Should AI be used in schools? What the evidence says in 2026
AI helps students learn when it acts as a tutor and hurts when it hands over answers. Here's what the studies show and how to use it well.

Should AI be used in schools? Yes, I think so, but only in tutor mode: set up to ask questions and give hints, with a teacher deciding when it's switched on. Used as an answer machine for homework, it can leave students worse off once it's taken away. The research on both sides is clearer than the shouting suggests.
Here's the short version for busy parents and teachers. The same model (GPT-4, in the two best-known studies) produced big learning gains in one setup and measurable harm in another. The difference wasn't the AI. It was the instructions wrapped around it and what the student was asked to do with it.
Should AI be used in schools? The short answer
Use it where it makes students think, and keep it away from work that's meant to prove they can think on their own. That means guided tutoring, practice feedback and teacher prep time, yes. Unsupervised chatbots doing take-home essays and problem sets, no.
Students aren't waiting for the debate to end. In a Pew Research Center survey of US teens aged 13 to 17, run from 25 September to 9 October 2025, 54% said they'd used chatbots to get help with schoolwork, and one in ten said they do all or most of their schoolwork with a chatbot's help. So the real question for most schools isn't "AI or no AI". It's which kind, for which tasks, and with what checks.
The mechanics: answer mode versus tutor mode
A general chatbot is trained to be helpful, which in practice means it finishes your task. Paste in a quadratic equation and you'll get the roots, neatly worked. That's great for a busy adult. For a 14-year-old who's supposed to be learning to factorise, it skips the exact step the homework exists to practise.
A tutor setup changes the system prompt (the hidden instructions the model reads before every reply) so it withholds final answers, asks what the student has tried, and gives one hint at a time. Some tools add more structure on top. The Harvard team behind the physics study below found a prompt alone couldn't reliably keep the tutor on track through multi-part problems, so their platform walked students through each part in order, and they fed it step-by-step solutions to cut down on wrong answers.
Several products now ship this mode by default or as a toggle:
- Khan Academy's Khanmigo is built to guide learners to the answer instead of giving it.
- ChatGPT added study mode on 29 July 2025, which uses guiding questions and step-by-step help.
- Google sells Gemini for Education to schools, with custom Gems that teachers can set up as a learning coach.
- MagicSchool is aimed at districts, with 80+ teacher tools and 50+ student tools behind teacher-controlled settings.
The weak point is obvious: a student who wants the answer can usually open a different, unrestricted chatbot in another tab. Tutor mode only works if the assignment and the classroom make it the path of least resistance.
Real examples from classrooms
Harvard physics: AI tutor beat active learning
In a randomised controlled trial with 194 students in a Harvard undergraduate physics course, published in Scientific Reports in June 2025, one group learned a topic from a custom GPT-4 tutor at home while the other got the same material in an active-learning class, and then the groups swapped for a second topic. The median learning gain in the AI group was more than double the in-class group's. Median time with the tutor was 49 minutes, against about 60 minutes of class.
The authors put the effect size somewhere between 0.73 and 1.3 standard deviations, which is huge for education research. Students also reported feeling more engaged. Still, it was two lessons, at Harvard, with a tutor its designers had tuned carefully, so I'd be cautious about assuming a generic chatbot in a crowded secondary school gets anywhere near that.
High school maths: the crutch effect
This is the study I'd make every school board read. A field experiment with nearly a thousand high school maths students, published in PNAS in 2025, tested this head-on. One group got a plain ChatGPT-style interface (GPT Base), another got a version with learning safeguards (GPT Tutor), and a third had no AI.
During practice, scores jumped: 48% higher with GPT Base and 127% higher with GPT Tutor. Then the AI was taken away for the exam. Students who'd had GPT Base scored 17% worse than students who never had access at all, while the safeguards in GPT Tutor largely cancelled that harm out. The authors' explanation is that students used the plain chatbot as a crutch.
Same model, opposite outcome.
Nigeria: after-school English with a teacher in the room
A World Bank pilot in Edo State ran a six-week after-school programme in June and July 2024, with students using generative AI as an English tutor under teacher guidance. Randomly assigned participants outperformed their peers by about 0.3 standard deviations, which the World Bank equates to nearly two years of typical learning. Girls, who'd started behind, seemed to gain more, and each extra day of attendance added to the effect.
A worked example with numbers
Take a class of 30 whose unaided test average is 60 out of 100. Here's roughly what the crutch-effect results would mean if they carried over (a big if, since one study doesn't set a rule):
| Setup | Practice score | Exam score without AI |
|---|---|---|
| No AI | 60 | 60 |
| Plain chatbot (GPT Base) | about 89 (+48%) | about 50 (-17%) |
| Tutor with safeguards (GPT Tutor) | capped near 100 (+127% in the study) | about 60 (harm largely gone) |
Look at the plain chatbot row. Homework marks rise by nearly 30 points while real skill falls by 10, and the teacher sees only the homework until exam day. That gap between practice scores and what students can actually do is, in my view, the single biggest risk of letting AI into schools without rules. It's also why grades on take-home work are becoming a poor signal of learning.
