How one high school is using red, yellow, and green lights to guide smarter AI use in class

AI homework adds an invisible layer of work for teachers

Students can pick up their phones and generate answers or essays in seconds. Teachers can sometimes spot common model errors, formulaic sentence patterns, or punctuation habits, but those traces can't reliably prove who wrote what. Basing judgments on whether something "looks like AI" leads to both false positives and missed cases.

Generative AI also adds new responsibilities for teachers. Beyond lesson planning, assignment design, and grading, they now have to determine whether a given usage violates policy, whether content is reliable, and what level of AI participation a task actually permits. Organizations like OpenAI and UNESCO can advocate for AI literacy, but classroom teachers still have to translate those principles into concrete instructions on the next assignment.

Classroom rules need to answer at least three practical questions:

• Can students accurately restate what's allowed and what's prohibited before they start an assignment?
• When a teacher encounters a new tool, can they apply existing rules rather than building a new blocklist of products?
• After a student uses AI, can the teacher see the student's verification, trade-offs, and revision process?

Cheshire Academy's approach starts with those three execution-level questions.

The school trains general skills first, without mandating a single tool

George Aiello, the school's technology coordinator, says most teachers already use AI in some form. That's his observation, not an independent survey. The school also doesn't require everyone to use it. Miriam Przybyla-Baum, who has taught French for nearly 30 years, has built up a large library of classroom materials and doesn't think she needs AI for lesson preparation.

Following consultant advice, the school focuses training on cross-product capabilities: how to write workable prompts, how to spot errors and bias, and how to verify generated output. Teachers can choose from ChatGPT, Perplexity, MagicSchool, or other tools. When products change, the evaluation methods remain relevant.

Teachers currently use AI mainly in the preparation phase:

• Brainstorming activities and lesson plans for new units;
• Generating supplementary exercises and worksheets for a grade level;
• Drafting rubrics, slide decks, or administrative materials.

Some teachers have considered having AI generate feedback directly for students but ultimately decided against it due to quality, personalization, and privacy concerns. That hesitation is telling: the closer the output gets to a specific student, the more immediate the impact of any error, and the heavier the required human review.

Red, yellow, green: baking permissions into every assignment

Cheshire Academy's 2026–27 student and family handbook, section 2.8.3, requires every assignment to be labeled red, yellow, or green—and all assignments default to red.

The three colors carry different permissions:

• Red: No generative AI allowed.
• Yellow: AI is allowed, but the teacher must define how and to what extent. The article gives the example of using spell-check while not asking a chatbot for answers.
• Green: Generative AI is allowed.

Yellow assignments demand the most from task design. Simply writing "AI may be used appropriately" doesn't resolve ambiguity. Students still don't know which stage they can use it, whether they can generate the main text, whether they need to disclose their process, or who verifies their citations.

The assignment cards below are an operational rewrite based on the Cheshire case. The school's original documentation doesn't include this table:

Red-light assignment card

• Allowed: Non-generative tools that the teacher explicitly permits.
• Prohibited: Using generative AI to create, modify, or answer any part of the task.
• Process evidence: A statement of independent completion; drafts kept when necessary.
• Accountability: Students are responsible for all content.

Yellow-light assignment card

• Allowed: Brainstorming, error-checking, translation, or coding assistance as specified by the teacher.
• Prohibited: Generating main text, answering on behalf of the student, or rewriting without permission.
• Process evidence: Note where AI contributed, and list the suggestions you adopted, revised, or rejected along with your reasoning.
• Accountability: Students verify facts, citations, formulas, and wording.

Green-light assignment card

• Allowed: Generation and collaboration as permitted by the task description.
• Prohibited: Hiding AI use, submitting unverified output, or entering protected information.
• Process evidence: A description of AI usage, key prompts or revision logs, and the final verification results.
• Accountability: Students own the final product; teachers review against the task's objectives.

A yellow-light writing assignment could be written directly as:

Spell-check and AI brainstorming are allowed; do not have a chatbot generate the main text. When you submit, mark where AI contributed and list three suggestions you adopted, revised, or rejected, with reasons. You are responsible for the final text, facts, and citations.

The color provides the boundary; the actions, evidence, and accountability make that boundary enforceable.

Students turn in more than answers — they turn in their judgment

Przybyla-Baum encountered students using Google Translate as a shortcut early on. Her French assignment didn't stop at "no translation tools." Students first had a large language model edit their own work, then went through each change to judge which ones were correct and which stripped away their personal voice.

The AI produces suggestions, but students still have to verify, choose, and explain. That lets teachers observe the learning process rather than just a clean final answer.

The article also describes two other approaches. One is anonymous peer review of AI-assisted work, where students mark which passages look AI-generated—this works as an exercise in language observation, not as evidence of cheating. The other is a student AI council, currently in pilot, where students create content, lead discussions, and consider how AI should be used to support learning and the school community.

These cases can be turned into ready-to-assign classroom exercises:

Exercise One: AI revision audit

1. Students submit their own original text.
2. Have the model suggest edits, without overwriting the original.
3. Mark each suggested change as "accept," "revise," or "reject."
4. Explain at least three decisions: whether it's correct, whether it's clearer, and whether it still sounds like your own writing.
5. Submit the original, the AI suggestions, the decision log, and the final draft together.

Exercise Two: Anonymous language diagnosis

1. The teacher anonymizes several AI-assisted texts.
2. Students flag factual errors, vague phrasing, sudden style shifts, and missing sources.
3. The class compares disagreements and discusses why a "feel" for language can't prove authorship.
4. Grade on the quality of evidence and revision suggestions, not on whether students correctly guess the author.

