The Irvington AI Framework

Clarity · Capacity · Community

Teach the tool. Test the thinking. Trust the process.

Irvington Township School District — Department of Education
Professional Learning Framework for Artificial Intelligence, Grades K–12

The One-Sentence Version

**We are not trying to stop students from using AI. We are trying to make sure that
when they use it, learning still happens — and that when they don't, we can tell the
difference.**

The Premise

AI did not create the problem of students completing six months of coursework in six hours.
It exposed it.

If a course can be finished by a machine in an afternoon, that course was measuring
compliance — submissions, seat time, completed boxes — rather than learning. The machine
simply became very good at compliance, very fast, for free.

This is uncomfortable, and it is also the good news: it means the fix is within our
control. We do not need better detection software. We need better evidence of learning.

The Three Pillars

        ┌───────────────────────────────────────────────────┐
        │                                                   │
        │   CLARITY          CAPACITY         COMMUNITY     │
        │   ────────         ────────         ─────────     │
        │   Everyone         Build what       Respond to     │
        │   knows what       AI can't         misuse as     │
        │   is allowed       fake             teaching,     │
        │   on THIS                           not policing  │
        │   assignment                                      │
        │                                                   │
        │   → Red/Yellow/    → Process-       → Conversation│
        │     Green on         weighted         before      │
        │     every task       assessment       accusation  │
        │                                                   │
        └───────────────────────────────────────────────────┘
                              ↓
              Teach the tool. Test the thinking. Trust the process.

Each pillar solves one of the three problems in the district's situation. They are not
sequential — they run together — but if you can only do one thing this year, do Clarity,
because it is free, it takes ten minutes per course, and it eliminates the most common
cause of AI conflict: the student genuinely did not know.

PILLAR 1 — CLARITY

"Everyone knows what is allowed on THIS assignment."

The problem it solves

The most common AI incident in a classroom is not defiance. It is ambiguity.
A student used a grammar checker. Another used a translation tool. Another asked a chatbot
to explain the prompt. Another had it write the whole thing. In most classrooms, all four
students received the same vague instruction — "don't use AI" — and formed four different
interpretations of what that meant.

A blanket ban is not clarity. It is an untested assumption that everyone defines "use AI"
the same way. They do not.

The tool: The AI Traffic Light

Put one of these three lines on every single assignment. Not on the syllabus once in
September — on the assignment, every time, the way you already put a due date.

🔴 RED — No AI

*"This task is AI-free. I need to see your unassisted thinking. Work is done in class,
on paper or on a locked device. If you're stuck, ask me — that's what I'm here for."*

Use when: you are establishing a baseline, assessing a foundational skill, or the
struggle itself is the learning. Early drafting. Timed writes. Math procedure practice.
First attempts at anything.

Make it possible to comply. A Red task that is assigned as homework on a laptop is not
a Red task; it is an honor-system experiment. Red tasks belong in class.

🟡 YELLOW — AI with disclosure

*"You may use AI for specific, named steps: brainstorming, outlining, checking grammar,
explaining a concept you're stuck on, or generating practice problems.
You may NOT use it to produce the text you submit.
At the end, add three lines: what tool you used, what you asked it, and what you changed."*

Use when: this is most of school. Yellow is the default setting for the majority of
work in grades 6–12.

The three-line disclosure is the entire mechanism. It costs the student 30 seconds and it
does four things at once: it makes honest use easy, it makes dishonest use a deliberate
choice rather than a drift, it gives you diagnostic information about how your students
think, and it teaches a professional norm they will need for the rest of their lives.

🟢 GREEN — AI encouraged

*"Use AI as hard as you can. The skill I'm assessing is your ability to direct it,
judge it, and improve on it. Submit your prompts and the AI's raw output alongside
your final work — I'm grading the distance between them."*

Use when: the learning target is AI fluency, critique, revision, or evaluation.
Also excellent for: comparing AI output against a primary source, fact-checking exercises,
producing a first draft that students must then dismantle.

The Green move that changes everything: grade the critique, not the output. Have the AI
write the essay, then have the student mark it up in a different color — every unsupported
claim, every vague sentence, every missing citation. A student who can red-pen an AI essay
understands the genre better than a student who wrote a mediocre one from scratch.

The Clarity Rule

If you have not told students which light is on, the light is Yellow.
You cannot penalize a student for a rule that existed only in your head.

10-minute implementation

1. Open your unit plan.
2. Next to each assessment, write R, Y, or G.
3. Paste the matching sentence into the assignment description.
4. Done. Repeat once per unit.

That's it. That is the single highest-return action in this entire framework.

PILLAR 2 — CAPACITY

"Build the things AI can't fake."

The problem it solves

If the final artifact is the only evidence of learning, AI wins — always, permanently, and
at no cost to the student. No detector fixes this, because the problem is not detection.
The problem is that we asked for a product and a product is exactly what a generative model
produces.

