AI for Lesson Plans: A Teacher's Guide

Lesson planning is the most common thing teachers use AI for, and the most common way they get it wrong. This guide covers the workflow end to end — from feeding the model your actual standard to differentiating for a mixed-ability class — and it starts with the rule that almost no vendor page will tell you: the student data you are about to paste in is probably protected by law.

The Challenge

Teachers do not have a creativity problem. They have a time problem, and a set of constraints that make generic AI output nearly useless.

• Prep happens after hours. Planning competes with grading, parent email, and going home. It is the task that expands to fill whatever time is left, which is usually none.
• A plan must map to a specific standard, in specific words, from a specific framework — Common Core, NGSS, a state framework, or a national curriculum. "Write a lesson about fractions" produces something no coordinator will sign off on.
• One class is not one learner. A single period may hold a reading-level spread of four or more years, students with IEPs, and English learners at several proficiency levels. A plan that serves the median student serves almost nobody.
• Materials have to be made, not just described. The plan is the easy part. The worksheet, the exit ticket, the rubric, the three differentiated versions of the reading passage — that is where the hours actually go.
• And student information is legally protected. This is the constraint that turns a productivity tool into a liability, and it is the one teachers are least warned about.

The scale of the time problem is documented. Gallup and the Walton Family Foundation surveyed more than 2,200 U.S. public school teachers and found that those using AI at least weekly save an average of 5.9 hours per week — about six weeks over a school year. Preparing to teach was the single most common use, reported by 37% of teachers at least monthly, followed by making worksheets and activities (33%) and modifying materials to meet student needs (28%).

A separate Gallup survey documents the gap, and it is wider than the adoption numbers suggest: just 18% of teachers report any formal guidance from their school on how AI should be used. Another 48% get only informal guidance, and 34% get none at all. The tools arrived faster than the guidance did.

How AI Solves It

Used well, a general-purpose model collapses the distance between "I know what I want to teach" and "I have the materials to teach it."

• It drafts to a structure you impose. Give it your standard's verbatim text, your time block, and your class profile, and it produces a plan in your required format rather than a generic one.
• It differentiates cheaply. Producing three versions of a reading passage at different reading levels is a long job by hand and a thirty-second job for a model. This is where the hours actually come back. (Reading levels, not Lexile measures — a Lexile is computed by MetaMetrics' proprietary algorithm, and a model asked for one will invent a number.)
• It generates the artifacts, not just the outline — exit tickets, rubrics, discussion questions, scaffolded sentence frames, a parent-facing summary.
• It plays the skeptical colleague. Ask it where the lesson will fall apart, which students will disengage, and what misconception the activity might accidentally reinforce. This is its most underused capability.
• It grounds in your documents. Tools like NotebookLM restrict the model to sources you upload, so a lesson can be built from your district's curriculum map rather than the model's memory of one.

What it does not do is replace your judgment about your class. Everything below assumes you are the one who checks.

Recommended Tools & Models

For lesson planning specifically, the differences that matter are: does it have a free tier, will it ground answers in your documents, and — decisively — can it legally hold student data.

Start with the top row, not the second. Almost every guide to this topic — including, until recently, this one — sends teachers to a consumer chatbot and then teaches them elaborate anonymization gymnastics to work around a restriction they did not have to accept. OpenAI's ChatGPT for Teachers is free through June 2027 for verified US K-12 educators, is self-serve, does not train on your inputs by default, and is built to hold student data inside a district workspace. Google extends enterprise-grade data protection to Gemini and NotebookLM across all Workspace for Education editions, including the free one. If either applies to you, the privacy section below becomes a safety net rather than a daily discipline. Check first.

Education-specific tools — MagicSchool, Diffit, Brisk, SchoolAI — wrap these same underlying models in a teacher-shaped interface and, importantly, several offer contracts that can satisfy district requirements. That contract, not the interface, is what you are actually paying for. Check whether your district has already licensed one before you buy anything yourself.

On models: any current frontier model plans a competent lesson. Claude Sonnet 5 and Gemini 3.5 handle long source documents well, which matters when you paste in a whole curriculum unit. GPT-5.5 is a strong generalist. The model is not the bottleneck — the prompt is.

