LPS JOINT COUNCIL · AUGUST 18, 2026
Think First
Partner with AI. Don't outsource your thinking.
PRE-POLL QR
(paste here)
90 seconds · anonymous · your phone
Russ Wilcox · pro bono, at the invitation of Dr. Julie Hackett
EXERCISE · PENCILS ONLY · ANONYMOUS
Estimate the total number of words a Lexington student writes by hand, K through graduation.
GUT
PLAN
WORK
ANSWER
FLIP
STEP BACK
There is no right answer. Order of magnitude is the game. Cross-outs are data — leave them.
9:00
SEND IT IN
One scanner per table. Youngest person scans.
UPLOAD QR
(paste here)
Front, then back. Whole page in frame.�
While you wait:
compare final estimates at your table.
Just among yourselves.
WHAT JUST HAPPENED
10:1?
100:1?
A room of experts. One question about the system you run. How far apart did this room land? The mirror knows — we'll see shortly.
When experts disagree on a question like this, it's rarely the arithmetic.
It's the definitions nobody wrote down.
Fermi-problem research scores by order of magnitude for exactly this reason — and every chatbot prompt has the same silent-definition problem.
THE EVIDENCE · 1 OF 4
The pen lights the brain up. The keyboard doesn't.
HANDWRITING
TYPING
256-channel EEG: writing by hand produced widespread theta/alpha connectivity — the brain's encoding signature. Typing did not.��"We urge that children be exposed to handwriting from an early age."
Van der Weel & Van der Meer (2024), Frontiers in Psychology — high-density EEG, N=36. Methodological commentary: Pinet & Longcamp (2025).
THE EVIDENCE · 2 OF 4
Transcribing is not thinking.
Laptop note-takers transcribe.
Longhand note-takers reframe.
Reframing is learning.
“The Pen Is Mightier Than the Keyboard”
Even with no distractions, laptop users performed worse on conceptual questions — because typing enables verbatim capture, and verbatim capture bypasses processing.��Generative AI performs the same bypass — at industrial scale.
Mueller & Oppenheimer (2014), Psychological Science. Replication: Morehead, Dunlosky & Rawson (2019) — effects smaller but directionally consistent.
THE EVIDENCE · 3 OF 4
Your brain on ChatGPT
Darker bar = stronger neural engagement · values are a relative index (100 = Brain-only), illustrating the study's reported ordering
Essay writing across three groups. Neural coupling scaled DOWN with external support.��LLM users couldn't quote their own essays minutes later. Ownership collapsed.��“Cognitive debt.”
Kosmyna et al. (2025), MIT Media Lab, arXiv:2506.08872 — EEG, N=54, preprint. Critique noted: Stanković et al. (2026). Values shown are illustrative of the reported ordering.
THE EVIDENCE · 4 OF 4 · THE PART NOBODY QUOTES
The group that thought FIRST —
then used AI —
showed HIGHER recall and re-engaged neural networks.
“Think First” isn't a slogan. It's an EEG finding. And you just lived it.
Kosmyna et al. (2025), session 4: Brain-to-LLM vs. LLM-to-Brain crossover.
CALIBRATION CHECK · WHAT YOU'VE BEEN TOLD
Two stories are being sold about AI in schools. Both are wrong.
THE BOOM
“AI will personalize everything — adopt now or fall behind.”
Reality: national surveys show use racing ahead of guidance; district plans chase efficiency — lesson plans, workload — without a learning vision; and almost nothing on the market is validated against learning outcomes.
THE DOOM
“Scientists proved ChatGPT is rotting kids' brains.”
Reality: the EEG study is a preprint of 54 adults writing essays — it measured engagement during a task, not brain damage. Its own fourth session showed the think-first group IMPROVING with AI. “Cognitive debt” is a metaphor, not a lesion.
What the evidence actually supports is neither panic nor party. It supports DESIGN — because the variable that decides the outcome is structure and sequence, and structure and sequence are exactly what a district controls.
Boom: RAND (2024–25) — “AI use quickly increasing but guidance lags”; AIR (2025) district-plans analysis. Doom: Kosmyna et al. (2025) read in full, incl. session 4; critique: Stanković et al. (2026).
NAMING THE RISK
The risk isn't wrong answers. It's outsourced thinking.
Cognitive offloading
Unstructured AI use consistently reduced argument quality; offloading was the measured mechanism.
Gerlich (2025), Societies
Metacognitive laziness
Learners handed the self-checking — reviewing criteria, judging drafts — to the tool, and self-regulation atrophied.
Fan et al. (2025), Brit. J. Ed. Tech.
思维惰性 · thinking inertia
China's Ministry of Education named the same risk in national policy — and regulates it.
MOE Usage Guide (2025)
THE STORY · TWO STUDENTS
Two ninth graders. Same morning.
BEIJING
AI literacy is on her schedule — 8 hours a year, mandated.
