Brainstorm: What This Training Actually Has To Solve

Working notes. Read this before the framework — it explains why the framework looks the way it does.

1. The problem statement contains three different problems

People talk about "AI in schools" as one issue. It isn't. Irvington's situation is three
problems that happen to share a trigger, and each needs a different move.

The single most important reframe for the whole day:

AI did not create the 6-hours-for-6-months problem. It revealed it.
If a course could be completed in 6 hours by a machine, the course was already
measuring compliance rather than learning. AI just made that visible and cheap.

That sentence is the pivot of the entire training. It moves the room from
"how do we stop students" to "what were we actually measuring" — which is a
question teachers can answer and control.

2. The hardest audience problem: the room is not on one page

A show of hands in any district PD on AI splits roughly four ways:

• Sunny — already using it, wants advanced material, will be bored by "what is a chatbot"
• Partly cloudy — curious, tried it twice, unsure what's allowed
• Stormy — thinks it is harmful to students and to the profession; often well-informed
• Hurricane — fears replacement, de-skilling, or simply doesn't want another initiative

The mistake almost every district makes

Treating Stormy/Hurricane as a compliance problem to be overcome with enthusiasm.
It backfires. Those teachers dig in, and they are frequently the most senior and most
influential voices in the building.

The move that works: make the skeptics the Red Team

Give the critical teachers the formal role of trying to break the AI — find the
hallucinations, the bias, the privacy leaks, the places it fails students. Then
publish their findings as a district asset.

Why this works:

• It requires zero attitude change. They can stay skeptical and still participate fully.
• Their skepticism becomes expertise rather than resistance.
• It is genuinely the most valuable work in the room. Guardrails come from people who

look for failure, not from people who are delighted.

• It publicly signals that "this is not a sales pitch," which buys credibility with the

entire middle of the room.

Nobody is asked to like AI. Everyone is asked to be competent about it.
That is the only ask the day makes, and it should be said out loud in the first five minutes.

3. Why detection-first strategies fail (and what to do instead)

Teachers will arrive wanting one thing: a reliable detector. The honest answer is that
one does not exist, and the training has to deliver that news early and then immediately
replace it with something that actually works — otherwise the room feels disarmed.

The problems with detector-first enforcement:

• Detector scores are probabilistic, not evidentiary. They produce false positives.
• False positives fall hardest on **multilingual learners and students who write in a

plain, formulaic register** — which in Irvington means the accusation risk is not
evenly distributed across the student body. This is an equity issue, not just an
accuracy issue.

• "Humanizer" sites defeat detectors in one click, so enforcement escalates into an arms

race the school always loses.

• One wrong accusation costs more trust than ten caught cases recover.

The replacement: process evidence beats product evidence.
A detector score is the weakest thing in the evidence stack. The strongest things are
free: an in-class writing baseline, document version history, a two-minute oral check,
and a draft the student can talk about.

This is the section where a skeptical room becomes an engaged room, because it gives them
something real to hold.

4. The "6 months in 6 hours" autopsy

This deserves its own module because it is the district's actual pain.

What's happening mechanically:

• Self-paced / credit-recovery platforms with no pacing gates
• Auto-graded item banks with fixed questions
• Product-only submission: no drafts, no process, no checkpoints
• Extensions and proxy sites that auto-answer inside the platform
• Seat-time policies that measure minutes logged rather than evidence of learning

Who can fix what — this matters, and the training must be honest about it:

Showing teachers this table prevents the most common PD failure mode: giving people
strategies they are not empowered to implement, then blaming them for not implementing them.
It also gives the Superintendent's office a clean list of what only they can unblock.

5. What makes it fun rather than another sit-and-get

Rules I'm designing against:

1. No slide goes more than 6 minutes without the room doing something.
2. Competition beats contemplation. Teachers are competitive. Scored, timed

activities with a leaderboard get more engagement than reflection prompts.

1. Let them beat the AI. The most satisfying moment in any AI PD is watching the

machine be confidently, stupidly wrong. Build that in deliberately and early.

1. Give back time in the first hour. A teacher who saves 40 minutes on parent emails

before lunch will stay engaged after lunch. Value first, ethics second — ethics lands
better on someone who has already felt the benefit.

1. Everything they build, they keep. Every activity outputs a copy-pasteable artifact

for their own course. No one leaves with only notes.

1. Anonymous by default. Nobody's wrong answer gets attributed. Especially important

for the skeptics and for teachers who feel behind.

Activity concepts (built into the portal)

The deliberate trap: "Spot the Bot" is rigged to humble everyone.
The samples are chosen so that confident teachers score around 50%. The debrief reveals
that the cues people trust — "too polished," "weird transitions," "no personality" — are
the same cues that describe a nervous multilingual writer. That single activity does more
to prevent a false accusation than any policy slide.

6. Format decisions

Why a self-contained HTML portal instead of Google Slides / Nearpod / Kahoot:

• No login, no accounts, no student data, no vendor. We are about to teach a session

that includes student-privacy law. Using a tool that harvests attendee data would be
an own-goal. Everything runs locally in the browser; nothing is transmitted.

• No IT ticket. Opens by double-click from a shared drive or a USB stick. Works if

the district Wi-Fi dies mid-session — which it will.

• Survives the session. Teachers keep the file. It becomes the year-round reference,

not a deck they never reopen.

• Free and forkable. Other Essex County districts can use it. No per-seat licensing.

Trade-off accepted: no cross-room live aggregation (no shared leaderboard across
devices). Mitigation: the facilitator collects counts by show-of-hands or a one-line
tally on the board, which is faster than a QR code anyway and keeps eyes up.

7. What I deliberately left out

• Specific product recommendations. Tool choice is a procurement decision with

privacy-contract implications under P.L. 2019, c.494. The training teaches evaluation
criteria instead of naming winners.

• A district AI policy. Drafting binding policy is the Board's job. What's here is a

classroom-level framework plus a checklist the Board can hand to counsel.

• "AI will replace teachers" debate. It's unwinnable in a 6-hour session and it

poisons the room. Addressed once, briefly and honestly, then set aside.

• Cheating statistics. Widely cited numbers in this space are unreliable and

contested. Better to use the district's own data than borrowed alarm.

8. Open questions for the district before delivery

These change the material, so answer them first:

1. How long is the session? Full day, half day, or a series of after-school blocks?

(The framework is built modular so it can be cut to 90 min / 3 hr / 6 hr.)

1. Which grade bands are in the room? K-5 and 9-12 need different examples; the

portal content is currently written for 6-12 with K-5 notes.

1. Is there an approved AI tool with a signed data agreement? If yes, name it in the

deck. If no, teachers must be told explicitly not to put student data into anything.

1. What is the current credit-recovery platform, and does it support pacing gates

and randomized banks? This determines whether the speedrun fixes are available.

1. Does the district want PD hours attached? If yes, the portal's certificate needs

the district PD code and an administrator signature line.

1. Who owns follow-up? A single day changes nothing without the 30/60/90 cadence.