Stanford HAI published the 2026 AI Index Report this year. It runs to over 400 pages, and Chapter 7 is the one about education.

I've read it so you don't have to. Here's what's worth knowing if you're responsible for AI in a school or trust.

The headline number

Only 6% of US teachers say their school or district has very clear, comprehensive policies or guidance on generative AI and academic integrity.

That figure comes from a RAND report (Doss et al., 2025) that the AI Index draws on. Go back to RAND and it gets starker. It's 3% for elementary teachers, 6% for middle school and 9% for high school.

Teachers of younger pupils were less likely to report clear, comprehensive guidance. That doesn't tell us why, but it does raise a question about where support is reaching.

The AI Index also reports that only about half of US middle and high schools have an AI policy. It gives figures of 28% permitting AI in some circumstances and 22% not permitting it. These add up to about half of schools; they aren't percentages of just the schools with policies.

The students are already there

Estimates of how many US secondary students use AI for schoolwork range from 50% to 84%, depending on which survey you read. The surveys use different questions and samples, so these aren't interchangeable measures.

They mostly use it for research, editing essays and brainstorming.

Here's the figure I'd put in front of any senior leadership team. 47% of students have wanted to use AI for schoolwork but weren't sure if it was allowed. Only 36% described their school's policy as extremely clear.

Students also have a view on what should happen next. In one survey, 65% of middle school and 73% of high school students said they should have access to AI tools and be able to use them for schoolwork. In another, 52% said schools should be required to teach them how.

When pupils don't understand the policy, they're left guessing.

A UK comparison worth making

One useful UK figure in the chapter comes from higher education. In HEPI's 2025 survey, 80% of UK university students agreed their university had a clear policy on generative AI in assessments, up 16 percentage points in a year.

That suggests universities made progress on policy clarity. It doesn't tell us whether every student knows how to apply the rules.

And it isn't a direct comparison with RAND's 6%. HEPI asked students whether they agreed a policy was clear. RAND asked teachers whether guidance on academic integrity was very clear and comprehensive. Different people, different questions and different settings.

I'd love to see equivalent evidence for UK schools and trusts. If you've got data on this, I want to hear from you.

Policy and resources

The report found schools with AI policies are more likely to be in wealthier, more urban communities. That's especially true of policies that allow AI use.

The computer science data shows another gap. In 2025, 91% of large US high schools offered foundational computer science. For small high schools it was 44%.

If clear AI guidance becomes another thing that depends on school size and postcode, trusts have a job to do. A trust-wide policy is one way to reach the small school as easily as the big one, provided staff have the support to put it into practice.

Three terms we keep blurring

This is the most useful page in the chapter, and it's a definitions page.

The report says public debate mixes up three different things:

  • AI in education is using AI tools to teach and learn.
  • AI literacy is understanding how AI works, how to use it and what the risks are.
  • AI education is AI literacy plus the technical skills to build AI systems.

Does your policy mainly cover tool use? What does it say about helping pupils understand AI and its risks? Ask your leadership team which of these your strategy covers. The answers will tell you a lot.

What students actually use it for

Anthropic analysed university students' education-related conversations with its Claude model. It classified most use under what Bloom's taxonomy calls higher-order skills: creating (39.8%) and analysing (30.2%). Applying (10.9%) and understanding (10.0%) came well behind.

Those categories describe the tasks in the conversations. They don't prove students developed those skills, or lost them.

The report reads the pattern as a possible warning. Students may be handing over the thinking they're meant to be developing. Another survey found 55% of US college students thought generative AI had a mixed effect on their critical thinking skills.

Bear in mind Anthropic makes the tool it's reporting on. Even so, the question for teachers is useful: which part of the thinking is the learner doing? Using AI to do the hard part can become the escalator. Learning still needs the stairs.

What other countries tried

The global section is short but useful, because it shows what happened when education systems acted.

The chapter describes South Korea launching AI textbooks in primary schools in March 2025, then reversing course a few months later after parent and teacher pushback.

It also describes Estonia's AI Leap programme piloting AI learning applications with 20,000 students and 3,000 teachers during 2025 to 2026.

In China, Beijing, Guangdong and Hangzhou began requiring AI education in the 2025 to 2026 school year. The chapter describes progression from AI literacy in primary to designing AI systems in secondary. The UAE also mandated AI education across grade levels from that school year.

South Korea is the example I'd study before a large rollout. The reversal raises questions about trust and readiness. I'd want to understand the objections before drawing a lesson for a UK trust.

Skills are growing outside the classroom

LinkedIn data in the chapter shows growth in the AI skills people list on their profiles. AI literacy skills are growing faster than engineering skills in most countries.

In the US, the fastest-growing AI literacy skills listed were AI prompting and Microsoft Copilot Studio.

That second one caught my eye. Copilot Studio is a tool for building agents. Its appearance on the list is a reminder that workplace AI skills already extend beyond using a chatbot.

Profile data doesn't tell us how people learned those skills, or independently verify their level of competence.

The report's own health warning

Credit to the authors, they're upfront about the gaps. They say AI's role in education is expanding faster than the data to track it. A lot of the figures are limited, out of date, or both. The global OECD data used here only runs to 2023, and there's little standardised data on AI teaching in schools.

Most of the school evidence discussed here is also American. Treat it as a prompt to examine your own setting, and check its relevance before using it in a UK policy paper.

What I'd do with this

If you lead AI in a school or trust, I'd ask three questions this half term.

Would your teachers describe your AI guidance as very clear and comprehensive? Would your pupils know what's allowed without having to guess? And does your policy work as well in your smallest school as your largest?

These are the questions we work through in the AI Leadership units I deliver at Primary Goal.

Where does your school or trust sit? What would your staff say if you asked them?

Sources