{"id":247426,"date":"2025-11-06T14:06:11","date_gmt":"2025-11-06T14:06:11","guid":{"rendered":"https:\/\/www.newsbeep.com\/uk\/247426\/"},"modified":"2025-11-06T14:06:11","modified_gmt":"2025-11-06T14:06:11","slug":"should-llms-be-allowed-in-the-classroom","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/uk\/247426\/","title":{"rendered":"should LLMs be allowed in the classroom?"},"content":{"rendered":"

This article is an on-site version of our The AI Shift newsletter. Premium subscribers can sign up here<\/a> to get the newsletter delivered every Thursday. Standard subscribers can upgrade to Premium here<\/a>, or explore<\/a> all FT newsletters<\/p>\n

Hello and welcome back to The AI Shift. This week: what\u2019s the positive case for AI\u2019s role in education? We have read (and also written<\/a>) plenty this year about the danger that AI poses to young people\u2019s ability to learn. So today we thought we\u2019d take a more constructive approach. Are there any useful lessons from history? And what do we know so far about whether \u2014 and how \u2014 AI can be used to enhance the learning process? Email us: aishift@ft.com<\/a>. <\/p>\n

Sarah writes<\/p>\n

I often meet academics who believe their students are leaning far too heavily on large language models for their essays or coding assignments. But a popular counter-argument compares this debate to one we\u2019ve had before: about the pocket calculator. These new machines also prompted a \u201cmoral panic\u201d about education in the 1970s, the argument goes. But young people do still acquire maths skills. In fact, they can now do more complex and better maths, thanks to the tools at their disposal. <\/p>\n

Sam Altman likes this analogy. He has described ChatGPT as a \u201ccalculator for words\u201d and argued<\/a> the fears that pocket calculators would ruin maths education proved completely overblown. <\/p>\n

I think there are important ways in which generative AI is not, in fact, like a calculator. For one thing, calculators are deterministic. Two plus two is always four, no matter how many times you try it, and no matter which company made the calculator. Generative AI is nothing like this. (For more on the problems with the calculator analogy, I recommend this piece<\/a> by political science professor Blayne Haggart.)<\/p>\n

But even if we do accept the calculator analogy, I don\u2019t think it implies \u2014 as Altman seems to suggest \u2014 that we can safely shrug off concerns about students\u2019 over-reliance on AI. Rather, educators have had to carefully figure out how and when to integrate calculators into the way people learn maths. Some of those lessons can be useful when we think about generative AI. <\/p>\n

Firstly, we don\u2019t let kids have access to calculators too early. My six-year-old is learning mathematical concepts and mental arithmetic at school. It will be many years before the curriculum allows her anywhere near a calculator. Secondly, many school maths exams feature papers in which calculators are allowed, and also exams in which they are not allowed. Thirdly, US research<\/a> suggests that students do better at college-level maths if their high-school teachers restricted calculator use until they mastered concepts with pen and paper methods. <\/p>\n

This tallies with various conversations I have had with thoughtful critics of AI in education. Lakshya Jain, a machine-learning engineer (whose essay<\/a> about AI use by his computer science students at UC Berkeley is well worth reading), told me a chunk of his students \u201cseem to believe that, \u2018you know what, it\u2019s fine if it gets done by AI, as long as it gets done\u2019. And I tell them: \u2018look, that may be fine when you\u2019re a professional, but it\u2019s not fine when you\u2019re a student, because as a student the point of getting a task done is not so I can get 400 partially-completed code-bases, it\u2019s so that you can build neural connections\u2019.\u201d <\/p>\n

John, I\u2019m quite persuaded by the view that AI might have its place in education, but not until students have mastered key skills and the ability to think critically without it. But do you think that\u2019s too conservative? Is there a case for embedding AI in education from an earlier age, and if so, how?<\/p>\n

John writes<\/p>\n

You\u2019ve raised several good reasons to be concerned here, Sarah. In a domain as important as education, starting from a position of precaution is always wise.<\/p>\n

I should also say that the examples you provide of people using AI to short-cut the learning process are borne out in larger-scale representative data. Starting last year the UK\u2019s Higher Education Policy Institute (HEPI) has been surveying university students<\/a> on whether and how they\u2019re using AI as part of their studies, and between 2024 and 2025 the share who report submitting LLM-written text in their assessments doubled from around one in eight to more than one in four. And this is just the share who admit this in a survey \u2014 the true level is surely much higher.<\/p>\n

<\/p>\n

Some content could not load. Check your internet connection or browser settings.<\/p>\n

