The debate about AI in education usually runs on vibes. One side says chatbots will give every child a personal tutor. The other says they will produce a generation that cannot write a paragraph without help. Both camps tend to skip the part where researchers actually tested it.
This guide starts with that evidence. You will get what students and teachers are really doing with AI, what four controlled studies found (including one where AI made results worse), an honest list of pros and cons, and where official policy stands as of October 2026. It ends with a practical checklist for schools and parents. If you want the basics of how these tools work first, read what an LLM is.
What is AI in education?
AI in education is the use of artificial intelligence to support learning, teaching and the running of schools and universities. Today that mostly means generative AI: tools built on large language models that can explain, quiz, draft and give feedback in plain language.
It helps to split it into three uses, because the evidence is different for each:
- Students learning with AI. Chatbots, AI tutors and study modes that explain concepts, give hints or check work.
- Teachers working with AI. Drafting lesson plans, quizzes, rubrics, parent emails and feedback, or getting real-time coaching while tutoring.
- Institutions using AI. Advising, scheduling, admissions support and analyzing which students need help. These uses raise the biggest fairness and privacy questions.
A 2026 OECD report on generative AI in education reaches a conclusion worth keeping in mind for the rest of this guide: according to a European Commission summary of the Digital Education Outlook 2026 (opens in a new tab), how a generative AI tool is designed and used matters more than whether it is used at all.
How students and teachers use AI today
Whatever a school's official policy, AI is already in the classroom. The latest large surveys show how far adoption has gone:
| Group | Finding | Source |
|---|---|---|
| US teens (13 to 17) | 54% have used AI chatbots for help with schoolwork; 10% do all or most of their schoolwork with chatbot help | Pew Research Center, survey Sept to Oct 2025 |
| US teens (13 to 17) | 59% say students at their school use AI to cheat at least somewhat often | Pew Research Center, survey Sept to Oct 2025 |
| UK undergraduates | 95% use AI in some form; 94% use generative AI to support assessed work | HEPI and Kortext, 2026 |
| US K-12 teachers | 60% used AI during the 2024 to 2025 school year; weekly users report saving about 5.9 hours a week | Gallup and Walton Family Foundation, June 2025 |
| US K-12 teachers | Only 19% say their school has an AI policy | Gallup and Walton Family Foundation, June 2025 |
| Lower secondary teachers (TALIS 2024 countries) | 37% used AI in their work in 2024 | OECD TALIS 2024, via the OECD Digital Education Outlook 2026 |
Two patterns stand out. First, students are ahead of their institutions. In the HEPI 2026 survey (opens in a new tab), fewer than half of UK undergraduates felt supported by their teaching staff on AI. Second, use is uneven. Pew's teen survey (opens in a new tab) found that 20% of teens in households earning under $30,000 a year say they do all or most of their schoolwork with chatbot help, compared with 7% in higher-earning households. That gap matters, because the research below shows unguided use is the kind most likely to hurt learning.
On the teacher side, the Gallup and Walton Family Foundation survey (opens in a new tab) found weekly AI users report saving about 5.9 hours a week. The same survey found most teachers were working without a school policy to guide them.
What the research says: four controlled studies
Opinion pieces about AI in schools are everywhere. Randomized studies, where some students get the AI and a comparable group does not, are rarer and far more useful. These four are among the most cited, and together they explain most of what we know.
| Study | Who | What happened | Main caveat |
|---|---|---|---|
| Harvard AI physics tutor (Scientific Reports, 2025) | 194 Harvard undergraduates in an intro physics course | Median learning gains were more than double those from an active-learning class, in less time | One course, one university, introductory material |
| GPT-4 in high school math (PNAS, 2025) | Nearly 1,000 high school students in Turkey | Open ChatGPT-style access raised practice scores 48% but exam scores fell 17% once AI was removed; a guardrailed tutor avoided the drop | One country, one subject |
| AI after-school tutoring in Nigeria (World Bank, 2025) | Students in a six-week after-school program in Nigeria | About 0.3 standard deviations of improvement after six weeks, with teacher support | Early results; teacher-supported program, not a chatbot alone |
| Tutor CoPilot (Stanford, 2025) | 700+ human tutors and 1,000 students from underserved communities | Students were 4 percentage points more likely to master math topics; 9 points for students of lower-rated tutors | AI coached the tutor; it did not teach students directly |
1. A well-designed AI tutor beat a strong classroom
At Harvard, researchers ran a crossover trial in Physical Sciences 2, the university's largest physics course. Each student learned one topic from a custom AI tutor at home and another in an active-learning class, then switched. The study in Scientific Reports (opens in a new tab) found median learning gains with the AI tutor were more than double, and the median student spent 49 minutes on the AI lesson against an assumed 60 minutes in class.
students learn significantly more in less time when using the AI tutor
The details matter. The tutor was built on research-based teaching principles, designed to guide students through problems, and given pre-written step-by-step solutions to work from. The authors also caution that they do not assume AI tutoring will always beat active learning, especially for work that needs complex synthesis and higher-order thinking.
