AI for educators: the real playbook from Brazilian classrooms in 2026

Public conversation about AI in education in 2026 is stuck between two unhelpful extremes. On one side: “AI will replace the teacher.” On the other: “ban ChatGPT from school.” Teachers on the front line — the ones who actually teach, grade essays, write weekly lesson plans — live a different reality: they’ve been using AI personally since 2023, under the radar, because nobody trained them and the school’s policy hasn’t been decided yet.

This post is for that teacher. No hype, no alarm. What works in a real classroom in 2026, what still doesn’t, and the 4 costliest mistakes we see.

Why this playbook is rooted in Brazilian classrooms (and why that matters elsewhere)

Most material on AI in education published in 2024–2025 came out of the US and addressed problems that aren’t the same as ours. But the four Brazilian constraints below also describe a wide chunk of public education globally — low-bandwidth schools, language-specific tooling, standardized exams as the budget driver, and curriculum mandates being added faster than teacher training. Translate the constraints to your country and the playbook mostly carries over.

1. National curriculum mandate (BNCC in Brazil; Common Core / National Curriculum / equivalent elsewhere). Brazil’s BNCC was updated in 2024 to add “critical and ethical use of artificial intelligence” as required content under digital culture. Curricula in Spain, Singapore, and several US states followed in 2025. AI literacy is no longer optional in many places.
2. The national exam funnels everything. In Brazil that’s ENEM (national university entrance). In the US, SAT/ACT. In the UK, A-levels. In India, JEE/NEET. Any AI tool that ignores the exam structure leaves teachers without ammunition for the parents-and-principals conversation.
3. Infrastructure inequality is the base problem. Teachers in well-funded private schools and teachers in underfunded public schools operate opposite realities. Most AI tools assume broadband + one device per student — still not universal anywhere.
4. Local language nuance matters. Accents, regionalisms, formal-register essay grading. Generic models trip on these. Models calibrated for the local language (Claude and Gemini both improved substantially for PT-BR in 2026, and similar trajectories happened for Spanish, French, Hindi, and Arabic) work meaningfully better than the 2023 free-tier baseline.

The catalog of what works

1 · Lesson planning anchored on what the national exam actually covers

You teach history and you have to plan the final year for exam-bound students. The good input: the last 5 national exam papers. Paste the PDF (or transcribed text) into Claude/Gemini and ask:

“From the 5 most recent national exam papers (2021–2025), list in order the top 10 history topics most frequently tested in objective questions. For each, give: typical question stem, which curriculum competency it maps to, and where my current textbook (textbook X) is weakest on it.”

This collapses planning from a half-day-per-month task to 30 minutes. You review, adjust, and get real focus.

Where it fails: if you don’t review critically, the model sometimes invents a “curriculum competency” — that’s a hallucination. Always cross-check with the official exam documentation.

2 · Essay grading with explicit criteria

National-exam-style essay grading follows a defined rubric (in Brazil’s ENEM, five competencies). A generic AI vendor doesn’t know your rubric. But you can teach the model in the prompt:

“You are an essay grader following the official 5-competency rubric: 1) command of standard written language, 2) grasp of the prompt, 3) selection and organization of arguments, 4) demonstration of linguistic mechanism mastery, 5) intervention proposal respecting human rights. Score 0–200 per competency. Justify every score by citing passages from the text. Do not invent errors — only comment on what’s actually written.”

Then paste the student’s essay. You get a reasonable first pass that you refine in 5 minutes. Compare with 25–40 minutes per essay for full manual grading.

Where it does NOT work: small classes (up to 8 students), short essays, or when you want deep formative correction. For critical-citizenship feedback, role-modeling, fine pedagogical intervention — humans are irreplaceable.

Ethics note: tell students that AI is part of the first-pass grading. Don’t hide it. In parallel, require a final human grade signed by you — AI assists, teacher decides. That’s what most education councils (and parents) expect.

3 · Differentiating exercises by classroom level

Same exercise, three versions: for the student below the class baseline, for the student at baseline, for the student running ahead. AI generates all three variants in 2–3 minutes from the original exercise.

“I use this math exercise [paste]. Generate 3 variants: A) simplified version for students who haven’t consolidated prerequisite X, B) standard class version, C) extension version with 1 extra challenge that bridges to the next unit’s content.”