Benefits and limitations
Both lists are real. Neither cancels the other out.
| Benefits of AI in schools | Where it breaks |
|---|---|
| Patient one-to-one help at any hour, for students who'd never get a private tutor | Students use it as a crutch, and unaided performance drops (17% in the high school maths study) |
| Self-pacing: fast students move on, slow ones don't get left behind | Confidently wrong answers; the Harvard team had to supply worked solutions to keep the tutor accurate |
| Instant feedback on practice problems and drafts | Cheating at scale: 59% of US teens think AI cheating happens at least somewhat often at their school |
| Teacher time saved on lesson plans, rubrics and quizzes | Student data privacy, especially with free consumer tools |
| Big gains in low-resource settings when a teacher supervises (Nigeria pilot) | Unequal access at home, which can widen gaps between families |
Detection isn't a reliable fix for the cheating problem, either. AI detectors produce false positives, and I'd be uneasy about any school that punishes students on a detector score alone (we cover this in our piece on AI detector accuracy and in how teachers check for AI). Changing the assignment works better than policing it: in-class writing, oral checks and graded drafts.
Teens have mixed feelings too. Among those who expect AI to hurt society, about a third (34%) named overreliance or lost critical thinking as the reason. That's students describing the crutch effect in their own words.
Practical takeaways for students, parents and teachers
A quick decision guide, based on the evidence above:
| Task | Use AI? | Setup |
|---|---|---|
| Practising maths or science problems | Yes | Tutor or study mode that gives hints, not answers |
| Explaining a concept a second or third way | Yes | Any chatbot, checked against the textbook |
| Writing a graded essay | Mostly no | Feedback on a draft you've written, if the teacher allows it |
| Take-home tests | No | Assume it's cheating unless told otherwise |
| Teacher lesson prep and quizzes | Yes | School-approved tool, no student personal data pasted in |
If you're a student, the test is simple: after a session with the AI, could you solve a similar problem with the laptop shut? If not, you've been copying, however it felt. Our list of AI tools for studying leans towards the ones that quiz you.
Parents: check which mode your child is using. Khanmigo costs $4 a month or $44 a year for families on its plans page, and the parent account lets you see your child's chat history and get moderation alerts. That visibility matters more than the price, honestly. For homework-solver apps that just hand over answers, like the ones in our Gauth AI review, I'd set clear rules about when they're allowed.
Teachers: Khanmigo's teacher tools are free, and platforms like MagicSchool (see our MagicSchool AI review) are built for district rollouts with admin controls. Start with your own prep work, where the risk is low, before putting AI in front of students.
Next up: bans, pilots and literacy rules
Policy is moving in two directions at once. New York City is imposing a one-year moratorium on student-facing generative AI from pre-K through eighth grade for the 2026-27 school year, covering around 600,000 children, with up to 50,000 high schoolers allowed into five pilot programmes. Norway banned generative AI in grades one to seven earlier, with only supervised use up to age 16.
In the US more broadly, federal executive orders between 2025 and early 2026 set up a national framework and a task force for K-12 AI instruction, and states have been passing laws requiring districts to adopt formal AI policies. In Europe, the AI Act treats AI used to grade, admit or monitor pupils as high risk, but those obligations were pushed back to 2 December 2027.
The AI literacy duty in Article 4 already applies, though.
My guess is that the split will settle along age lines: little or no student-facing AI for young children, and supervised tutor-mode tools for teenagers, paired with lessons on how the tools fail. That seems like the right call to me, though I'd want to see more long-run studies than the handful we have. For the wider picture, see our explainers on AI in education and AI in schools pros and cons.
FAQ
How can AI help students?
AI can act as a patient tutor that explains a topic several ways, quizzes you, and gives instant feedback on practice problems at any hour. It helps most when it's set to give hints instead of answers. In a Harvard physics trial, a carefully designed AI tutor more than doubled median learning gains compared with an active-learning class.
How is AI being used in education?
Teachers use it to draft lesson plans, rubrics and quizzes. Students use chatbots for research, maths help and editing; 54% of US teens in a Pew survey said they'd used them for schoolwork. Schools are also rolling out tutor tools such as Khanmigo.
Why is AI good for education?
It brings one-to-one tutoring to students who couldn't otherwise get it. A World Bank pilot in Nigeria found six weeks of supervised AI tutoring produced gains it equated to nearly two years of typical learning.
Why is AI bad for students?
Used as an answer machine, it lets students skip the thinking that builds skill. In a field experiment with nearly a thousand high school maths students, those who practised with plain ChatGPT-style help scored 17% worse on exams without it than students who never had access. Wrong answers, privacy and cheating are real problems too.
Can AI help students learn?
Yes, probably, when it's built and used as a tutor inside a teacher-run class. The same model without those guardrails boosted homework scores but hurt exam results, so the setup matters more than the model.



