AI is already in Britain's classrooms. The rules, the evidence and the money are still catching up

Ask a Year 10 pupil in England whether they use AI for schoolwork and the honest answer, most of the time, is yes. Ask their teacher and, increasingly, so is theirs. What neither can say with much confidence is what the rules are, whether it helps, or who is in charge. That gap, between a technology that arrived in classrooms through pupils' phones and an education system still deciding how to respond, is the subject of a sharply worded report published this week by the historian Sir Anthony Seldon and the venture capitalist Tim Bunting, which described the current state of affairs as "chaotic and anarchic" and called for a rapid response unit inside the Department for Education.²

The government's reply is that it is already moving. It has published guidance for schools, safety standards for products, training modules for staff, and in April it invited edtech and AI companies to bid for a share of a programme to build AI tutoring tools for disadvantaged pupils, backed by £23m for testing AI and edtech in classrooms. This explainer sets out what is actually happening in UK schools, what the policy says, what the evidence shows, and what the argument this week is really about.¹⁶

What pupils are already doing

The Seldon and Bunting report, published by the polling and policy consultancy Public First under the title Britain's Greatest Education Opportunity Ever, draws on a survey of 4,021 people aged 11 to 24 in England, carried out in August. It found that 87% had used an AI tool, that 63% of secondary pupils had used AI for homework and 77% to understand something difficult, and that 38% believed they were more skilled with AI than their teachers. Perhaps the most striking finding was about trust: 47% said they would trust AI over a human to check whether something was true, against 21% who would trust the human. Fewer than one in ten, 9%, said their school or university prepared them very well for a world with AI in it.²³

Those figures fit with what the regulator Ofcom has been finding. Its 2026 Children and Parents report, published in May, says around half of online 8 to 17-year-olds use AI tools, often for learning or schoolwork, and its qualitative Children's Media Lives study found every child it followed had encountered AI-generated content, with several using chatbots to generate ideas, answer questions and in some cases complete homework outright. Parents are wary: in a Parentkind survey of more than 5,000 parents in April 2025, just 12% thought AI should be used in the classroom.¹³¹⁴

What teachers are doing

Teacher Tapp, the daily survey app used by thousands of teachers in England, reported this spring that by autumn 2025 a majority of teachers were using AI weekly and only 12% had never used it. The uses are mostly behind the scenes: 54% had used AI for lesson preparation, 50% for quizzes, 46% for marking support, 41% for letters to parents and 39% for pupil reports. Just 11% had used it live in a lesson, and only 14% had used it to plan their most recent lesson, which the analysis takes as a sign that teachers treat it as a starting point rather than a substitute for their own judgement. Primary teachers lean on it for planning and reports; secondary teachers for quizzes and exam-style questions. Among the minority who avoid it, the main reasons given were reliability (56%), academic integrity (46%) and fear of pupils becoming dependent (46%), not a lack of confidence with technology.⁹

Confidence is a different matter from use. A weighted Teacher Tapp survey of more than 9,000 teachers commissioned by the Royal Society in March 2026, and submitted to MPs on the Education Committee, found that only 14% felt confident across all three strands of AI literacy: understanding the technology, using it in practice, and its human and ethical dimensions. Just 18% had received AI-related training, 15% had clear whole-school guidance, and 23% said they had received no support at all. Fewer than 2% said AI literacy was taught across most subjects in their school.¹⁰

At the far end of the spectrum, a few schools are experimenting well beyond planning tools. The BBC reported in December that Great Schools Trust, which runs academies in Liverpool, Warrington and Bolton, was using AI to mark assessments and mock exams and preparing to trial deepfake videos of teachers delivering personalised feedback to each pupil, with the trust's chief executive insisting the aim was "not to replace teachers". A Lancashire academy's decision to teach top-set maths through a remote teacher based 300 miles away prompted strikes. Both cases show the boundary being tested is not whether teachers use AI to save time, which is now routine, but whether technology starts to stand in for the teacher in front of the class.¹³

