How to encourage smarter AI use in the classroom

2026-08-26 · MIT Technology Review

Background: Classrooms Were Caught Off Guard

When chatbots became widely available, students suddenly had always-on tools that could answer homework questions or draft essays in seconds. Teachers can often detect AI use—models make distinctive mistakes, and some writing patterns can look artificial—but detection alone does not solve the problem. Instead, generative AI added pressure to educators already stretched by lesson planning, assignment design, and grading.

Although organizations such as OpenAI and UNESCO encourage classroom AI adoption, many teachers still lack practical clarity: what to allow, what to restrict, and how to preserve learning quality.

Case Study: Cheshire Academy’s Method-First Approach

Cheshire Academy, a private boarding and day school in Connecticut with about 400 students in grades 9–12, did not mandate AI adoption. Still, according to librarian and technology coordinator George Aiello, most instructors use AI in some form.

The school uses a patchwork of tools:

  • General chatbots: ChatGPT, Perplexity
  • Education-focused platform: MagicSchool

This mixed setup is intentional. Based on consultant advice, the school trained staff on transferable AI practices rather than prescribing one product. Training covered prompt-writing techniques and, equally, model limitations—especially inaccuracy and bias.

Current Use: Mostly Teacher-Facing Tasks

Teachers commonly use generative AI to prepare instructional materials, including:

  • Lesson planning
  • Assignment and quiz creation
  • Grading rubric generation

Some educators are interested in AI-assisted student feedback, but none have implemented it yet due to concerns about:

  • Feedback quality
  • Personalization
  • Privacy

Pedagogical Example: Turning AI Into a Reflection Tool

French teacher Miriam Przybyla-Baum does not rely on AI for her own material creation, largely because she has built extensive resources over nearly 30 years. However, she has long seen students use tools like Google Translate to shortcut language assignments.

She developed reflection-based activities:

1. Students let an LLM edit their homework, then evaluate which edits are valid and which erase their voice.

2. Students anonymously peer-grade AI-assisted assignments and annotate which parts appear AI-generated.

The goal is not only compliance, but student judgment: understanding what AI can and cannot do in language learning.

Schoolwide Governance: From Experiments to Shared Norms

Cheshire Academy is piloting a Student AI Council, where students create media and lead discussions on healthy AI use. The intent is to make students active participants in defining beneficial AI norms for their community.

The school has also adopted a traffic-light labeling system for assignments:

  • Green: AI fully allowed
  • Red: No AI allowed
  • Yellow: Partial allowance (e.g., spell-check allowed, chatbot messaging prohibited)

This converts vague expectations into explicit, enforceable rules.

Tool Snapshot: MagicSchool’s Value and Limits

As generative AI entered the mainstream, Cheshire previewed MagicSchool for staff. Its key strength is breadth in one platform:

  • Multi-subject, multi-grade question and assignment generation
  • Dedicated rubric generator with ready-to-use tables
  • Presentation, lesson plan, and administrative report support

Teachers can provide structured inputs (grade level, question count, question type, and aligned documents). Still, not all teachers are comfortable generating student-facing text with LLMs, mainly due to effectiveness and accuracy concerns.

MagicSchool offers free and paid tiers; individual plans with unlimited access and full records cost just under $100 per year. Many teachers continue using general-purpose tools from Anthropic, Google, and OpenAI for planning and administrative work.

Practical Takeaways

  • Start with shared methods and risk literacy, not a single mandated tool.
  • Replace blanket bans with reflective pedagogy that builds student discernment.
  • Use clear policy signals (like traffic-light labels) to reduce ambiguity.
  • Prioritize real instructional pain points and iterate through small pilots.

Source