The answer is to collect evidence AI cannot generate: **process, presence, and personal
context.**

The Evidence Stack

Ordered from weakest to strongest. Notice where the software sits.

  STRONGEST   ┌──────────────────────────────────────────────┐
      ▲       │ 6. The student can explain it out loud       │  ← unfakeable
      │       ├──────────────────────────────────────────────┤
      │       │ 5. In-class baseline you watched them write  │  ← unfakeable
      │       ├──────────────────────────────────────────────┤
      │       │ 4. Version history / revision trail          │  ← very hard to fake
      │       ├──────────────────────────────────────────────┤
      │       │ 3. Drafts, notes, annotated sources          │  ← hard to fake
      │       ├──────────────────────────────────────────────┤
      │       │ 2. Personal / local / classroom-specific     │  ← hard to fake
      │       │    content only they could know              │
      │       ├──────────────────────────────────────────────┤
      │       │ 1. An AI-detector percentage score           │  ← NOT EVIDENCE
   WEAKEST    └──────────────────────────────────────────────┘

Everything above the bottom line is free. You do not need to purchase anything to make
your classroom AI-resilient. You need to change what you collect.

The Five Moves

Move 1 — The September Baseline

In the first two weeks, have every student write for 20 minutes, in class, on paper or a
locked device, on a low-stakes personal prompt. Keep it.

This is the highest-leverage 20 minutes of your year. It gives you a calibrated sense of
each student's unassisted voice, vocabulary, sentence rhythm, and characteristic errors.
Later, when something feels off, you are comparing against that student's actual baseline
rather than against your general impression of how a teenager writes.

Repeat in January and April. The drift between baselines is also your growth data.

Move 2 — The Two-Minute Oral Check

For any significant submission, spend two minutes with a rotating sample of students:

"Walk me through how you got to this paragraph."
"Why did you choose this source over the other one?"
"What did you cut, and why?"
"What was the hardest part?"

A student who did the work answers instantly and with texture. A student who did not cannot
fake it, and — importantly — you have not accused anyone of anything. You asked about their
work, which is a normal thing a teacher does.

Announce in advance that oral checks are routine and random. The deterrent effect of a
known, ordinary, non-punitive check is larger than the detection effect.

Logistics: 5 students × 2 min = 10 minutes while the class does independent work.
You do not need to check everyone. You need everyone to know they might be checked.

Move 3 — Make the Prompt Un-Googleable

Anchor tasks in something the model cannot have seen.

The pattern: local + personal + specific + tied to a shared classroom event.
AI can still help — and that's fine — but it cannot do it without the student, which is
the entire goal.

Move 4 — Grade the Process, Weight the Product Less

A rubric that puts 100% of the points on the final artifact is an invitation.

Suggested redistribution for a major task:

A student who uses AI on the final product but genuinely did every other stage has
learned the thing. That student should pass. That is not a loophole — that is the point.

Move 5 — Teach the Disclosure Habit

Require the three-line AI note on every Yellow task, all year, until it's automatic.

AI USE NOTE
Tool: ____________________
I asked it to: ____________________
What I changed or rejected: ____________________

That third line is where the learning lives. "I changed nothing" is itself a data point
worth a conversation. Over a year, this builds the single most transferable AI skill a
student can leave Irvington with: the habit of accounting for their tools.

PILLAR 3 — COMMUNITY

"Respond to misuse as teaching, not policing."

The problem it solves

The fastest way to destroy a classroom is to make it adversarial. Once students believe
you are hunting them, three things happen: honest students get anxious, dishonest students
get better at hiding, and the students who most need help stop asking for it.

And the enforcement-first approach carries a specific equity risk that Irvington cannot
afford to ignore: **AI-detection tools produce false positives at higher rates for
multilingual writers and for students who write in a plain, formulaic register.** In a
district as linguistically diverse as ours, a detector-driven discipline process would not
distribute its errors evenly. It would concentrate them.

One wrong accusation costs more trust than ten correct catches recover.

The Conversation Protocol

When something feels off, use this. It takes four minutes and it is designed so that you
are never wrong — because you never make a claim.

Step 1 — Gather process evidence first. Not a detector score. Version history, the
baseline, the drafts, the disclosure note.

Step 2 — Open with curiosity, not accusation. The exact words matter:

❌ "Did you use AI on this?"
→ A yes/no trap. Invites a lie, and if they say no you have nowhere to go.
✅ "I'd love to hear how you put this together — walk me through your process."
→ Not an accusation. Impossible to fail if they did the work.
Most cases resolve here, in both directions.

Step 3 — Name what you observe, not what you conclude.

✅ *"This doesn't sound like the writing I saw from you in September, and I want to
understand why. Help me out."*

You are describing a fact (a difference) and asking for their account. You have not called
anyone a cheater.

