AI in Education (2026): What Works, What the Evidence Shows, and What Schools Should Do
AI in education means using artificial intelligence, mostly chatbots and tutoring systems built on large language models, to help students learn and teachers teach. By 2026 it is part of daily school life: a Pew Research survey published in February 2026 found that 54% of US teens had used an AI chatbot for schoolwork. In the classroom, RAND reported that 53% of English, maths, and science teachers and 54% of their students used AI for school in the 2024–25 academic year.
The core finding from the evidence is straightforward: AI that tutors helps, but AI that does the work can hurt learning. This distinction matters. When AI asks questions, gives hints, and keeps the student thinking, it tends to improve outcomes. When it simply generates the answer, it may improve homework scores but fail to build lasting understanding. The OECD Digital Education Outlook 2026, released in January, confirmed this pattern. Students using general-purpose GenAI produced better outputs, but “this advantage disappears – and sometimes reverses – in exams when access is removed.” Purpose-built pedagogical tools, by contrast, showed sustained gains.
This guide grades the evidence behind these claims. We look at randomized trials, lab studies, and large-scale surveys. We separate company claims from independent results. We examine who is using AI, how they are using it, and what the data says about cheating, equity, and policy. It ends with a practical checklist for schools and universities.
How AI is used in education
AI is not a single tool. It is a set of applications that serve different roles in the educational ecosystem. Understanding these roles helps explain why some uses improve learning while others do not.
- AI tutors for students: Systems that guide students through practice problems, ask Socratic questions, and provide hints without giving the final answer. Examples include Khan Academy’s Khanmigo and purpose-built tutors in university courses.
- Homework help with general chatbots: Students using ChatGPT, Gemini, or Claude to explain concepts, check grammar, or generate drafts. This is the most common form of AI use among students today.
- Teacher planning and admin: Educators using AI to draft lesson plans, create rubrics, write emails, or summarize student work. A randomized trial by the UK Education Endowment Foundation found teachers spent 31% less time planning with AI guidance.
- Feedback and grading support: AI tools that provide initial comments on essays or code, allowing teachers to focus on higher-level feedback. This is often used as a first pass, not a final grade.
- Coaching for human tutors: AI tools that assist human tutors by suggesting strategies or tracking student progress. The “Tutor CoPilot” study showed students were 4 percentage points more likely to master topics with this support.
- Accessibility and language support: Tools that translate text, simplify language, or provide text-to-speech for students with disabilities or those learning in a second language.
- School administration: AI used for scheduling, resource allocation, or analyzing enrollment data to identify at-risk students.
Beyond these general categories, major tech companies have released edu-specific versions of their chatbots. OpenAI launched “study mode” in ChatGPT in July 2025, which provides step-by-step Socratic guidance rather than direct answers. Anthropic introduced “learning mode” in Claude in April 2025, designed to engage students in dialogue. Google released “Guided Learning” in Gemini in August 2025. These products aim to reduce the “do the work” problem by forcing interaction. However, they remain commercial products with varying degrees of independent validation.
What the evidence says: tutoring vs doing the work
The central tension in AI education is between augmentation and automation. Augmentation means AI supports the student’s thinking. Automation means AI replaces it. The evidence suggests that augmentation leads to better learning, while automation leads to better homework but potentially worse exam performance.
The OECD Digital Education Outlook 2026 (January 2026) provides a clear summary of this dynamic. It notes that students using general-purpose GenAI produce better outputs, but “this advantage disappears – and sometimes reverses – in exams when access is removed.” Purpose-built tools, designed with pedagogical guardrails, show sustained gains. This distinction is critical for educators who confuse output quality with learning depth.
Trials where AI tutors helped
The strongest evidence for AI’s benefits comes from randomized trials where the AI acted as a tutor.
A study published in Scientific Reports in June 2025 by Kestin et al. from Harvard examined physics students. In a randomized crossover trial involving 194 students, those who used a purpose-built AI tutor learned “significantly more in less time” than those in an active-learning class. The median learning gain was over double, and the median time spent was 49 minutes versus 60 minutes for the control group. However, the study had limitations: it covered only one course and two topics, the authors built the tutor themselves, and they used an immediate post-test only. Still, the effect size is notable.