Exercise Three: Student rule review

Students collect the yellow-light assignments most often misunderstood this term, identify vague language in the rules, propose rewording, and create a one-page student-facing guide. The student AI council meets monthly to compile new tools, recurring misunderstandings, and real conflicts, which teachers then use to update examples.

All three exercises train observable actions: spotting errors, comparing options, preserving personal voice, and explaining decisions. Those skills hold up even when the model changes.

MagicSchool can speed up prep, but it can't take over validation

MagicSchool is the education-specific tool highlighted in the article. It brings quizzes, worksheets, rubrics, presentations, lesson plans, and administrative reports into one platform. Teachers generating assignments can specify grade level, number of questions, question type—multiple choice or short answer, for instance—and upload reference materials so the output stays aligned with the curriculum.

As of 2026/8/25, MagicSchool's official pricing page shows: the personal Free tier at $0; Plus at $8.33 per month billed annually (total $99.96) or $12.99 month-to-month. Plus includes unlimited generation, full output history, editing, and advanced features. Enterprise pricing for schools and districts requires a quote. Prices and features will keep changing; these figures are only meant to give a sense of the procurement scale.

The convenience of education-specific platforms comes primarily from templates, history, and school-level administration. It doesn't eliminate four categories of review:

Brainstorming and additional question types

• Benefit: Rapidly expands the range of options, especially helpful for newer teachers with limited question banks.
• Risk: Repetitive content or mismatched difficulty.
• Check: Teachers filter and rewrite to match curriculum goals.

Worksheets, lesson plans, and rubric drafts

• Benefit: Saves time on structuring materials.
• Risk: Factual errors or distorted grading standards.
• Check: Go through each item against the textbook and learning objectives.

Student-facing materials

• Benefit: Quick generation of instructions, example problems, or draft feedback.
• Risk: Stiff tone, lack of personalization, misleading students.
• Check: Teachers rewrite the final text and check for privacy.

Formulas, citations, and factual claims

• Benefit: Fast generation.
• Risk: Hallucinations slip directly into instructional materials.
• Check: Recompute formulas, open the original sources, and verify facts sentence by sentence.

Before purchasing anything, ask at least five questions: where does student data go, can the school control accounts and permissions, is output history traceable, can teachers edit before sending, and who is responsible when something goes wrong. No amount of features replaces this set of governance questions.

Related readingDartmouth Put AI Grading to the Test: Students Called It Rigid — But Users Scored HigherIn a 151-student trial, short-answer questions moved scores more than multiple choice, while almost no one touched the AI help sidebar→

Three recommendations from the article, turned into concrete teacher actions

MIT Technology Review closes with three practical recommendations. Rewriting them as work actions makes them easier for teachers to adopt directly.

1. Acknowledge that students will try AI and teach boundaries in advance

Homework can't be fully monitored by the school. In the first week of class, teachers can show students a reliable answer, a wrong answer, and a fabricated source for the same question, so they can see how variable model output is. Then use the three-color assignment card to explain what's allowed in your course. Setting the rules before the first conflict is less work than guessing at student motives after the fact.

2. Demonstrate your own verification process

If a teacher sends out AI-generated materials without checking them, students learn that "having an answer is enough to ship it." In class, do a public review once: delete vague sentences, open the cited sources, recompute a formula, and explain why you replaced a particular question. What students see is the complete workflow.

3. Draft with AI, then revise or replace with human judgment

AI works well for helping newer teachers brainstorm lesson plans, expand question banks, and generate initial rubrics. Before any material goes to students, run through this checklist:

• Does every citation open and support the claim it's attached to?
• Has every formula and numerical value been recomputed?
• Have factual statements been checked against textbooks, papers, or official sources?
• Do the difficulty, wording, and examples fit the current grade level?
• Does the content contain student names, grades, health information, or other data that shouldn't go into a model?
• Would you be willing to put your name on this material and accept responsibility for it?

If any item fails, the material isn't done yet.

To replicate this approach, a school needs five things at minimum

The red-yellow-green system can be piloted in a week, but it can't just be a poster on the wall. The minimum deployment package includes:

1. One-page course rules: define red, yellow, and green, state the default status, and explain how violations are handled.
2. Per-assignment instruction blocks: for each task, write down the color, allowed actions, prohibited actions, process evidence, and the responsible person.
3. One judgment practice session: let students actually go through generating, verifying, accepting, correcting, and rejecting.
4. Teacher delivery checklist: before anything is published, manually check all AI-generated citations, formulas, facts, and student feedback.
5. Feedback and revision mechanism: teachers and students regularly collect the rules that are most often misunderstood, update the examples, and avoid rewriting the policy every time a new product appears.

A pilot is not just about counting how many times AI was used. At minimum, track four outcomes: whether students can state their permissions before they start, how many boundary disputes arise from yellow-light assignments, the quality of the judgment evidence students submit, and how much time teachers spend preparing and reviewing. Add a spot-check on output quality, recording errors in facts, citations, and formulas. Only then does a school have the data to decide whether the system is worth scaling.

Existing reports do not include this data. They do not show that the red-yellow-green system improved grades, reduced cheating, or cut teacher workload. The student AI committee is still a pilot, and teachers generating personalized feedback directly with AI has not been rolled out either. Cheshire Academy shows how the system could work, but actual results still need to be measured by the school.

Schools can test this on the next assignment: first mark the color, then list allowed actions, require judgment evidence, and finally assign a responsible person. When all four elements are present, AI in the classroom shifts from a vague stance to an inspectable set of task rules.