Step 4 — Listen for the actual reason. In practice, the reasons are usually:
overwhelmed, behind, didn't understand the task, afraid to ask, working a job, caring for
siblings, or genuinely thought it was allowed. Very rarely: contempt for the course.
The response should match the reason.

Step 5 — Redirect to learning. The consequence should teach the thing that was skipped.

✅ *"Here's what we're going to do: you'll redo this with me during lunch on Thursday,
and we'll do it in stages so I can see your thinking. This one doesn't count against you."*
❌ Zero, referral, phone call home, done.

A zero teaches "don't get caught." A redo teaches the content.

Step 6 — Escalate only for pattern or defiance. A repeated, deliberate pattern after
a clear conversation is a different matter and follows the district's academic-integrity
code. A first incident is a teaching moment, full stop.

The Amnesty Reset

When you introduce the traffic-light system, open with a clean slate — out loud:

*"The rules were fuzzy before. That's on me, not you. Starting now, every assignment tells
you exactly what's allowed. I'm not going back and re-litigating anything from before
today. Going forward, I'll believe you if you tell me, and I'll be disappointed if you
don't."*

This costs nothing and it converts a large number of students in a single sentence,
because it removes the incentive to keep hiding past behavior.

Building the Friendly Room

• Use AI openly in front of them. Project it. Let it fail. Say *"that's wrong, here's

how I know."* A teacher who models critical use is far more persuasive than one who
forbids it.

• Never use a detector score as the basis of an accusation. Use it, at most, as a

private prompt to go look at the process evidence.

• Praise disclosure loudly. When a student writes an honest AI-use note, say so.

"This is exactly right — you told me what you used and what you changed."

• Let students be the experts sometimes. They know the tools. Ask them. A student who

teaches you something is not a student who is trying to beat you.

• Separate the tool from the student. The problem is never "you're a cheater." It's

"this assignment didn't get your thinking, and your thinking is what I'm here for."

The Speedrun Problem: A Direct Answer

"Six months of coursework in six hours"

This is a systems problem, not a classroom-management problem, and it cannot be fixed
by teachers alone. Here is the honest division of labor.

What the teacher controls

What the building administrator controls

What the district controls

Say this out loud in the training: *"Rows 2 and 3 are not your job. We are showing
them to you so you know they are being handled, and so you can stop feeling responsible
for a hole you did not dig and cannot fill."*
Teachers disengage from PD when they're handed strategies they lack the authority to
execute. Naming the boundary buys enormous credibility.

Detection: The Honest Briefing

Teachers will ask for a reliable detector. Here is what to tell them.

What is true

1. No AI detector is reliable enough to justify a disciplinary consequence on its own.

They output probabilities, not proof.

1. False positives are not randomly distributed. Multilingual writers, students with

plain or formulaic prose styles, and students who use grammar-assistance tools are
flagged more often. In Irvington, that is an equity problem with a name.

1. Bypass is trivial. "Humanizer" and paraphrasing services exist specifically to

defeat detectors, cost nothing, and take one click. Any enforcement strategy built on
detection is an arms race the school loses by design.

1. The best signal is free and requires no software: a September baseline, version

history, and a two-minute conversation.

What a detector score is good for

One thing only: a private prompt to yourself to go look at the process evidence.
It is the beginning of your own investigation, never the end of it, and it is never
shown to the student or a parent as proof.

The landscape teachers should be aware of

This is for assignment design awareness, not for surveillance. The point of knowing
these exist is to understand why product-only assessment cannot be secured — which leads
back to Pillar 2.

• Chatbots and assistants — general-purpose, now built into phones, browsers, search

results, and office software. "Did you use AI" is becoming unanswerable because AI is
ambient.

• Browser extensions — sidebar assistants that read and answer whatever is on screen,

including inside an LMS or a quiz.

• "Homework helper" apps — photograph a problem, receive a worked solution.
• Humanizer / paraphrase services — rewrite AI text specifically to defeat detectors.
• Proxy and mirror sites — used to reach blocked services from a school device.
• Autocomplete everywhere — predictive text in docs, email, and phone keyboards is

already AI assistance that no policy meaningfully covers.

The conclusion to draw: you cannot build a wall high enough. You can build assessments
that don't need a wall.

The Differentiated Room

The staff will not arrive at the same place. Design for that explicitly.

The AI Weather Report

Open the session by having everyone self-place. Anonymous. No judgment attached.

The Red Team

The most important design decision in this training.

Teachers who are critical of AI are assigned the formal role of finding where it fails:
hallucinations, bias, privacy problems, bad pedagogy, harm to students. Their findings get
written up and distributed district-wide as the Irvington AI Failure Log.

Why this works:

• It asks for no change of opinion. You can think AI is bad for kids and be the most

valuable person in the room.