In developing countries, the results are equally promising. A World Bank trial in Edo State, Nigeria, ran from June to July 2024. It used a Microsoft Copilot-based tool based on GPT-4 in after-school settings over six weeks. The study found an effect size of about 0.3 standard deviations overall, with an effect of 0.23 in the main English outcome. The World Bank states this outperformed 80% of education trials in developing countries and equals 1.5–2 years of normal schooling. The effects were larger for girls. It shows that AI tutoring can work in schools with few resources, not only at elite universities.
Another randomized trial, preregistered and revised in November 2025, studied “Tutor CoPilot” from Stanford. This tool coaches human tutors rather than replacing them. Students were 4 percentage points more likely to master topics when their human tutors used the AI tool. Students of lower-rated tutors gained 9 points. The cost was about $20 per tutor per year. This suggests AI can amplify the impact of human educators, particularly those who are less experienced.
Contractor & Reyes (preprint 2026) conducted a proctored randomized trial with undergraduates. They found that gains held for students who used AI as a tutor (“augmentation”) but faded when AI wrote for them (“automation”). This reinforces the idea that the mechanism of use determines the outcome.
Trials where AI hurt or did little
Not all trials show success. Some reveal that AI can harm learning if not carefully managed.
A field experiment published in PNAS in June 2025 by Bastani et al. studied about 1,000 high-school maths students in Türkiye. When students used plain GPT-4 (“GPT Base”) for practice, their practice grades rose 48%. However, on the unassisted exam, those students scored 17% lower than the control group. This is a clear case of AI doing the work: students got better grades on practice but performed worse on actual exams. When the researchers added tutoring guardrails (“GPT Tutor”), practice grades rose 127% and the harm was largely removed (exam scores were about equal to the control group). The title of the paper, “Generative AI without guardrails can harm learning,” captures the risk.
In the US, a cluster randomized trial over two years in 18 Tennessee middle schools tested Khanmigo. The working paper by Oreopoulos & Low, released in August 2026, found an effect size of about 0.06–0.08 standard deviations per year. This is similar to Khan Academy practice without AI. However, 96% of students tried the tool, but they used it in only 17% of exercise sessions where they had made an error. The authors conclude that “the binding constraint appears to be engagement.” If students don’t use it, it doesn’t help.
A randomized trial of 2,379 students and 30 instructors at a US public university (Liu et al., working paper October 2026) found that access to a course-integrated generative AI tutor “reduced final grades by 0.37 standard deviations.” The losses were larger for first-generation students. If that pattern holds, poorly designed AI tutors could widen gaps between students rather than close them.
The “Your Brain on ChatGPT” study and its limits
In June 2025, the MIT Media Lab published a preprint study titled “Your Brain on ChatGPT.” It involved 54 participants who wrote essays using an LLM, a search engine, or nothing. EEG recordings showed lower brain engagement with the LLM, and LLM users struggled to quote their own essays later. This suggests that AI use may reduce cognitive processing. However, the study has significant limitations: it is a small sample, it only tested essay writing, and it was a lab setting. It does not show brain damage, nor does it prove that AI use is bad for all types of learning. It is a signal worth monitoring, not a verdict.
| Study | Type | Setting | Result |
|---|---|---|---|
| Kestin et al. (Harvard) | Randomized crossover | 1 course, 194 students | Significant learning gains in less time; immediate post-test only. |
| World Bank (Nigeria) | Randomized trial | After-school, 6 weeks | 0.3 SD effect; outperforms 80% of dev-country trials; larger for girls. |
| Tutor CoPilot (Stanford) | Randomized | Human tutors coached by AI | +4 percentage points topic mastery; +9 for students of lower-rated tutors; about $20 per tutor a year. |
| Bastani et al. (PNAS) | Field experiment | ~1,000 students in Türkiye | Plain AI: +48% practice, -17% exam. Guardrailed AI: +127% practice, neutral exam. |
| Khanmigo (Oreopoulos & Low) | Cluster randomized | 18 TN middle schools | 0.06–0.08 SD effect; low engagement (17% of error sessions). |
| Liu et al. (US University) | Randomized | 2,379 students, 30 instructors | -0.37 SD final grades; larger losses for first-gen students. |
| Contractor & Reyes | Proctored randomized | Undergraduates | Gains held for “augmentation”; faded for “automation”. |
| MIT Media Lab | Lab study (preprint) | 54 participants | Lower EEG engagement with LLM; difficulty quoting own essays. |
How many students and teachers use AI
Adoption is widespread, but the numbers vary by region and demographic. It is important to note that these surveys are not directly comparable due to differences in question wording, timing, and sample populations.