What the rules actually say

For schools in England, there is no law on AI in the classroom and no requirement to use it. The Department for Education's position paper, first published in 2023 and last updated in August 2025, says generative AI can help teachers with planning, resources, marking and administration, but that pupils' use must be supervised, filtered and appropriate to their age; that schools remain responsible for accuracy and safeguarding; and that AI tools must not be trained on pupils' work without permission. Keeping Children Safe in Education, the statutory safeguarding guidance, now treats generative AI as part of a school's online-safety duties. The department has also funded free training modules, written by Chiltern Learning Trust and the Chartered College of Teaching and updated in May, covering what AI is, how to use it safely, and where it can and cannot help.⁶

For the companies selling to schools, the department has published product safety standards: filtering and monitoring, no harmful or age-inappropriate content, security and data protection, testing before release. The January 2026 update added standards on cognitive and emotional development, mental health and manipulation, reflecting worries that chatbots built for engagement are not the same as tools built for learning. The standards are voluntary, but the AI tutoring programme requires bidders to meet them, and the government's AI incubator is building what it calls a sovereign education benchmark to test whether models teach safely and accurately before they are trialled.¹⁴⁷

Ofsted, for its part, visited 21 early-adopter schools, colleges and trusts and published its findings in June 2025 under a title borrowed from one of them: "The biggest risk is doing nothing". It found that almost every setting had an AI champion on staff, that leaders saw data protection, bias, safeguarding and intellectual property as the foundations of responsible use, and that the most common uses were the same as Teacher Tapp reports: planning, resources, assessment and parent communication. Ofsted says schools are under no obligation to adopt AI, but it expects leaders to know that pupils and staff are probably already using it and to manage that.⁸

The AI tutoring programme

The government's flagship intervention is the AI Tutoring Tools Pioneer Programme, announced by the Department for Science, Innovation and Technology and the DfE on 16 April. Up to eight companies were offered £300,000 each to co-design tutoring tools with schools over the summer term, for English, maths, science and modern languages in Years 9 and 10, with the aim of supporting up to 450,000 disadvantaged pupils by 2027. Bidders had to meet the product safety standards, could not publish pupil data and could not train models on pupils' work without parental permission. The procurement notice makes clear the contracts are for research and development rather than for buying tutors: the outputs are meant to define what a safe, curriculum-aligned AI tutor looks like, with the findings shared through a DfE "reading room" so that other suppliers do not have to repeat the work.¹⁵

The government's Incubator for AI, which runs the programme with the DfE, now says eight pioneers have been selected and are working with partner schools to iterate their products, with trials to be run through the DfE's EdTech Testbed. As of this week, it had not named them on its website, and we could not find a public list. Alongside the tutors it is building a machine-readable digital national curriculum with Oak National Academy, so that tools can plug directly into what is being taught, and it has set up an educational AI taskforce of pedagogy, safety and AI experts. The £23m for testing AI and edtech sits with a wider £325m for school connectivity to 2029-30.¹⁴

Does any of it work?

The honest answer is that the evidence is early and mixed, and much of what is marketed to schools has never been tested. The strongest UK result so far comes from the Education Endowment Foundation's randomised trial of Maths-Whizz, an adaptive online tutor, in 63 primary schools with more than 10,000 pupils in Years 2 to 5. Pupils allocated to the programme made the equivalent of one additional month's progress in maths, and pupils eligible for free school meals made two. The EEF has designated it a promising programme, a label given to around one in ten of the interventions it evaluates. The tool is adaptive rather than generative, and it was used in addition to normal teaching, not instead of it.¹¹

On the teacher side, an EEF trial in which secondary science teachers used ChatGPT with a short guide found they spent about 31% less time preparing lessons, roughly 25 minutes a week, with no measurable drop in the quality of the resources. The EEF is now running a much larger trial of Aila, Oak National Academy's free AI lesson assistant, with around 450 Key Stage 2 teachers in 86 schools, evaluated by NFER, with results expected this autumn. That will be the first large test of whether an AI planning tool built for the English curriculum actually changes anything for pupils rather than just for teachers' evenings.¹²

Against this sit the worries that run through the Royal Society's evidence, the Seldon report and Ofsted's fieldwork: that pupils will offload thinking rather than learn, that tools hallucinate, that access will be unequal, and that engagement-optimised chatbots may harm younger users. None of these has yet been tested at scale in British schools. The Seldon report's own finding that nearly half of young people would trust an AI over a person to check a fact is, depending on your view, either the strongest argument for teaching AI literacy or the strongest argument for keeping AI out of the classroom.²¹⁰

