“Does it matter?… I don’t think it does.”
These were the words of Jensen Huang, chief executive of Nvidia, in September 2026. They were in response to the findings of new research on pupils learning with AI.
Eighteen months earlier, few would have thought to ask the question.
2025: The Promise
In 2025, there was a justifiable case for folding generative AI into schools. The promise went well beyond reducing workload for teachers; harnessing the technology was seen as the path to revolutionise education and transform learning itself, with a personal tutor for every child enabling them to make learning gains never seen before.
Sal Khan claimed AI would democratise and scale one-to-one learning with patient Socratic tutors and the evidence at the time appeared to back his theory. In May 2025, a meta-analysis by Wang and Fan, published in a Nature Portfolio journal, combined 51 studies and concluded that ChatGPT had a large positive effect on learning performance and improved students’ higher-order thinking. It became one of the most influential studies on AI in education, being accessed almost half a million times and cited by more than 500 papers. EdTech companies seized upon it to sell AI products and it appeared in both vendor sales decks and policy memos. This same year also saw the publication of a World Bank pilot in Nigeria reporting that learning with Microsoft Copilot resulted in close to two years of progress in only six weeks. A Harvard trial with university students, also published in 2025, reported them learning twice as much in less time with an AI tutor, providing corroborating evidence for Khan’s claims.
The evidence at the time looked compelling and reinforced the warning that failing to embrace AI was tantamount to failing pupils, leaving them unprepared for an AI-driven workplace. AI in schools began to feel inevitable. The teachers and policymakers supporting it were reading the evidence in front of them, and they were right to take it seriously.
The Collapse
Within two months of its launch, ChatGPT became the fastest-growing consumer product in history. New models now arrive at such a pace that the usual independent research and regulation cannot keep up.
The promise led to a flood of enthusiasm and a scattered, contradictory evidence base, rife with poorly designed studies and research funded by the companies selling the tools.
But buried in it is a growing body of independent, well-designed studies that give us some much-needed clarity. And what it shows us does not support the case made for AI in schools.
Evidence for the benefits of learning with general-purpose chatbots, the kind that most children now use, rested on numerous small, short-term studies. The most influential evidence, however, came from the Wang and Fan meta-analysis and the Nigerian pilot. Despite its reach, the Wang and Fan meta-analysis was retracted in April 2026 over discrepancies serious enough for the editor to no longer have confidence in its conclusions. One leading educational researcher said it “should not have been published in the first place”, yet the 500-plus papers that cited it remain in print, carrying its claims into new research and sales pitches long after the original was withdrawn. The randomised Nigerian pilot still stands, but the claims attributed to it rested on an unfair comparison; the pupils using AI were given extra after-school sessions with Microsoft Copilot but were compared with pupils given no additional sessions at all. Crucially, the researchers never checked whether the learning gains lasted.
AI tutors built specifically for education have fared no better. The Harvard study involved fewer than 200 university students across just two lessons, with an AI tutor fed the correct answers in advance. And again, the students were only tested the moment the lesson ended.
Larger and longer trials with school pupils have also failed to demonstrate learning gains beyond what a pupil can achieve without AI.
In a trial of almost 1,000 secondary pupils by Bastani and colleagues, those using a Socratic AI tutor produced no better results than pupils working with textbooks. Khan Academy’s own AI tutor, Khanmigo, was finally put to the test in a two-year trial across 18 American middle schools. It too failed to produce gains any greater than ordinary practice using Khan Academy without AI.
In October 2026, results came out from a trial in ten English state schools. It was run by the London-based company Eedi with Google DeepMind, and 1,179 pupils took part. The AI tutor was as good as human tutors at correcting maths errors as they happened. After ten weeks, though, the pupils who used the AI tutor did no better on a standard maths test than pupils who simply completed practice questions in the app with no tutoring at all.
These are not a few selected studies. In March 2026, Stanford researchers set out to assess ‘the evidence that does exist’ on AI in schools. They reviewed over 800 papers but found only 20 to be robust enough to base their findings on. Of those, only one study checked whether pupils remembered what they had learned days later, and again, it found no benefit for children using AI as a learning tool. What that study did find was something far more alarming, and it is now emerging as a pattern across new research.
2026: New Research, New Understanding
If AI use fails to produce lasting learning gains for children, the case for restructuring education around it cannot stand, and investing in it is impossible to justify. But the latest evidence points to a deeper problem:
AI can make pupils appear to be learning while leaving them knowing less than those who never used it.
In the same Bastani trial, pupils who worked through practice questions using a ChatGPT-style chatbot, unsurprisingly, got more questions right than those with no access to AI. The real test came when they sat an exam without it. They scored 17% lower than students who never used AI and only practised with textbooks.