In the US, Pew Research data shows a rapid rise in teen usage. In 2023, 13% of teens used ChatGPT for schoolwork. By late 2024, that number doubled to 26%. A 2025 survey published in February 2026 found that 54% of teens used any AI chatbot for schoolwork. One in ten do all or most of their schoolwork with chatbot help. Notably, 59% of teens think AI cheating is a regular occurrence at their school.
RAND reports that 25% of US teachers used AI in 2023–24. In 2024–25, this rose to 53% of English, maths, and science teachers and 54% of students. Students using AI for homework rose from 48% in May 2025 to 62% in December 2025. Concerns about critical thinking are also rising: the share saying AI harms critical thinking rose from 54% to 67% in the same period.
Gallup’s 2025 survey of 2,232 US public school teachers found that six in ten teachers used AI for work. Weekly users estimate they save 5.9 hours a week, though this is self-reported data.
Across the countries in the OECD’s TALIS 2024 survey, 37% of lower-secondary teachers used AI and 72% were concerned about academic integrity. In the UK, HEPI reports that 92% of UK undergraduates used AI in 2025, rising to 95% in 2026. In 2026, 94% use generative AI to support assessed work.
Anthropic’s April 2025 company report analyzed 574,740 student conversations with Claude. It found that nearly half (~47%) were “direct” conversations seeking answers with little back-and-forth. That resembles the “automation” pattern that the trials above link to weaker learning.
| Survey | Who | Finding |
|---|---|---|
| Pew Research | US teens | 54% used AI chatbots for schoolwork (2025 survey); 10% do all/most work with AI; 59% think cheating is regular. |
| RAND | US teachers & students | 53% of teachers and 54% of students used AI (2024–25); 62% of students used AI for homework (Dec 2025); 67% say it harms critical thinking. |
| Gallup–Walton | US teachers | 6 in 10 used AI; weekly users save 5.9 hours/week (self-reported). |
| OECD TALIS | Lower-secondary teachers | 37% used AI; 72% concerned about academic integrity. |
| HEPI | UK undergraduates | 95% used AI in 2026; 94% use it for assessed work. |
| Anthropic | Students (Claude users) | 47% of conversations were “direct” (seeking answers with little back-and-forth). |
Company and national programs
Adoption is also driven by company and national programs (the figures below are the organizations’ own). Khan Academy’s Khanmigo had about 2 million users in the 2024–25 school year, free for US teachers through Microsoft since May 2024. OpenAI launched ChatGPT Edu for universities in May 2024. Anthropic’s Claude for Education was adopted campus-wide by Northeastern University, LSE, and Champlain College in April 2025. Google released Gemini “Guided Learning” in August 2025. The AFT teachers’ union launched a National Academy for AI Instruction in July 2025, funded by Microsoft, OpenAI, and Anthropic, to train 400,000 educators over five years.
Internationally, Estonia’s “AI Leap” program started in September 2025, providing AI tools for 20,000 upper-secondary students and 3,000 teachers in phase one. Greece agreed with OpenAI in September 2025 to pilot ChatGPT Edu in secondary schools. El Salvador announced Grok for 1 million+ students with xAI in December 2025, though there is no independent evaluation yet.
AI for teachers: time saved and how
The promise of AI in education for teachers is not that it will replace them, but that it might give them their time back. The evidence suggests it can, but the savings are real only if teachers use the tools as intended.
The strongest signal comes from a randomized trial conducted by the UK Education Endowment Foundation (EEF). Published in December 2024, the study involved 259 science teachers across 68 schools. When teachers used ChatGPT alongside a practical guide, they spent 31% less time planning lessons compared to the control group. The average time dropped from 81 minutes to 56 minutes a week. Crucially, blind experts did not rate the quality of the lessons as lower. This trial suggests that AI is most effective when it handles the heavy lifting of structure and drafting, leaving the teacher to apply pedagogical judgment.
Self-reported data backs up the efficiency gains. A Gallup–Walton Family Foundation survey of 2,232 US public school teachers in spring 2025 found that six in ten teachers used AI for their work. Weekly users estimated they saved 5.9 hours per week. While self-reported time savings can be inflated, the consistency with the EEF trial indicates that administrative and planning tasks are indeed being compressed.
Typical tasks where teachers report saving time include lesson planning, creating differentiated materials for diverse learners, and drafting feedback on student work. AI is also useful for routine administrative duties, such as summarizing meeting notes or generating email drafts to parents.
