The Teacher's AI Toolkit in 2026: What Actually Works and What's Still a Demo

September 24, 2026

aiedtecheducationgeminicopilotpolicy

Last school year the question was whether AI belonged in a classroom at all. This year it's which platform already has it and who signs off on the policy. That's a different conversation.

AI stopped being a separate service teachers open on purpose. It moved inside the tools schools already run: Google Classroom, Microsoft 365, and in Ukraine the state-run Mriia ecosystem. At roughly the same moment Ukraine passed a law that makes dishonest use of AI its own category of academic misconduct, alongside plagiarism.

The numbers first, because without them this turns into forecasting

Gallup published results in May 2026 from a survey of 2,069 US public school teachers. Six in ten use AI for work, three in ten use it weekly, saving an estimated 5.9 hours per week.

Now the second number from the same study: 18% receive any formal written guidance from administrators. Roughly a third get no guidance of any kind, formal or informal. Adoption ran about two years ahead of governance.

Where schools encourage AI is worth reading closely. Most often it's adapting materials to individual student needs (58%), making worksheets and assignments (54%), and preparing to teach (53%). Least often it's one-on-one instruction and tutoring (35%). The logic holds up: lesson prep happens at the teacher's desk, where a bad AI output costs five minutes of editing. Direct interaction with a child carries a different price.

England looks similar in shape. Per the DfE's Technology in Schools Survey, 44% of teachers used generative AI for school activities, 35% for lesson planning, 5% for marking. Planning yes. Assessment almost never.

Ukraine's picture differs because the driver is a state platform rather than a vendor. As of 20 July 2026, Mriia had over 4,300 connected schools and roughly 640,000 users.

Four platforms, and the teacher isn't the one choosing

This is the year's real shift. A year ago a teacher picked a service and either paid for it or lived with the free tier. Now the baseline assistant arrives bundled with the school licence, and the IT administrator picks it.

Gemini in Google Classroom. More than 30 educator features, from brainstorming to personalising content. Since April 2026 it's available in every language Classroom supports and across all Workspace for Education editions, including the free Fundamentals tier. From 10 August 2026 the Gemini tab opened to K-12 students of any age where an admin has granted access: they can turn class materials into flashcards or practice quizzes.

Microsoft Copilot, the Teach module. Lesson plans, quizzes, rubrics, plus modifications to language, reading level, difficulty and standards alignment. No paid Copilot licence required — a Microsoft 365 Education licence is enough. Over the summer Microsoft added standards-aligned unit plans covering more than 50 countries: supply subject, grade, language, duration and context, and Teach drafts an overview, essential questions and weekly detail.

Worth being honest about quality here. Med Kharbach, who tested Teach systematically, notes the plans skew toward descriptive, expository formats — students receive information rather than investigate or create. If your teaching leans inquiry-based or project-based, expect to rework most of the output. That's less a verdict on the module than a description of what AI generation looks like when no pedagogical intent goes in.

ChatGPT for Teachers and Claude for Teachers. Both free, both limited to verified US K-12 educators. Anthropic launched Claude for Teachers on 14 July 2026 with a Learning Commons connector mapped to academic standards in all 50 states. Outside the US these are market signals, not options.

The practical takeaway is blunt: if your school runs Google, your assistant is Gemini; if it runs Microsoft, it's Copilot Teach. Both are free within the education licence. There is no third zero-budget option right now.

Mriia and its first AI feature

In February 2026 Ukraine's Ministry of Education shipped the state ecosystem's first AI capability: a test generator for teachers. An AI timetable generator for administrators and full homework workflows inside the app were announced for autumn 2026.

Announced is the operative word. Actual availability in a given school should be checked after the updates ship, not from the press release. And an individual teacher can't run a gradebook independently of their institution — the education features work for staff at connected schools.

Mriia's advantage isn't technological. Ukrainian language, alignment with the national curriculum reform, no cost, and state guarantees around student data. No global vendor offers that combination here.

What changed in how the tools are used

Three things that weren't standard a year ago.