2. An open chatbot helped in practice and hurt on the exam
In Turkey, researchers gave nearly 1,000 high school math students one of three setups: a standard GPT-4 chat interface, a GPT-4 tutor with safeguards that gave hints instead of answers, or no AI. According to the paper in PNAS (opens in a new tab) and Wharton's summary (opens in a new tab), students with the open chatbot scored 48% better during practice, then 17% worse than the no-AI group on an exam taken without it. The guardrailed tutor group scored 127% better in practice and about the same as the control group on the exam.
The authors' explanation is simple: without guardrails, many students used the chatbot as a crutch and skipped the thinking. This is the single most important finding for anyone deciding whether to allow AI homework help.
3. A low-cost program in Nigeria produced large gains
A World Bank team ran a six-week after-school program in Benin City in June and July 2024, with students using a generative AI tutor for English while teachers guided the sessions. The World Bank's write-up (opens in a new tab) reports gains of about 0.3 standard deviations, which the authors equate to nearly two years of typical learning, and says the program outperformed 80% of interventions in a database of randomized trials in developing countries. Each extra day of attendance was linked to larger gains.
4. AI made human tutors better
Tutor CoPilot did not talk to students at all. It suggested responses to human tutors during live math sessions, which the tutors could use, edit or ignore. The Stanford randomized trial (opens in a new tab) found students were 4 percentage points more likely to master topics, rising to 9 points for students whose tutors were rated lower, at an estimated cost of about $20 per tutor per year.
Benefits of AI in education
Based on the evidence above, these are the benefits with the strongest support:
- Tutoring at scale. Human one-to-one tutoring is expensive, so most students never get it. Well-designed AI tutors can offer a version of it to far more students.
- Immediate, patient feedback. A student can ask the same question five ways at 11 p.m. without embarrassment.
- Teacher time back. Weekly AI users in the Gallup survey report saving about 5.9 hours a week.
- Stronger support from less experienced staff. Tutor CoPilot's biggest gains came from helping the weakest tutors.
- Accessibility. Text-to-speech, translation, simplified explanations and reading support can help students with disabilities and English learners take part in the same lesson.
- Better-designed tools are arriving. Major AI companies now ship learning modes that hold back answers: ChatGPT study mode (opens in a new tab) (July 2025), Claude's Learning mode (opens in a new tab) for universities (April 2025) and Gemini's Guided Learning (opens in a new tab) (August 2025). Independent evidence on these specific features is still limited.
Risks and downsides of AI in schools
- Over-reliance and lost skills. The Turkish study is the clearest warning: unguided use raised practice scores and lowered exam scores. The OECD report links heavy chatbot reliance to what it calls metacognitive laziness.
- Cheating and assessment. In Pew's survey, 59% of teens say students at their school use AI to cheat at least somewhat often. Take-home essays and problem sets are now hard to grade as evidence of individual skill.
- Wrong answers stated confidently. Language models can produce plausible errors. Purpose-built tutors such as the Harvard one were given pre-written solutions rather than relying on the model alone.
- Student data privacy. Students may type personal information into tools that store it. UNESCO's guidance calls on governments to mandate data privacy protection.
- Unequal access and unequal use. Better-resourced schools can buy and train for guided tools, while other students may rely on free chatbots with no guardrails.
- High-stakes decisions. AI that scores exams, decides admissions or flags cheating can affect a student's future. The EU treats these uses as high-risk, as covered below.
Should AI be used in schools? The case for and against
This is a values question as well as an evidence question, and reasonable people land in different places. Here is the strongest version of each side.
| Argument for | Argument against |
|---|---|
| Students already use AI, so schools should teach them to use it well rather than pretend it is not there | Schools should protect time for students to build skills without shortcuts, especially in early grades |
| Guided AI tutors have outperformed strong teaching in controlled trials | Those trials used purpose-built tools; many students use general-purpose chatbots |
| AI can narrow gaps by giving tutoring to students who cannot afford it | Heavy reliance on chatbots is more common among lower-income teens, which could widen gaps |
| Teachers save hours each week for higher-value work | Time saved can become pressure to do more with fewer staff |
| AI literacy is a job skill students will need | Long-term effects on attention, writing and reasoning are not yet known |
The official guidance reviewed for this article generally takes a middle position: allow AI where it serves a clear learning purpose, require human oversight, protect student data, and adjust by age. The Manhattan Institute (opens in a new tab) has proposed a grade-by-grade approach (AI for critical thinking in elementary school, AI literacy in middle school, personalized learning in high school), while UNESCO has suggested considering an age limit of 13 for generative AI in the classroom.