Differentiation has been pedagogically desirable for 30 years. The production cost made it prohibitive. AI collapses that cost.

4 · Personalized written feedback from bullet points

You graded 80 essays over the weekend. You have to write individualized feedback for each student. Instead of everyone getting “great” / “watch agreement” / “study punctuation,” paste your bullets per student and ask:

“For each student, write a 4–6 line paragraph in [target language], respectful and formative tone, addressed to the student, mentioning 1 strength and 1 area to develop. Use only the bullets I provided — don’t invent anything else.”

Result: each student gets feedback that reads as human and personalized. Because it is. You just didn’t type each paragraph from scratch.

5 · Formative quiz in Google Forms / Microsoft Forms from a PDF

You just finished the lesson on the Industrial Revolution. You want an 8-question quiz at the end to check comprehension. Paste the lesson PDF + objectives:

“Generate 8 questions for a short formative quiz on the Industrial Revolution. Mix of 5 multiple-choice (4 options each, justify the correct answer) + 3 short open. Target Bloom’s taxonomy level ‘apply’, not ‘remember’. Local-language school register.”

Paste into Forms in 3 minutes. Run in the last 7 minutes of class. You have a real thermometer of what stuck.

6 · Real-time accommodation for inclusion students

Common scenario: a deaf student (sign language), a student with dyslexia, a recent immigrant (still acquiring the school’s primary language). Generative AI in 2026 does decent sign-language adaptation (locale-specific models), simplification to lower reading levels, or translation to a heritage language (Spanish, Haitian Creole, Arabic — depending on your local immigration patterns).

You paste the original text and request the appropriate variant. In parallel, always keep the original available.

The 4 mistakes that kill the initiative

Mistake 1: letting AI pass/fail a student

In most jurisdictions, the policy is clear (in Brazil since 2025; in the EU under AI Act Article 50 from Dec 2026): high-stakes summative assessment that decides progression or retention needs accessible human review. AI assists, teacher decides. Schools that invert that order hit legal trouble and lose family trust.

Mistake 2: using AI as an unmediated tutor for the student

“Student chats with ChatGPT alone on their phone during class” sounds modern and in practice is a disaster. Without teacher mediation, the student copies, doesn’t learn; or they get bad information and adopt it with confidence because AI “sounds confident.”

What works: student uses AI with a structured task pre-defined by the teacher, inside class, with a debrief after. What fails: “take out your phone and ask the AI.”

Mistake 3: trusting automated AI-text detection

In 2026, automated detectors of “text written by AI” remain unreliable. False positives punish honest students; false negatives let copiers through. Don’t use as evidence.

What works: redesign the assessment so that AI use is part of legitimate practice (with explicit criteria) or so that direct copying is structurally infeasible (in-class assessments, oral defenses, long projects with checkpoints).

Mistake 4: generic 4-hour AI training disconnected from each teacher’s context

A 4-hour “AI in Education” course for 200 teachers without direct hands-on in each one’s context → adoption near zero by day 60. What works: 60 minutes in groups of 8–12, with each teacher bringing a real problem (one class, one lesson plan, one essay) and building the solution live with a facilitator.

SkilLab runs that format in Brazilian schools with 60–70% adoption by day 90, vs ~20% from the generic course.

The golden rule for 2026

Before any AI-assisted educational task, ask: is the AI doing the work the student was supposed to learn, or the work the teacher was supposed to scale?

• Student’s work (reasoning, writing, debating) → AI stays out, or enters under heavy mediation.
• Teacher’s work (volume grading, repetitive planning, scalable feedback) → AI accelerates dramatically.

The confusion between these two is the root of nearly every serious problem with AI in education. Solve that, and the rest is a catalog of practices.

Where to go deeper

• AI for students — the flip side: how students should (and shouldn’t) use AI to study.
• Workshop Claude Cowork — practical training format for teachers and school administrators, 60–90 minutes in small groups.
• Productivity AI cluster — other tools teachers use (Copilot, Workspace, NotebookLM).

---

By Ivan Prado · SkilLab AI · May 2026. Translated and culturally adapted from the PT-BR original.