The rule that comes before any tool

FERPA-protected education records cannot go into a consumer AI account. Not ChatGPT Free, not Plus, not Pro; not Claude Pro; not a personal Gemini account.

The reason is contractual, not technical. FERPA permits disclosure to a "school official" with a legitimate educational interest, and a vendor can qualify — but 34 CFR 99.31(a)(1)(i)(B) sets three conditions: the vendor performs a service the district would otherwise use its own employees for; it is under the district's direct control as to how it uses and maintains the records; and it is bound by § 99.33(a) against redisclosing them. On top of that, the district's annual FERPA notification must state the criteria for who counts as a school official — a vendor cannot be one if the district never said vendors can be.

Note what is not on that list: a signed contract. The Department of Education is explicit that FERPA does not require a written agreement, though it calls one a best practice. The written agreement you keep hearing about is mostly a creature of state law, not FERPA — and that law is often stricter. New York's Education Law § 2-d mandates a written data privacy agreement with any contractor receiving student data; California's SOPIPA binds the vendor directly whether or not your district signed anything; roughly forty states have something. FERPA is the floor, not the ceiling.

Either way, consumer accounts fail every version of the test. No consumer terms of service put a district in direct control of education records or bind the vendor against redisclosure, and no district's annual notice designates a consumer chatbot as a school official.

In practice, that means the following must never be typed into a consumer chatbot:

• Student names, initials, or student ID numbers
• IEP or 504 plan contents
• Grades, test scores, or assessment results tied to an individual
• Behavior incidents, disciplinary records, counseling notes
• Anything that would let a reader identify a specific child

COPPA governs the vendor, not you: it requires verifiable parental consent before an operator collects personal information from a child under 13, and the 2025 amendments added a separate consent requirement for disclosing that information to third parties. The FTC considered codifying a "school authorization" exception and declined to finalize it — so districts still rely on FERPA and state law, which is where your obligations actually live.

The workaround is easy, and it costs you nothing. Describe the need, never the student:

❌ "Marcus reads two grades below level and has an IEP for dyslexia. Adapt this passage for him."

✅ "One student in this class reads roughly two grade levels below the target and has a documented reading disability. Produce an adapted version of this passage: same content and vocabulary goals, decodable structure, shorter sentences."

The model produces exactly the same adapted passage. You have simply not handed a child's protected record to a company that has made you no promises about it.

Step-by-Step Implementation

1. Gather your three inputs first

Most bad AI lesson plans come from a thin prompt. Before opening any tool, collect:

The standard, in its exact published words. Go to your framework and copy the text. Do not cite it by code alone.

This matters more than it sounds. If you write CCSS.ELA-LITERACY.RL.5.2 and nothing else, the model will reconstruct the standard from memory — and it will often be subtly wrong. It will then build a perfectly coherent lesson on that wrong foundation, and the error is nearly invisible because everything downstream is internally consistent. Paste the words.

Your class profile, anonymized. Class size, period length, reading-level spread, number of English learners and their proficiency band, number of students with accommodations and the type of accommodation. No names. No IDs.

Your constraints. Available technology, whether students can take materials home, your required lesson-plan format, and what came immediately before this lesson.

2. Generate the plan

The last item is the one to keep. Asking a model to predict its own lesson's failure mode reliably produces the most useful paragraph in the output.

3. Differentiate the materials

This is where the time actually comes back. Take the core reading passage or problem set and ask for parallel versions.

That last constraint is what separates differentiation from tracking. Without it, models tend to produce three unrelated worksheets, and the class can no longer have one conversation.

4. Verify before you teach

Never skip this. Three specific things go wrong, and all three are invisible at a glance.

The standard. Compare the objective to the standard's real text, side by side. Did the model drift toward an adjacent skill? (Standards drift is the most common failure and the hardest to spot, because the lesson is coherent — just about the wrong thing.)

The facts. Every date, name, formula, historical claim, and scientific mechanism. Models produce fluent, confident, wrong statements — see hallucinations. A wrong date in a history lesson is a wrong date you taught thirty children.

The reading levels. When a model says "this is at a 3rd-grade reading level," treat that as a guess, not a measurement. It cannot reliably compute Lexile scores. Check the output yourself, or run it through the readability tool your district already uses.