First, grounded AI. Instead of "write me a quiz on this topic," the teacher hands the model their own materials. Gemini Notebook lets students sync teacher-provided materials and build interactive study guides from them. Copilot Notebooks do the same at no extra cost for Microsoft 365 Education. The assistant works from the syllabus and the textbook rather than the open web, and hallucinations drop noticeably.

Second, the teacher controls how much AI students get. Microsoft's Student AI Guidelines let teachers set the permitted level per assignment, from no AI at all to open Copilot Chat, with Learning Groups for differentiated expectations. That's a move from a school-wide allow/deny policy to an instruction at the level of a single task.

Third, reusable assistants instead of rewriting prompts. ChatGPT has Custom GPTs, Gemini has Gems, Claude has Projects. Same pattern each time: take the task you repeat most, load your own materials into it, freeze it as a tool. For most teachers that's unit planning, tiered exercises and feedback drafts.

The year's most important caveat comes from the OECD

The Digital Education Outlook 2026 put into words the thing that usually gets lost in classroom AI debates.

When AI is used without pedagogical guidance, outsourcing tasks to it simply improves performance with no real learning gains. The student submits better work and learns nothing. The report also identifies the mechanism: heavy reliance on generative AI reduces metacognitive engagement, the ability to monitor your own understanding.

That leads to a fairly demanding definition of the teacher's role. In the OECD's framing, teachers aren't managers who permit or prohibit AI; they're professionals who recognise when a student's pattern of AI use starts undermining learning and adjust it. The instruments for that are not technical — have students generate questions, explain their reasoning, reflect on the process.

The report also draws a useful line on creative work: fast uses of AI that prioritise immediate output versus slower uses that support iteration and reflection. The second serves learning. The first serves submission.

The legal frame: no longer a recommendation

Ukraine's Law No. 4742-IX "On Academic Integrity" took effect on 1 February 2026 and came into force on 31 July. It's the country's first law regulating AI use in academic work at the statutory level, and one of the first in Europe.

The law defines 13 categories of violation. Dishonest use of AI-generated results is listed as its own category, alongside plagiarism, fabrication, falsification and non-independent completion of assignments.

For a school that means concrete work: an internal academic integrity policy with AI rules, a commission, a classification of violations and an appeals procedure.

The Ministry's instructional guidance for the 2026/2027 school year frames it more softly but points the same way: AI is a supporting tool for teacher and student, not a replacement for thinking, professional pedagogical judgement or independent work.

One detail in that guidance deserves wider reading. The Ministry states plainly that the format, length, structure and presentation of calendar-thematic and lesson plans is the individual teacher's decision, and that imposing uniform mandatory standards at school, city, district or regional level has no basis in law. In practice that legalises AI-assisted planning in whatever form the teacher chooses.

For European context: the AI Act's high-risk obligations for education systems — admissions, evaluating learning outcomes, assigning students to levels, exam monitoring — moved from 2 August 2026 to 2 December 2027 under the Digital Omnibus. The Article 50 transparency duty, disclosing that a person is interacting with an AI system, applied from 2 August 2026 with no deferral at all.

A minimum stack, and what to do with it this term

Baseline assistant: Gemini or Copilot Teach, depending on what the school runs. Source-grounded work: NotebookLM or Copilot Notebooks, because that's where the model stops inventing. Two or three custom Gems or GPTs for the tasks that repeat weekly. Canva or Google Vids for visuals. For Ukrainian schools, Mriia for the gradebook and quizzes in Ukrainian.

And a school policy. Without one, the rest of the stack is personal risk carried by each teacher individually.

Four things worth doing this term: check what AI is actually enabled in your school licence (often more than staff realise); build one assistant for your most repeated task; define the permitted AI level for at least three assignment types; read your school's academic integrity policy, and if there isn't one, that's no longer a next-year problem.

Separately: Google is running a free training programme for educators covering Gemini, Gems, NotebookLM, Gemini in Classroom and Google Vids, with the first sessions from 28 September 2026.

AI detectors are deliberately absent from that list. They're unreliable, they produce false positives, and they can't be the sole basis for accusing a student. An assignment designed to resist AI works better than a detector: oral explanation, drafts, defending the reasoning behind an answer.