Until 2026, the studies showing the harm to pupils had only measured it over days or weeks. In June 2026, we began to see what years of everyday AI use does to them. The research by Strömberg and colleagues followed more than 26,000 secondary students over the course of two and a half years and the findings were not reassuring. Those who began using AI for their homework completed it 30% faster than before and their marks rose by 18%. But in exams, sat without AI, their scores fell. The damage appeared within six months across monthly tests that covered more recently acquired knowledge. In the entrance exams, however, which covered years of material, the damage took two years to show in full, at which point average scores had fallen by up to 24%. The more the pupils used AI, the more their unaided scores dropped, and despite having access to freely available tutoring tools, most preferred chatbots that simply provided the answers. The researchers warned that existing short-term studies may be underestimating the long-term cost.
In October, researchers in the US published a randomised trial with 2,379 university students. The students given the AI tutor built into their course got lower final grades than the ones who weren’t. The difference was large, 0.37 of a standard deviation, and was biggest for students who were the first in their family to go to university.
The one study in Stanford’s review that tested what pupils remembered days later was carried out in England and published in 2026. Kreijkes and colleagues compared Year 10 pupils who learned with a chatbot against pupils who only took notes. They found that even though pupils in the AI group preferred using the chatbot, finding it more helpful, it was the pupils who took notes with no AI who understood and remembered significantly more.
Similarly concerning findings had been accumulating across smaller and earlier research, mostly involving adults. These were early warning signs of the risks now emerging across larger trials involving pupils. One study found university students who studied with ChatGPT remembered less 45 days later than those who didn’t, while another showed adults who used AI for only ten minutes performing worse and giving up more quickly once it was taken away. In another study of over 3,000 adults, those without AI said ‘I don’t know’ to difficult questions 44% of the time, whereas those with access to AI only said it 3% of the time. When the AI was wrong, the user’s accuracy dropped threefold while their confidence, despite the answer being incorrect, more than doubled. In other research, surveys have linked heavy AI use with weaker critical thinking, with younger users showing the heaviest reliance and the lowest scores. This concern is now shared by 66% of secondary teachers. Researchers are now circling this phenomenon at its most severe: cognitive surrender, where a person stops evaluating and accepts the AI’s answer instead. An adult with enough knowledge of their own can sometimes catch the error. A child who is still learning it is far less likely to.
Even those who build this technology had noticed. As early as July 2025, Sam Altman, OpenAI’s chief executive, told a Federal Reserve conference: “There’s young people who say things like,
‘I can’t make any decision in my life without telling ChatGPT everything that’s going on. It knows me, it knows my friends. I’m gonna do whatever it says.’
That feels really bad to me.”
Others are less troubled. The study the journalist Ezra Klein put to Jensen Huang was Strömberg’s. Huang agreed with its findings before asking whether it mattered. We are, he said, “going to lose some… intellectual dexterity” but become better “systems thinkers”. When Klein pressed that there must be some set of skills that matter, Huang replied, “Yeah, but maybe not those. We’re going to discover new ones.” It was a familiar promise and one that remains unproven: AI freeing up higher-order thinking. Wang and Fan claimed the same in their now-retracted meta-analysis. But even on Huang’s own terms, it is a promise only realised if the pupil stays plugged in.
Lost learning is not the only cost, and it is not shared equally.
AI tutors have been heralded as the solution to support the most vulnerable and disadvantaged students in the UK, yet new research highlights that they may actually further entrench inequality. In March 2026, Tan and colleagues gave AI models identical essays, changing only a brief description of the student. The feedback changed with it. Responses for girls were more likely to contain language emphasising feelings, empathy and solidarity such as “Girls like us” and “I love your confidence in expressing your opinion!” while responses for boys were more likely to contain language around reasoning and evidence. White students were assumed to be more competent writers and given feedback to develop their arguments while ethnic minority students received less constructive criticism, more reference to their culture and in some cases, more corrections to basic spelling and grammar. Lower-attaining, EAL and SEND students were also more likely to receive surface-level feedback while higher-attaining students were coached on the ideas. Models have also been shown to give less accurate answers to users who come across as less educated or as non-native English speakers. AI does not even need to be told explicit information about a pupil. Other studies have found that models assume a user’s identity from far subtler clues, such as a name, a style of writing or even the type of school a pupil attends, and then respond in ways that reflect racial stereotypes. The children these AI tutors were meant to lift are the same ones they are holding down.