What Seldon and Bunting want, and what the critics say

The report's ten-point plan would go much further than the government has. It asks for a rapid response unit on AI inside the DfE by 1 November, led by outside experts and modelled on the Covid vaccine taskforce; a national framework, with every school required to have a governor-approved AI strategy by September 2027 and AI champions among both staff and pupils; Ofsted to assess those strategies and downgrade schools without one; a prime ministerial speech and a public information campaign; and a "two-lane" assessment system that keeps terminal exams but adds continuous, AI-supported assessment of what pupils can do. The authors argue the education system is "trying to teach for an AI-less world that no longer exists", and warn of "psychological absenteeism" if pupils stop seeing school as a legitimate use of their time.³¹⁹

The most contested part is data. The report proposes a federated model in which anonymised pupil data is held at school level and moves with the pupil, institutional "data lakes" for predictive analytics, and an ambition to exchange data with other high-performing countries within five years. It imagines collecting performance down to sub-question level, keystrokes and even the temperature of the exam room. Data protection specialists have pointed out that it does not address the lawful basis for such collection, the role of the Information Commissioner, consent, parental rights or the rules on international transfers. The report itself acknowledges a risk of "dehumanisation" if data displaces relationships.¹⁹

The DfE's response has been to list what it is already doing. A spokesperson said the government takes the opportunities and challenges of AI "incredibly seriously", pointed to the curriculum refresh, new vocational routes, the AI tutoring tools and the £23m for testing what works, and did not commit to a taskforce, a national framework or mandatory school strategies. Teaching unions, meanwhile, have their own red lines. The National Education Union's advice is that AI must not be used to intensify workload or to replace educators, and that teachers should keep professional judgement over whether and how to use it in their classrooms; nine unions have jointly asked to be involved "at every stage" of AI adoption.²¹⁸

Exams and the curriculum

For GCSE and A-level coursework, the rules come from the Joint Council for Qualifications, whose guidance says pupils must acknowledge any AI use, that passing off AI-generated work as their own is malpractice, and that sanctions can include disqualification. Exam boards have been guarded about using AI to mark, and Ofqual, the regulator, is among those that have given evidence to the Commons inquiry on how AI should and should not be used in assessment.²⁰

Longer term, the curriculum is changing. The government's response to the Curriculum and Assessment Review, published in November 2025, accepted that pupils should learn about AI: citizenship will become compulsory in primary and secondary schools with media literacy at its core, the computing curriculum will be refreshed to include AI and data, the computer science GCSE will be replaced with a broader qualification, and a data science and AI qualification for 16 to 18-year-olds is being considered. None of it arrives before 2028, which is the gap Seldon and Bunting are pointing at: the pupils using ChatGPT for tonight's homework will have left school by the time the curriculum catches up.¹⁵

Scotland, Wales and Northern Ireland

Education is devolved, and the other nations have moved in their own ways. Scotland published guidelines and guardrails for AI in schools in March 2026, written jointly by the Scottish Government and the EIS teaching union and grounded in children's rights, with the ministerial foreword stating that AI "can never and should never act as a replacement for human interaction within classrooms"; Qualifications Scotland has a separate position on AI in assessment. Wales has a national approach built around Hwb, its digital learning platform, with guidance, a policy template for schools and an Estyn review of how generative AI is being used. Northern Ireland's Department of Education has issued guidance by circular to schools. The result is four sets of guidance and, so far, no binding rules anywhere in the UK.¹⁶¹⁷

What to watch

Four things will show whether this autumn's argument changes anything. First, the Education Committee's AI and EdTech inquiry, which began taking oral evidence on 8 September and will report in the coming months; its recommendations will land on a department that has so far preferred guidance to mandates. Second, the results of the NFER trial of Aila, due this autumn, and the first classroom trials of the eight AI tutoring pioneers, which will be the first real test of whether tools built to the DfE's standards do anything for attainment. Third, whether the government names the pioneers and publishes what the reading room learns. And fourth, whether Ofsted's new inspection framework says anything about AI at all, which is the lever Seldon and Bunting want pulled hardest.⁴⁸¹²

The one thing every side agrees on is that waiting is not an option, because pupils are not waiting. A 12-year-old at a Liverpool academy told the BBC this week that she already uses AI to help with her homework. Whether her school, and the system around it, has a plan for that is the question the report was written to force.²