However, the time savings come with a caveat. The RAND Corporation highlights an equity gap in teacher training. By fall 2024, 67% of low-poverty districts had trained their teachers on AI, compared to only 39% of high-poverty districts. Teachers who get no training are less likely to know how to prompt well or check AI output, so the time savings reported in surveys may not reach the schools that need them most.
What should stay human? The EEF trial and other studies suggest that the “human in the loop” is essential for quality. AI can generate a lesson plan, but it cannot read the room. It cannot adjust an explanation based on a student’s facial expression or prior knowledge in real-time. Similarly, while AI can draft feedback, the personalization and encouragement that drive student motivation remain a human strength. Over-reliance on AI for grading or feedback can lead to generic comments that fail to address specific student needs.
In short, AI in education for teachers works best when used as a co-pilot for preparation, not as an autopilot for instruction. The 5.9 hours saved by Gallup users are real, but they depend on training and intentional use. Without adequate training, particularly in high-poverty schools, the time savings may not materialize, and the quality of instruction could suffer.
Cheating, AI detectors and assessment
The question of cheating has dominated the conversation around AI in schools, but the evidence suggests that traditional detection methods are flawed. The scale of AI use in assessed work is significant. According to HEPI, 94% of UK undergraduates used generative AI to support assessed work in 2026. In the US, Pew Research reported in February 2026 that 59% of teens think AI cheating is a regular occurrence at their school. Turnitin, a leading detection service, reported in April 2024 that about 11% of over 200 million papers had at least 20% AI-written text, and 3% had at least 80%. These are detector flags, not proven misconduct.
The problem with AI detectors is their reliability. A study by Liang et al., published in Patterns in 2023, found that seven different AI detectors flagged 61% of essays by non-native English writers as AI-generated on average. The same detectors were near-perfect on essays by US students. This high false-positive rate for non-native speakers makes detectors risky for high-stakes decisions like discipline or grading. OpenAI, the company that built ChatGPT, withdrew its own AI text classifier in July 2023 because of its “low rate of accuracy.” It caught only 26% of AI-generated text and wrongly flagged 9% of human-written text.
So, what should schools do instead of relying on detectors? The evidence points to a shift in assessment design. Instead of trying to catch AI, schools should redesign tasks to make AI use transparent or irrelevant.
One effective approach is in-class writing. If students write their essays or solve problems in class, under supervision, the use of AI is minimized. Oral checks, such as asking students to explain their reasoning after submitting written work, can verify understanding. An oral check confirms that the student understands the work, even if AI helped with a first draft.
Another approach is to allow AI use but require disclosure. Students can use AI to brainstorm or draft, but they must submit a process folder showing their drafts, the prompts they used, and how they edited the AI output. This shifts the focus from the final product to the learning process. It also teaches students to use AI as a tool for augmentation rather than automation.
The OECD Digital Education Outlook 2026 (January 2026) notes that while students using general-purpose GenAI produce better outputs, “this advantage disappears – and sometimes reverses – in exams when access is removed.” This suggests that assessments should measure deeper understanding, not just the ability to produce a polished text. Purpose-built pedagogical tools show sustained gains, implying that assessments should align with the learning process, not just the output.
In summary, the era of the high-stakes, take-home essay as the sole measure of learning is ending. Schools should move towards assessments that value process, oral verification, and in-class performance. AI detectors can be used as a flag for further review, but not as the final judge of cheating.
Rules and policy in 2026
As AI becomes embedded in schools, policy is catching up. The regulatory landscape is fragmented but increasingly structured.
United States
On April 23, 2025, President Trump signed Executive Order 14277, “Advancing Artificial Intelligence Education for American Youth.” This created a White House Task Force on AI Education and a Presidential AI Challenge. In July 2025, the Department of Education clarified that federal grant funds could support AI for instructional materials, tutoring, and advising, provided they comply with the Family Educational Rights and Privacy Act (FERPA). By mid-2026, more than 35 US states had issued K-12 AI guidance. New York City, which blocked ChatGPT on school devices in January 2023, reversed the ban in May 2023, adopting a more nuanced approach.
European Union
The EU AI Act, amended by the 2026 Digital Omnibus, classifies AI used in education for admissions, grading, steering learning, or detecting cheating in tests as “high-risk.” These obligations apply from December 2, 2027. AI-literacy duties have applied since February 2025. From then on, the providers of such systems, and to a lesser degree the schools that use them, face duties such as risk management, documentation and human oversight.