Changing Course
These findings do not stand alone. Even before the publication of most of the alarming studies, in January 2026, the Brookings Institution, one of the world’s most influential policy institutes, published a year-long review drawing on more than 400 studies and over 500 interviews across 50 countries. It concluded that “the risks of utilizing generative AI in children’s education overshadow its benefits”. The OECD’s PISA 2025 results came out in September. Fifteen-year-olds who used AI chatbots for schoolwork, for summarising texts or drafting their writing, scored about 20 points lower in science than students who didn’t. Such a drop equates to roughly a year of schooling. The OECD added that this doesn’t prove AI caused the gap, and part of it could be that lower-attaining students use AI more, but it does clearly line up with what other recent trials are demonstrating. Any position statements or school-based AI policies written before 2026 likely reflect the best evidence available at the time, but that evidence has now changed, and attitudes and responses must change with it.
Whether or not AI can help a pupil complete their next assignment is not the relevant question. What we must ask is, after a thousand assignments completed with AI support, can that pupil become more capable of learning and problem-solving on their own? Longer-term studies already point to an answer, but waiting for conclusive proof would mean running this experiment on children in schools now.
The burden of proof should not belong to the most vulnerable in our society. It belongs to those who build and sell this technology.
Children’s use of AI has soared. In the US, the percentage of teenagers using ChatGPT for schoolwork doubled in a single year. Now, 97% of UK children aged 8-17 are using AI, including voice assistants and AI summaries. Outside of school, ChatGPT, Gemini, Copilot, Grok and Perplexity are free to access on any browser or as free apps. Most of these services set a minimum age of 13, but these age limits only exist on paper: a February 2026 age-gating study found none of the chatbots tested blocked under-13s. Children who use WhatsApp have access to Meta AI, as do those on Instagram or Facebook. Snapchat has My AI, a chatbot that sits among a child’s conversations with their friends. It can be cleared from the feed, but everything a child has already disclosed to it remains with Snapchat. Even a Google search is now conversational: the AI Overview that appears above the results lets a child ask follow-up questions, each building on the last, and there is no setting to switch it off.
Children will use these tools at home, on their own terms, whatever schools decide, and this technology will continue to outpace our safeguards.
In September 2026, Apple and Meta launched AI agents that act on a user’s behalf. Apple’s new Siri can search a user’s messages and emails and act across their apps, and is already out in beta for anyone aged 13 or over with a recent iPhone. Meta’s Muse goes further: it can send emails and make purchases, keeps working on tasks while the user is elsewhere and starts conversations of its own. It is wrapped up in the form of a plush toy-like character, a suspect design choice for an adult product, and it sits in WhatsApp as just another contact. Muse is only for adults, but WhatsApp itself is officially for over-13s despite being used by more than half of 8-12 year olds. These agents will plausibly be able to do a child’s homework along with much of what children currently do online. For now, children’s access to agents is limited; most remain behind paid subscriptions or are built into newer iPhones that most children don’t have. But in 2026, they began arriving free inside messaging apps. Muse has so far been rolled out only in the US and Canada but can be expected to reach the UK in due course. Once agents are as accessible as conversational AI, their use by children will likely become as prevalent. No study has measured the harms these agents pose to children and none will for years, but if conversational AI is leading children to outsource their thinking, then we must pre-empt the possibility of these agents leading children to outsource their agency.
Access to generative AI does not just occur during homework outside of school hours; it is increasingly available throughout the school day. The vast majority of people in the UK search the internet through Google Search and its regular use by pupils continues to put them in contact with AI Overviews and conversational AI Mode. In November 2026, Google announced it won’t offer schools a version of Google Search without AI, so any school that intends to remove AI will need to give up on Google Search. Copilot can also be accessed by teenage students through their school Microsoft accounts when schools enable it. The percentage of secondary teachers who say a pupil has used AI without permission in their lessons has risen from 15% in 2023 to 58% this year, yet two-thirds of teachers say their school has no policy on pupils’ use of AI.
Much of the pressure to adopt AI comes from outside schools. But we can’t wait for regulators, or even the developers themselves, to catch up with the evidence. School leaders and staff can act now.
They are in a unique position to draw a line around what they can control, which is the tools they put into pupils’ hands and the settings they leave switched on.
With no demonstrated lasting benefit to weigh against the risks, pupils’ use of generative AI should be paused in primary and secondary schools until the evidence shows otherwise. For this technology to be placed into children’s hands, we need independent trials with children showing they learn more with AI than they would from good teaching without it. We need robust evidence that those gains last and that it does not leave pupils worse off once it is removed. Jensen Huang asked whether it matters that children are losing these skills. For every parent and teacher who believes children should learn to think for themselves, it does.
