Is AI good for students? A balanced look at the evidence

You've probably read that AI is going to transform education, and you've probably read that AI is destroying critical thinking. Both are pitched with the same certainty, and neither matches what the primary evidence shows in September 2026.

The honest read is more useful and less quotable. AI use among UK and US students is now mainstream: more than half of UK 8 to 17 year olds have used AI (Ofcom, 2026) and 70% of US teens use it for schoolwork (Common Sense Media, 2026). But the evidence on whether it helps learning is thin, and the strongest signals in the data are about how students use it, not whether they should. This piece pulls together what the primary sources say without hyping either direction, so parents, teachers and students can decide for themselves.

The uptake numbers are the least contested part of this discussion, because four independent surveys point the same way.

Ofcom's 2026 Children and Parents: Media Use and Attitudes report (21 May 2026) found that 56% of UK 8 to 17 year olds had used AI, rising to two-thirds of 16 to 17 year olds, and half of those users (51%) used it for schoolwork or homework. Ofcom also found teenage AI users split on whether they would trust an AI-written news article as much as a human-written one.

HEPI's Student Generative AI Survey 2026 (fieldwork December 2025, published 12 March 2026, n=1,054) reported that 95% of UK undergraduates used AI in some form in 2025/26, and 94% used generative AI for assessed work. The share pasting AI-generated text directly into assessed work rose to 12%, from 8% the year before and 3% in 2024.

Pew Research's How Teens Use and View AI (24 February 2026, n=1,458 teens aged 13 to 17) found that 54% of US teens used chatbots for schoolwork help. Common Sense Media's Teens in the AI Era (August 2026) put the schoolwork figure at 70%.

So the volume question is settled. AI is embedded in how students study. What's less clear is whether that's helping.

This is the part of the debate where certainty most often outruns the data.

Ofsted's June 2025 study of 21 early-adopter schools and further education colleges is the most substantial UK piece of work on AI in classrooms so far. Its conclusion is that "robust and reliable evidence of AI's impact on educational outcomes is limited", with most existing research described as "explorative, short-term and in limited domains". Ofsted also confirmed there is no stand-alone section of school inspection focused on AI at this stage.

The Department for Education's Generative Artificial Intelligence in Education paper (updated 12 August 2025) sits alongside this. Its position is that generative AI "cannot replace the judgement and deep subject knowledge of a human expert", and pupil use should be "carefully considered and assessed". The DfE expects schools to comply with the age restrictions set by AI tools, which vary: some require users to be 13 or over with parental consent, while others, such as Claude, are 18+.

What we don't have, at time of writing, is a large-scale primary study showing AI use improves grades at GCSE, A-Level, or AP. Any headline claim that AI "improves learning outcomes" isn't currently supported by primary evidence at the school level.

The most credible use case isn't for students at all; it's for teachers.

The Walton Family Foundation and Gallup's 2025 research, Six Weeks a Year, found that 32% of US teachers used AI at least weekly, and those who did saved an average of 5.9 hours per week, roughly six weeks over the school year. Ofsted's early-adopter research found that reducing teacher workload was one of the two main reasons schools introduced AI, alongside supporting pupils' learning.

For students directly, the evidence is thinner but not empty. HEPI 2026 found that 49% of UK undergraduates said AI had improved their student experience, and 68% viewed AI skills as essential for thriving today. The most common student uses are explaining concepts, summarising material and suggesting research ideas. Those are patterns that map onto legitimate learning activities, provided students do the retrieval themselves rather than outsourcing it.

Two concerns come out of the primary evidence in a form that survives scrutiny.

The first is offloading, which shows up clearly in the usage data: Common Sense Media's August 2026 report found that 70% of US teens used AI for schoolwork, and 63% of US teens who use AI say they get answers to schoolwork from AI; 25% say they use those answers as-is, while others rewrite them (31%) or improve them with what they know (35%). That's a use pattern, not a proven cognitive harm, and it's worth being careful about the distinction. But the pattern is what generates the worry. If students are consistently using AI to skip the retrieval and thinking parts of a task, the parts that build memory and understanding in the first place, that's a real reason for parents and teachers to pay attention.

HEPI's report quoted a UK undergraduate saying "I'm not using my brain at all". That single quote gets repeated a lot, and it's one person's experience, but it captures something the wider data supports: students themselves notice the offloading, and some of them don't like what it does to their thinking.

The second is trust and accuracy. Ofcom's 2026 report found that 40% of teenage AI users would trust an AI-generated news article as much as or more than one written by a human. That's some healthy scepticism given how confidently AI hallucinates. But many of the Common Sense teens get answers from AI, which means relying on those answers to some extent. The gap between how much students trust AI and how much they should is where mistakes happen.

The DfE's position for schools is cautiously supportive, not enthusiastic. Its paper is careful to say generative AI can't replace teacher judgement, and it flags safeguarding, data protection, and age-gate concerns as active issues schools have to manage. It doesn't recommend blanket adoption or blanket bans.

For exams and coursework, JCQ's AI Use in Assessments guidance (30 April 2025) is the binding document. Undeclared use of AI-generated content in coursework or non-exam assessment counts as malpractice. Ofqual recorded 100 cases of AI-related plagiarism in summer 2025, up from 85 in 2024. That was three-quarters (75%) of all student plagiarism cases but only 2% of the 5,025 student malpractice cases. Mobile phone and communication device offences remained the largest category at 44.3%.

In the US, College Board's AP Exam Terms and Conditions state: "No part of your Test answers, work, or submission ... may be generated in whole or in part using any AI tools, including but not limited to ChatGPT." Its separate AI guidance sets course-specific rules for AP Seminar, AP Research, AP Art and Design and AP Computer Science Principles. The two frameworks are not equivalent, and readers shouldn't treat them as such.

So is AI good for students? The honest answer is: it depends on how it's used, we don't yet have strong evidence either way on outcomes, and the biggest risks and benefits are pattern-of-use questions rather than tool questions.

For parents, the practical version is that an outright ban at home may be hard to enforce given how widespread AI use is, and official guidance for schools favours careful, supervised use over blanket bans. Understanding how your child is using it, whether they're doing the thinking or offloading it, matters more than whether they're using it at all.

For teachers, the workload use case is the strongest evidence-backed one. The Walton/Gallup survey, in which teachers who use AI weekly report saving about 5.9 hours a week, is the clearest benefit signal so far, though it is self-reported. For direct student use in class, the DfE's "carefully considered and assessed" framing is the right register.

For students, the risk isn't the tool. It's the pattern of using it to skip thinking. If AI is doing the retrieval, the problem-solving, or the writing for you, you're not building the skill your exam measures. If it's helping you understand something you couldn't otherwise, and then you're doing the retrieval yourself, that's genuinely different.

For a wider view of what's happening across schools, exams and policy this year, see AI in education 2026: What teachers, parents and students should know.