UNESCO and China
UNESCO’s guidance on generative AI in education (September 2023) suggested an age limit of 13 for classroom use. UNESCO published AI competency frameworks for students and teachers in 2024. In China, the Ministry of Education issued guidelines in May 2025: primary pupils may not independently use open-ended AI generation, and teachers may not use AI to replace core teaching. Beijing requires at least 8 hours of AI education a year in schools from autumn 2025.
Children’s safety and privacy
California’s SB 243, in force January 2026, requires companion chatbots to tell minors they are talking to an AI and to follow suicide-prevention protocols. The FTC ordered seven companies, including OpenAI, Meta, and Character.AI, to report on chatbot effects on children in September 2025. Character.AI barred under-18s from open-ended chat in November 2025. The updated US children’s privacy rule (COPPA) required compliance from April 2026.
Note: This is not legal advice. Schools should consult legal counsel for specific compliance.
A practical checklist for schools and universities
-
Write a clear AI policy by task. Define when AI is allowed, when it is required, and when it is prohibited. Specify if students must disclose AI use. Ambiguity leads to inconsistency.
-
Prefer tutoring-mode tools that ask rather than tell. The trials above favor tools that give hints and questions over tools that hand over answers. Even then, check that students actually use them: in the Khanmigo trial, engagement was the limiting factor.
-
Redesign assessment. Move towards in-class writing, oral defenses, and process-based assessments. Allow AI in take-home tasks but require disclosure and reflection on how AI was used.
-
Don’t rely on AI detectors for discipline. Detectors have high false-positive rates, especially for non-native speakers. Use them as a flag for review, not as proof of cheating.
-
Train teachers, including in high-poverty schools. The RAND Corporation found a significant training gap between low- and high-poverty districts. Invest in professional development that focuses on prompt engineering and evaluating AI output.
-
Protect student data. Ensure that AI tools comply with FERPA, COPPA, and local privacy laws. Review data privacy policies of AI vendors carefully.
-
Respect age limits. Follow UNESCO’s suggestion of 13 for classroom use, or adhere to local regulations. Be cautious with open-ended chatbots for younger children.
-
Measure learning, not usage. Focus on whether students are learning, not just how much AI they use. Use metrics like exam performance and engagement, not just AI usage rates.
-
Involve parents. Communicate your AI policy clearly to parents. Explain how AI is being used in the classroom and why. Transparency builds trust.
FAQ
How is AI used in education?
AI is used for lesson planning, differentiation, feedback drafts, and tutoring. The UK Education Endowment Foundation found that teachers using AI saved 31% time on planning. Pew Research found that 54% of US teens used an AI chatbot for schoolwork in 2025.
Does AI help students learn?
It can, if it works as a tutor. In a Harvard randomized trial, students using a purpose-built AI tutor learned more in less time than in an active-learning class, and a World Bank trial in Nigeria found gains of about 0.3 standard deviations in six weeks. However, if AI does the work, learning may suffer, as seen in the Bastani et al. study where plain GPT-4 lowered exam scores by 17%.
Is using ChatGPT for homework cheating?
It depends on the policy. If the task is to produce a final product, using ChatGPT may be cheating. If the task is to learn, using AI as a tutor is acceptable. The OECD notes that AI advantage disappears in exams when access is removed.
Can teachers detect AI-written work?
Not reliably. Liang et al. found detectors flagged 61% of non-native writers as AI. OpenAI withdrew its classifier due to low accuracy (26% catch rate). Detectors are flags, not proof.
Will AI replace teachers?
No. AI is a tool. The EEF trial showed teachers still needed to apply judgment. AI handles drafting, but teachers handle pedagogy. The human element in motivation and real-time adjustment remains crucial.
What is an AI tutor?
An AI tutor is a tool that asks questions and gives hints, keeping the student thinking. Examples include Khanmigo and the study or learning modes of ChatGPT, Claude and Gemini. Purpose-built tutors have shown gains in several trials, but results vary: a two-year Khanmigo trial found only a small effect because students rarely asked it for help.
Is AI in schools regulated?
Yes. The EU AI Act classifies some educational AI as high-risk, with rules starting December 2, 2027. In the US, FERPA and COPPA apply. California’s SB 243 requires chatbots to identify themselves to minors.
What age should students start using AI?
UNESCO suggests 13 for classroom use. China’s guidelines bar primary pupils from using open-ended AI generation on their own. Schools should follow local regulations and consider maturity levels.