Sources
Opening
2025: The Promise
- Sal Khan (2023), How AI could save (not destroy) education, TED
- Wang and Fan (2025), The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking, Humanities and Social Sciences Communications
- AI Weekly (2026), Springer Nature pulls cited ChatGPT-in-education meta-analysis
- De Simone et al. (2025), From chalkboards to chatbots: Transforming learning in Nigeria, World Bank
- Kestin et al. (2025), AI tutoring outperforms in-class active learning, Scientific Reports
The Collapse
- Reuters (2023), ChatGPT sets record for fastest-growing user base
- Humanities and Social Sciences Communications (2026), Retraction note: Wang and Fan
- Computing (2026), Publisher withdraws study claiming ChatGPT boosts learning
- TechSpot (2026), Major ChatGPT education study retracted after flaws found
- Plagiarism Today (2026), Study claiming AI helps students learn retracted
- De Simone et al. (2025), Policy Research Working Paper 11125, World Bank
- Bastani et al. (2025), Generative AI without guardrails can harm learning, PNAS
- Oreopoulos and Low (2026), One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment, NBER Working Paper 35620
- The 74 (2026), Five Things You Should Know About New AI Tutor Study (Eedi and Google DeepMind trial)
- Barnard, Fesler and Loeb (2026), Artificial Intelligence in K–12 Schools, Stanford SCALE
- Stanford SCALE (2026), The Evidence Base on AI in K-12: A 2026 Review
2026: New Research, New Understanding
- Strömberg, Lei and Wu (2026), The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, CEPR Discussion Paper 21577
- Strömberg, Lei and Wu (2026), The generative AI learning penalty in secondary school, VoxEU
- Liu, Sweet et al. (2026), The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial, EdWorkingPaper 26-1598
- Kreijkes et al. (2026), Effects of LLM use and note-taking on reading comprehension and memory, Computers & Education
- Barcaui (2025), ChatGPT as a cognitive crutch, Social Sciences & Humanities Open
- Liu et al. (2026), AI assistance reduces persistence and hurts independent performance, preprint
- Marcoccia, Quattrociocchi and Capraro (2026), AI advice suppresses people’s willingness to say “I don’t know”, preprint
- Gerlich (2025), AI tools in society: Impacts on cognitive offloading and the future of critical thinking, Societies
- Lee et al. (2025), The impact of generative AI on critical thinking, CHI 2025
- NEU (2026), State of Education: AI
- Shaw and Nave (2026), Thinking: Fast, Slow, and Artificial, preprint
- Business Insider (2025), Sam Altman is worried some young people have an “emotional over-reliance” on ChatGPT, Federal Reserve conference, July 2025
- DfE (2026), 450,000 disadvantaged pupils could benefit from AI tutoring tools
- Tan, Phalen and Demszky (2026), Marked Pedagogies: Examining linguistic biases in personalized automated writing feedback, LAK 2026
- Poole-Dayan, Roy and Kabbara (2026), LLM targeted underperformance disproportionately impacts vulnerable users, AAAI 2026
- Haim, Salinas and Nyarko (2024), What’s in a Name? Auditing large language models for race and gender bias
- Hofmann et al. (2024), AI generates covertly racist decisions about people based on their dialect, Nature
- Warr, Pivovarova, Mishra and Oster (2024), Is ChatGPT racially biased? The case of evaluating student writing
Changing Course
- Burns, Winthrop et al. (2026), A New Direction for Students in an AI World, Brookings
- OECD (2026), PISA 2025 Results (Volume I): Future-Ready Students
- Schleicher (2026), Generative AI and student learning: What the PISA data reveal, OECD Education and Skills Today
- Pew Research Center (2025), About a quarter of U.S. teens have used ChatGPT for schoolwork
- Childnet and Nominet (2026), Almost all young Brits use AI: Safer Internet Day 2026, UK Safer Internet Centre
- Actions Speak Louder Than Chats (2026), Investigating AI chatbot age gating, preprint
- The Star (2025), WhatsApp to get major AI update in the UK and you can’t turn it off
- Snapchat Support, How do I unpin or remove My AI?
- TechCrunch (2026), Google now lets users jump from AI Overviews into AI Mode conversations
- Apple (2026), Siri AI, a profoundly more capable and personal assistant, is here
- Meta (2026), Introducing Muse
- Engadget (2026), Meta reveals its AI agent that can shop, send emails and plan trips on your behalf
- TechCrunch (2026), Everything new coming to Meta’s AI agent Muse
- Neoteo (2026), Meta bars minors from Muse
- Ofcom (2026), Children’s Online Experiences
- StatCounter (2026), Search engine market share, United Kingdom
- AP News (2026), Google will cut off search access for students at schools that reject AI
- Teacher Tapp (2026), Neurodivergent teachers, cheating with AI, devices in schools




