by trusting the system without verification – at which point you are not really teaching; you are outsourcing.

This is the friction I wrote about in The Metacognitive Trap, but seen from the other side of the desk. Generic AI promises to remove cognitive load from pupils and administrative load from teachers. In both cases, the load doesn't actually go away – it just gets relocated. For pupils, it relocates into the exam hall, where the OECD's evidence shows performance drops sharply once the tool is removed. For teachers, it relocates into a verification step that they cannot ethically skip, because the DfE has been clear that final responsibility always rests with them.

There is one more thread worth pulling. The NEU survey found that just 7% of teachers use AI for marking. The Wensleydale pilot is one of the early UK state-sector trials. Independent schools watching this space should resist the temptation to leap ahead of the evidence. The state sector is – for once, and for good reasons – running the controlled experiment. We can learn from those trials before our bursars start signing purchase orders.

A short summary, then. The principle from earlier in the week was that IT must say no to tools that strip the friction out of learning. The data this week is that the friction is not always being stripped out – sometimes it is just being moved, repackaged, and in some cases dropped onto the same teachers who were promised relief. None of which means we should reject AI in our schools. It does mean we should stop accepting the marketing version of the story and start asking who, exactly, is doing the cognitive work in any given lesson.

If the answer is the pupil, fund it. If the answer is the machine, refuse it.

See you in the digital staffroom.