5. Build a prompt you can reuse

Once a prompt produces a lesson you would actually teach, save it. Replace the specifics with brackets and keep it in a document. Your second lesson takes fifteen minutes instead of an hour, and the tenth takes five. The compounding return is in the template, not in any individual output.

Real-World Examples

Example 1: The thin prompt versus the loaded one

Before — a typical first attempt:

"Create a 5th grade lesson plan about the water cycle."

Output: A generic plan. Objective: "Students will understand the water cycle." A vocabulary list of four words. "Show a video." "Have students draw a diagram." "Discuss as a class." No timing, no differentiation, no assessment that would reveal anything. Nothing here that a teacher could not have written faster themselves.

After — the same request, loaded:

Standard, pasted verbatim: NGSS 5-ESS2-1 — "Develop a model using an example to describe ways the geosphere, biosphere, hydrosphere, and/or atmosphere interact." (Note what this standard actually asks for. It is not "the water cycle" standard — that is MS-ESS2-4, and it is middle school. A 5th-grade water lesson has to be framed as hydrosphere–atmosphere interaction, or you have already drifted off the standard before the model has written a word.) Class profile: 28 students, 45-minute period, reading levels spanning grade 2 to grade 7, six English learners at intermediate proficiency, three students with attention-related accommodations. Prior lesson covered states of matter. Chromebooks available; materials must work at home too.

Output: A minute-by-minute plan whose hook surfaces the specific misconception that clouds are made of water vapor (they are made of condensed droplets — the distinction is the entire lesson). Three versions of the diagramming task. Sentence frames — "The water changes from ___ to ___ because ___". An exit ticket whose second question is designed to catch the vapor/droplet confusion. And a predicted failure: "Students will draw the cycle as a closed circle and conclude water never leaves the system; be ready to ask where the water in their glass came from."

What changed: Nothing about the tool or the model. Only the prompt. The second output is a lesson; the first was a description of one.

Example 2: Differentiation that stays in one conversation

Before: A teacher asks for "an easier version" of a reading passage on the Silk Road. The model returns a passage with the vocabulary stripped out — no merchant, no caravan, no trade route. The struggling readers can now decode the text, and they can no longer participate in the class discussion, because the discussion is about merchants and caravans.

After: The same request, with the constraint added — "same key vocabulary, defined inline; a student on version A must be able to answer the same discussion question as a student on version C."

The model returns a passage with short, decodable sentences that still says merchant, and defines it in place: "A merchant is a person who buys and sells things. Merchants traveled the Silk Road." The vocabulary survives. The discussion holds together.

What changed: One sentence in the prompt, which encoded a pedagogical decision the model had no way to infer.

Example 3: The privacy near-miss

Before: A teacher wants a reading passage adapted for a specific student and types the natural thing: "Adapt this for Marcus — he's reading at a 3rd grade level, has an IEP for dyslexia, and gets frustrated and shuts down when text is dense."

That is a student's name, disability, reading level and behavioral profile, pasted into a consumer account with no FERPA agreement.

After: "Adapt this for a student reading roughly two grade levels below target, with a documented reading disability, who disengages when text is visually dense. Keep the key vocabulary, shorten sentences, increase white space."

What changed: The adapted passage is identical. The legal exposure is gone. This is the whole trick, and it costs nothing.

Industry-Specific Applications

Elementary. The highest-value use is generating decodable text at controlled reading levels — a genuinely time-consuming task by hand. Verify the levels; do not trust the model's self-report.

Middle and high school. Strongest for producing worked examples, distractor-rich multiple-choice items, and Socratic question ladders. Ask for the wrong answers a student would plausibly choose and why — it is an excellent diagnostic.

Special education. Enormous potential, and the sharpest legal edge. IEP contents are exactly what FERPA protects. Plan against the category of need, never against the child's record, unless you are inside a district-contracted tool.

English language learners. Sentence frames, cognate lists, and parallel first-language glossaries are quick wins. Have a fluent speaker check any output in a language you do not read — quality drops off outside high-resource languages, and you cannot audit what you cannot read.

Higher education. FERPA still applies. The consumer-account rule does not relax because the students are adults.

