What AI in the classroom actually means in 2026, from a teacher building tools

I build AI tools for teachers, so I have a weird seat for this. I also know how a real school week feels when the marking stack is still sitting there and a vendor deck promises to fix everything by Friday.

The AI-in-education conversation has turned into fog. Every week brings another district partnership, another platform claiming it saves teachers 10 hours, another essay about homework dying. If you're a teacher or parent trying to decide what to trust, the noise is exhausting.

Here is the sorting system I use now: 3 piles, 6 concrete anchors, and a 4-question vendor test. By the end, you should know which claims deserve a second look and which ones can go straight into the recycling bin (digital or otherwise).

The map: overpromised, underdelivered, real

I use 3 piles because "is AI good for schools?" is too mushy to help anyone. Sort the claim by what it asks the school to trust, who gets to override it, and what happens when the output is wrong.

Overpromised: AI as the new infrastructure of school

The big-company pitch in 2026 is simple: AI as the layer underneath every part of school. Anthropic announced in January 2026 a partnership with Teach For All targeting 100,000+ teachers across 63 countries, reaching 1.5 million students. Google, Microsoft, and OpenAI all have similar-scale partnerships in motion. Ben Williamson at the University of Edinburgh has been watching this and calling it what it is: a positioning of AI companies "like a think tank and consultancy, setting out ambitions to be 'core infrastructure' of education." Once something becomes infrastructure, opting out gets expensive.

The companies building these tools are real businesses with sales targets, roadmaps, and investors who like permanent contracts. That matters. A school AI decision made on a 3-year contract can lock in choices that look strange 5 years later. If a vendor pitches the tool as infrastructure, ask harder questions before the paperwork gets warm.

Underdelivered: AI replacing teacher judgment

The Brookings Institution's 2026 global task force report, "A New Direction for Students in an AI World," is the calmest large-scale assessment I've read. It involved 500+ students, teachers, parents, leaders, and technologists across 50 countries, and reviewed 400+ studies. The headline finding is sobering:

At this point in its trajectory, the risks of utilizing generative AI in children's education overshadow its benefits. (Brookings, January 2026)

I read that as a warning about conditions, not a ban. Right now, with the way most schools are using generative AI, the costs are bigger than the gains. Brookings wants schools to change the conditions: prosper, prepare, protect. That frame is built for leaders making policy, not vendors trying to close a contract before lunch.

The RAND Corporation's 2026 update to their school-AI-adoption survey reported that student homework AI use jumped from 48% to 62% between May and December 2025, while school training and policies on AI fell visibly behind. Adoption is happening. Governance is not. That gap is the underdelivered story of 2026: every classroom in the world has students using AI tools that the school never decided to deploy.

Real: narrow, supervised, in the teacher's hand

The useful stuff is narrower and less glamorous. AI works in education when 3 conditions hold: the task is narrow, the AI's output is supervised by a teacher, and the teacher has the authority to override anything the AI suggests. EdTech Hub published a piece in March 2026 surveying 800+ teachers across 29 countries on how AI is being used in low- and middle-income schools, and the finding worth quoting is that 81% of teachers wanted to be involved in designing AI tools, not just using them. A pilot in Pakistan that took that "teacher in the loop" framing seriously hit 98% teacher integration with reported workload reduction. That's a real, narrow, useful win.

The categories I've watched land well in 2026 are practical: AI for grading consistency on a locked rubric, AI for parent-summary drafting, AI for worksheet adaptation, AI for first-pass feedback on practice exercises. Tasks where the teacher reads the output, decides whether to keep it, and the AI is auxiliary. The best tools stay close to the teacher's hand.

What parents actually think

A March 2026 Education Week survey of parents reported that 97% of parents had at least one concern about AI in schools, with an average of three concerns each. Cheating was the top worry, but the second- and third-ranked concerns were biased evaluation of student work and data privacy. 83% of parents wanted written AI policies; 74% wanted human connection preserved in teaching. Parents can live with useful tools. What they hate is mystery software touching their child's work with no policy around it.

The Center on Reinventing Public Education at ASU surveyed what parents actually know about generative AI in their child's school and found, unsurprisingly, very little. Most parents form opinions about classroom AI without basic information from the school about what's actually being used. In 2026, leaders need a plain note home more than another AI strategy PDF.

The four-question vendor checklist

The next time a vendor pitches you their AI tool, in person or in a school-board deck, run it through these 4 questions. If the answer gets slippery, move on. There is too much choice in 2026 to reward fog.

1. What exact task does this tool do, and where are the boundaries? A clear answer here is the most important signal. "It helps with everything" is the marketing version of "we don't really know what we built."
2. What does the teacher review before anything is final? Every honest classroom AI has a teacher-review step. If the tool is positioned as fully autonomous, it's the wrong tool for grading or any artefact that goes on a transcript.
3. What happens to student data? Is it used to train the model? Is it sold? Is it shared with parents, with other teachers, with other schools? The parent survey above shows this is the number 2 concern after cheating; the school that can't answer it cleanly will lose the parents.
4. What's the off-ramp? If the school stops paying, what happens to the data, the rubrics, the per-student records? Vendors who position themselves as "infrastructure" will struggle with this question. Vendors who position themselves as a useful tool will answer it in one sentence.

The first two questions filter out 80% of pitches in my experience. The third filters out the worst of the privacy-laundering startups. The fourth filters out the infrastructure-ambition vendors who want to bolt themselves permanently into the school. What's left is usually pretty good.

What I tell colleagues asking me about AI

This is the four-line version I give when a colleague asks me whether they should use ChatGPT, or Claude, or whatever specific tool just arrived in their inbox.

• Use it for the boring, repetitive parts of your job. Drafting parent emails, summarising student writing, generating practice questions. Low stakes, easy to verify.
• Don't let it produce the score itself unsupervised. Grades go on transcripts. Transcripts shape lives. The bar for "I trust this enough to put it on a transcript" is high, and a general chatbot doesn't clear it.
• Tell parents what you're using and why. The Education Week and CRPE data is unambiguous: parents want clarity. Even a one-paragraph note in your weekly newsletter is worth more than nothing.
• When in doubt, choose the narrower tool. Single-purpose tools are usually built with the teacher's review step front-and-centre. General platforms are usually built to be sticky, not safe.

That's the version I trust. Sort claims into the 3 piles. Read the Brookings report if you're a school leader. Run vendor pitches through the four questions. Tell parents what you're doing. The next time someone tries to sell you AI as the new infrastructure of education, you'll have sharper questions ready, and a good vendor will give you better answers.

Sources cited above: Brookings Institution, "A New Direction for Students in an AI World," January 2026; RAND Corporation, "AI Use in Schools" research report 2026; Education Week, parent survey on AI in schools, March 2026; Anthropic / Teach For All partnership announcement, January 2026; Ben Williamson, University of Edinburgh, public commentary 2026; EdTech Hub research piece, March 2026; Center on Reinventing Public Education research brief, 2026.