\"\"\/<\/a><\/p>\n

So I won\u2019t argue against any of that on the risk side, but I do think it\u2019s worth taking a look at some of the \u2014 admittedly tentative \u2014 evidence that AI might also be able to add real and durable value for both students and educators.<\/p>\n

First, education research consistently finds that the best results<\/a> come when instruction can be tailored to an individual\u2019s prior knowledge<\/a> and capabilities, with regular feedback provided along the way: in other words, one-to-one tutoring. It goes without saying that most students aren\u2019t able to get personal tuition, so could LLMs\u2019 flexibility, interactivity and capacity to synthesise and reformulate information provide them with an AI equivalent, and would the results be comparable?<\/p>\n

We\u2019ve now had several experiments testing this. The results are mixed, though there are some consistent patterns. Generally, where students are simply allowed unsupervised use of off-the-shelf generative AI to assist with their studies, long-term learning outcomes suffer because AI tends to be used primarily as a crutch to substitute for rather than complement<\/a> their own learning. Most troublingly, in one study, when students who had been allowed to use ChatGPT during the practice period then sat a test without it, they scored 17 per cent worse<\/a> than students who had never used it at all.<\/p>\n

But equally solid is the finding that where students were directed in using LLMs specially modified to have guardrails \u2014 typically using prompts carefully crafted by the teachers to ensure that they guide students towards key learning outcomes and do not simply provide answers \u2014 the results were generally positive<\/a> and often durable, especially when usage was also supervised in the classroom<\/a>.<\/p>\n

The second area where AI holds promise is in marking written assignments, and in doing so freeing up arguably the most scarce resource in education: teachers\u2019 time. The idea of robots judging human prose may sound preposterous, but the evidence here is strong. Humans are notoriously unreliable at assessing writing against a rubric<\/a>, and suffer from biases like being overly influenced by handwriting quality. AI is now exceptionally good at parsing even the most spidery scrawl.<\/p>\n

The result is that in large scale experiments<\/a> carried out by No More Marking, a British education technology firm that uses AI and other statistical techniques<\/a> alongside human assessment to provide standardised grading for thousands of students\u2019 writing assessments, artificial and human intelligence generally make the same judgments. When they differ it tends to be the human who slipped up.<\/p>\n

As Daisy Christodoulou<\/a>, director of education at No More Marking and author of three books on evidence-based education policy and practice put it to me, \u201cLLMs in the classroom do hold some promise, but they\u2019re highly variable and can introduce new problems, whereas using AI to automate the more tedious parts of a teacher\u2019s job \u2014 freeing up teachers to teach \u2014 is a clearer win.\u201d<\/p>\n

So what have we learned?John<\/p>\n

Last week<\/a> you wrote, Sarah, about how AI is acting as an amplifier of organisations\u2019 strengths and weaknesses. I do worry a little that the same mechanism might play out in education if students aren\u2019t properly supervised in their use of LLMs: the most motivated will 10x their learning (a wonderful thing!), but the least engaged will not merely stagnate; they\u2019ll go backwards.<\/p>\n

Sarah<\/p>\n

I wonder if that\u2019s going to become a running theme in this newsletter, John. I find the idea that AI could be useful in marking students\u2019 work interesting. I\u2019m sure teachers would love to get some time back, but I wonder if it might be a hard sell with students: \u201cyou can\u2019t use AI to write this, but I will use AI to mark it.\u201d In fact, we\u2019re seeing this same tension arise between job applicants and hirers right now (more on that next week).<\/p>\n

Recommended reading<\/p>\n

An interesting take<\/a> on chatbots and the recent Dutch election (Sarah)<\/p>\n

HEPI\u2019s collection of essays<\/a> by higher education experts on AI and the future of universities is full of thought-provoking ideas (John)<\/p>\n

Recommended newsletters for you<\/p>\n

The Lex Newsletter \u2014 Lex, our investment column, breaks down the week\u2019s key themes, with analysis by award-winning writers. Sign up here<\/a><\/p>\n

Working It \u2014 Everything you need to get ahead at work, in your inbox every Wednesday. Sign up here<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"This article is an on-site version of our The AI Shift newsletter. Premium subscribers can sign up here…\n","protected":false},"author":2,"featured_media":247427,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[554,733,4308,86,56,54,55],"class_list":{"0":"post-247426","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-artificial-intelligence","10":"tag-artificialintelligence","11":"tag-technology","12":"tag-uk","13":"tag-united-kingdom","14":"tag-unitedkingdom"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/247426","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/comments?post=247426"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/posts\/247426\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media\/247427"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/media?parent=247426"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/categories?post=247426"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/uk\/wp-json\/wp\/v2\/tags?post=247426"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}