Best Practices

• Paste the standard's exact text. The single highest-leverage habit in this guide.
• Describe the need, never the student. Anonymized descriptions produce identical output with none of the exposure.
• Ask for the failure mode. "Where will this lesson fall apart?" is the most useful question you can ask a model about your own plan.
• Constrain differentiation to a shared discussion. Otherwise you get three worksheets and no class.
• Verify facts, standards and reading levels. Every time. These are the three things that break.
• Keep the model out of grading individual student work unless your district has a contracted tool for it. Student work is an education record.
• Save your prompts. The return compounds in the template, not the output.
• Check your district's policy. Most states have now published AI guidance, and district rules vary. Only 18% of teachers have had any formal guidance — the absence of training is not the absence of a rule.

Common Pitfalls

Standards drift. The model produces a lesson adjacent to your standard rather than on it. Because everything downstream is coherent, this survives a casual read. It is caught only by comparing the objective to the standard's real text, word by word.

Invented specifics. Confident, fluent, wrong. A date, a formula, a historical actor, a scientific mechanism. The fluency is precisely what makes it dangerous — see hallucinations.

Reading levels that are vibes, not measurements. A model asserting "this is 3rd-grade level" is guessing. It has no reliable way to compute readability. Verify externally.

Differentiation that fragments the class. Three versions with different vocabulary means three groups who cannot talk to each other. Constrain for a shared conversation.

Homogenized voice. Ten AI-planned lessons in a row start to sound like one another — same rhythms, same phrasings, same tidy three-part structure. Students notice before you do.

The privacy leak that feels helpful. Every instinct in teaching says: give the tool the full picture of the child so it can help the child. That instinct is exactly what FERPA prohibits here. It will feel wrong to withhold context. Withhold it anyway.

Trusting the education-tool badge. A tool marketed to teachers is not automatically FERPA-compliant. What makes it compliant is a signed agreement with your district — not the logo, not the pricing page. Ask your administrator which tools are actually covered.

Measuring Success

Time. Track prep minutes for two weeks before and after. The Gallup benchmark for weekly AI users is 5.9 hours saved per week across all tasks; lesson prep is the largest single component. Your number is the one that matters.

Whether the plan survived contact. Did you teach the lesson as generated, or rewrite it? A plan you rewrote at 11pm saved you nothing. Rewrites usually mean the prompt was underspecified, not that the tool failed.

Exit ticket data. The point is student learning, not teacher convenience. If AI-planned lessons are not moving exit-ticket results, the time saved is being spent on lessons that do not work.

Differentiation reach. Are your below-level and extension students both participating in the same discussion? That is the test that separates real differentiation from three parallel worksheets.

Zero privacy incidents. The metric you notice only when it fails. Audit your own prompt history occasionally: has a name crept in?

Cost Analysis

Tools. For US K-12 staff, ChatGPT for Teachers is free through June 2027 and is the one free tier that also clears the compliance bar. Failing that, the free tiers of ChatGPT, Claude, Google Gemini and NotebookLM are sufficient for everything in this guide that involves no student data. Paid consumer tiers run about $20/month and buy speed and higher usage limits — not compliance. A consumer plan at any price still cannot hold student data.

Time to set up. Two to three hours to build and refine your first reusable prompt template. That is the real investment, and it is a one-time cost.

What you get back. The Gallup figure — 5.9 hours per week for weekly users, roughly six weeks over a school year — is the documented benchmark. It is an average across all AI use, not a promise about lesson planning alone, and it applies to teachers who use these tools weekly. Occasional use does not produce it.

What it does not buy. Compliance, judgment, or knowledge of your students. Those remain yours.

Related Tools

• ChatGPT — fast drafting of plans, worksheets and rubrics
• Claude — long documents and careful differentiation
• Google Gemini — drafting inside Google Workspace
• NotebookLM — grounding lessons in your own curriculum documents
• Microsoft Copilot — drafting inside Word and PowerPoint
• Gamma — turning a finished plan into slides

Related Models

• Claude Sonnet 5 — strong on long source documents and nuanced adaptation
• Gemini 3.5 — large context, useful for whole-unit curriculum documents
• GPT-5.5 — capable generalist for drafting and materials
• Claude Haiku 4.5 — fast and cheap for bulk material generation

If you are new to writing prompts, start with prompt engineering — the difference between a thin prompt and a loaded one is the difference between a generic plan and a usable one.