AI in education policy: where things stand in 2026
Policy is moving fast. Here is the current picture as of October 2026.
United States
- Federal direction. Executive Order 14277 (opens in a new tab), signed April 23, 2025, set a national policy to promote AI literacy, created a White House Task Force on AI Education and directed agencies to prioritize AI in teacher-training grants.
- Federal funding guidance. In July 2025 the US Department of Education (opens in a new tab) said existing federal grant funds can support AI for high-quality instructional materials, high-impact tutoring and college and career advising, and stressed privacy and parent engagement.
- Edtech evidence standard. On August 20, 2026, the Department issued a Dear Colleague letter on education technology and screen time. As EdSurge reported (opens in a new tab), it asks states and districts to buy tools with evidence of effectiveness, especially from randomized controlled trials, and says edtech should not be judged only by screen time. It creates no new federal rules.
- States. At least 28 states and the District of Columbia had issued K-12 AI guidance by July 2025, according to Stateline (opens in a new tab). In 2026, states began passing laws: ExcelinEd's tracker (opens in a new tab) lists enacted measures including Utah's AI literacy standards for middle school, Maryland's teacher professional development requirement and Oregon's design rules intended to protect minors.
International
- UNESCO. Its guidance on generative AI in education (opens in a new tab) calls on governments to regulate these tools, mandate data privacy protection and consider an age limit of 13. When UNESCO surveyed more than 450 schools and universities (opens in a new tab) in 2023, fewer than 10% had formal guidance.
- European Union. The EU AI Act (opens in a new tab) classifies AI used to decide access or admission, evaluate learning outcomes or monitor students for prohibited behavior during tests as high-risk, which brings stricter obligations.
- OECD. The Digital Education Outlook 2026 urges governments to protect foundational skills and independent thinking, invest in evaluated educational AI tools and set standards for privacy, safety, bias and transparency.
AI in higher education
Universities face the same questions with less supervision and higher stakes for grades. With 94% of UK undergraduates in the HEPI survey using generative AI for assessed work, banning it is not realistic, and HEPI's report puts changing assessment practices at the center. Practical options include in-person exams and oral defenses, marks for process such as drafts and reflections, and explicit rules on when AI is allowed. Tool makers are responding too: Claude for Education launched with Northeastern University, LSE and Champlain College among its first partners, according to Anthropic (opens in a new tab). HEPI also reports growing student anxiety about misconduct accusations, which is a reason to be careful with AI detection tools.
How to use AI in the classroom well: a checklist
For school leaders, teachers and parents, the evidence points to a short list of practical rules:
- Start from a learning goal, not a tool. Ask what students should be able to do without AI at the end, then decide where AI helps them get there.
- Prefer guided modes over open answers. Use tutors or learning modes that ask questions and give hints. Turn off or discourage answer-dumping for practice work.
- Assess without AI where it counts. If a skill matters, check it in a setting where AI is not available, as the Turkish study did.
- Keep a teacher in the loop. The strongest results came from tools that supported teachers or were run inside teacher-led programs.
- Vet privacy before rollout. Know what student data a tool collects, where it is stored and whether it trains models. Follow your district or national rules.
- Teach AI literacy directly. Students should learn how these models work, why they make mistakes and how to check them. Our guide to writing better prompts is a useful starting point for older students.
- Write the policy down. With only 19% of US teachers reporting a school AI policy, a clear one-page rule on allowed and banned uses removes a lot of guesswork.
- Review with evidence. Check whether the tool improves results you care about, and drop it if it does not. That is also what the Department of Education's 2026 letter asks.
Here is a hypothetical example of the difference. A student stuck on a quadratic equation asks an open chatbot and gets the worked answer to copy. The same student using a study mode is asked what they have tried, gets a hint about factoring and has to finish the problem. Both feel helpful in the moment. Only the second builds the skill the test will check.
The bottom line
AI in education is no longer optional in practice: most students already use it. The research says the outcome depends on design. Tools that guide students to think, or that help teachers teach, have produced some of the largest gains seen in education research. Tools that simply hand over answers can leave students worse off once the AI is gone. Schools that set clear goals, choose guided tools, test skills without AI and protect student data will get the benefits with fewer of the risks.
To understand the technology behind these tools, read our explainers on large language models and RAG, a technique education tools can use to ground answers in course material. For more guides, browse the AI hub.