• Skepticism is reframed from resistance into expertise.
• It is genuinely the most valuable work happening that day. Guardrails are built by

people looking for failure.

• It signals to the entire middle of the room that this is not a sales pitch — which is

the only way the middle will trust anything else that gets said.

The only ask this training makes of anyone: don't be sold, be competent.
Nobody has to like AI. Everyone has to be able to recognize it, evaluate it, and teach
students to do the same. That's a professional obligation regardless of your opinion.

For the Hurricane group specifically: lead with time, not ethics

A teacher who is frightened or hostile will not be moved by a slide about the future of
work. They will be moved by getting 40 minutes of their Sunday back.

Start them on the lowest-risk, highest-relief use cases — which involve **no student data
and no student-facing output**:

• Drafting a parent email about a difficult topic
• Rewriting a rubric at three reading levels
• Generating 20 practice problems at graduated difficulty
• Turning lesson notes into a sub plan
• Drafting the first version of a recommendation letter
• Summarizing a 40-page curriculum document into what changed

Value first. Ethics second. Ethics lands better on someone who has already felt the benefit.

Guardrails: The Non-Negotiables

These are the hard lines. They should appear on a single laminated card in every classroom.

🚫 Never put these into a general-purpose AI tool

• Student names or any identifying detail
• Grades, IEP/504 content, behavior records, health information
• Anything from a student record system
• Parent contact information or family circumstances
• Anything you would not post on the hallway bulletin board

Why: it may be retained, used for training, or exposed. Beyond the privacy harm, it may
place the district outside its obligations under student-privacy law.

Instead: de-identify. *"A 7th grade student who is struggling with fractions and
becomes frustrated quickly"* gets you the same useful answer as a name, with none of the risk.

✅ Always

• Verify every fact, citation, statistic, quote, and link before it reaches a student

or a parent. Fabricated citations look completely real.

• Read every word before you send or post it. Your name is on it; the model's is not.
• Disclose when it matters — to families, in professional work, in anything evaluative.
• Use district-approved tools where they exist. Tools handling student data require a

signed agreement; that is a legal requirement, not a preference.

• Check for bias in anything describing people, cultures, families, or abilities.

Models reproduce the patterns in their training data.

⚠️ Never let AI be the decider

AI does not decide: grades, discipline, placement, referrals, retention, or anything
affecting a student's record or trajectory. It may draft. **A human decides, and the human
is accountable.**

🧠 The Hallucination Rule

**AI is a confident, fluent, well-read colleague who sometimes makes things up
completely and will never tell you which parts.**

It does not know when it is wrong, and it will defend a fabricated citation with the same
tone it uses for a true one. Fluency is not accuracy. Treat every output as a first draft
from a bright intern who does not fact-check.

Implementation: 30 / 60 / 90

A single day of PD changes nothing without a cadence. This is the minimum viable follow-through.

Days 1–30 — Clarity

• Every teacher adds R/Y/G labels to one full unit
• Every teacher runs the September Baseline write with every class
• Guardrail card posted in every classroom
• District publishes the approved-tool list (or states plainly that there isn't one yet)
• Red Team begins logging failures
• Admin: pull completion-timestamp reports from the credit-recovery platform

Days 31–60 — Capacity

• Each department redesigns one major assessment to be process-weighted
• Oral micro-defense piloted in at least one course per department
• Version-history requirement added to major submissions
• Failure Log v1 published district-wide
• Admin: pacing gates enabled; proctored checkpoints scheduled
• Optional 60-minute follow-up clinic — bring a real assignment, leave with it rewritten

Days 61–90 — Community

• Conversation Protocol in use; department chairs collect what's working
• Student-facing AI literacy lesson delivered (library media specialist co-leads —

this connects directly to the K-12 information literacy requirement)

• Families informed: one short letter explaining the traffic lights
• Framework reviewed and revised with staff input
• District: device extension policy and credit-recovery policy reviewed with the Board

Measuring whether it worked

Do not measure "AI incidents caught." That metric rewards surveillance and punishes
the classrooms doing the best work.

Measure instead:

• % of assignments carrying a traffic-light label (target: >80% by day 90)
• % of teachers who ran a September baseline (target: 100%)
• Number of voluntary student AI-disclosure notes (should go UP — that's honesty surfacing)
• Speedrun completions in credit-recovery (should go DOWN)
• Teacher-reported hours saved per week (the retention metric)
• Staff confidence, pre/post survey *(track the Stormy/Hurricane group separately —

movement there is the real signal)*

The Closing Line

**We are not the last generation of teachers before AI.
We are the first generation of teachers who get to decide what it's for.**

*Framework prepared for the Irvington Township School District Department of Education.
Legal and compliance items should be confirmed with district counsel and the district's
data-privacy officer before adoption. See 03-STANDARDS-CROSSWALK.md.*