Her essay is handwritten; a smart pen digitizes every stroke; AI feedback arrives by afternoon.
When she uses AI, she documents it — the rule, not a confession.
The chatbot she uses is legally barred from making her dependent.
MASSACHUSETTS
His district has no AI policy. His teacher decides alone.
AI is banned in period 2, required in period 5, unmentioned in period 6.
At home: unsupervised ChatGPT. Homework in seconds. Nobody knows.
No law anywhere obliges the chatbot to protect his thinking.
The difference isn't the students.
THE STORY · THE DOCUMENTS BEHIND THEIR DAYS
The difference is the documents.
应当 · “shall”
Two national guides, published the same day — May 12, 2025. One mandates AI literacy. One restricts AI use. 1,400+ Beijing schools. Elementary open-generation: banned.
“choose to”
“The Department is not recommending or requiring the use of AI in schools — those decisions remain with local leaders.”
One system wrote “shall.” The other wrote “if you choose to.” Both sentences are now shaping a nine-year-old's mind.
PRC MOE: 《中小学生成式人工智能使用指南》 & 《通识教育指南》 (May 12, 2025) · Beijing Work Plan (2025–27) · Mass. DESE, Guidance for AI in K–12 Education (2025)
THE STORY · THE INTEGRITY SPLIT
One system documents. The other detects.
BEIJING: DOCUMENTATION
Students record how they used AI — which tool, for what, at which step. Transparency is the assignment. Trust is the default; the process is the evidence.
AMERICA: DETECTION
We buy detectors that don't work, accuse students we can't prove, and teach an arms race instead of a discipline. Suspicion is the default; the product is the evidence.
The worksheet you filled out today is the third way: make the thinking visible, and integrity takes care of itself.
WHERE MASSACHUSETTS STANDS
“The Department is not recommending or requiring the use of AI in schools — those decisions remain with local leaders.”
Massachusetts DESE, Guidance for AI in K–12 Education (2025)
“Without clear guidance… educators are left to their own perspective… students [have] drastically different experiences from classroom to classroom.”
DESE AI Task Force Report
The strongest AI framework in Massachusetts will be written by whichever district writes it.
This room can be that district.
While we talked,
an AI read your handwriting.
Let's see what it found.
THE REFRAME · WHAT YOU JUST WATCHED
That wasn't a chatbot.
ASK-AND-GET · generative
You ask. It answers.
The thinking travels out with the question — and the answer expires with the chat.
This is the only AI most people have ever met.
TASK-BASED · extraction
You assign: transcribe this pencil. Find the assumptions. Locate the misconception. Cite the evidence for every claim.
It reads work YOU made — and never writes a word of your thinking.
Same technology. Opposite relationship to your mind.
WHY EXTRACTION CHANGES EVERYTHING
Answers expire. Structure compounds.
K
2
4
6
8
10
12
every ordinary assignment adds a point — no extra tests, no screens
a record of how a student thinks — growing richer every year, traveling teacher to teacher
The chat log is a moment. The structure is a career.
WHY YOU CAN TRUST IT
Every number, three clicks from the pencil.
SCORE
“Adaptivity: 82”
→
TRANSCRIPT
“wait, high school is mostly typed now”
→
PHOTOGRAPH
the actual graphite
If the AI can't cite the pencil, it can't make the claim. Fabricated evidence is dropped automatically.
ONE STUDENT, OVER TIME
One student. One semester. Every breakthrough on record.
27
hours of voluntary
Socratic dialogue
162
exchanges — the AI never
once gave the answer
31
documented breakthroughs
(score jumps after struggle)
11:14 PM
when it happened —
no human tutor available
Five minutes in: “this actually makes sense.” Molecular polarity: 30 → 100 in ninety minutes. And he came back after the test was over.
He wasn't lazy. The method was wrong — and a method can be fixed.
WHAT THE PENCIL BECOMES
Unstructured thinking in. Structured intelligence out.
IN: THE PENCIL
free-text reasoning
cross-outs and margin notes
assumptions and becauses
confidence circles
the assumption you'd bet against
→
OUT: THE PROFILE
7 reasoning axes, scored
misconceptions, named + located
revision behavior, counted
calibration, tracked
every claim linked to the pencil
→
OVER TIME
growth curves per axis
persistent vs. resolved gaps
any subject, same schema
follows the student
teacher to teacher
No test was taken. No screen was used. The evidence is schoolwork that was happening anyway.
THE QUESTION THIS RAISES
TODAY
90 educators
9 minutes of pencil
one snapshot of a room's mind
A STUDENT
the same axes, every month
K through graduation
a thinking profile that travels
teacher to teacher, year to year
Not a test score that follows them. A record of how they think — built from their own pencil.
What would September look like?
Today you met three students. The first one was you.
AI reads the work.
Never the child.
Your turn to be counted.
POST-POLL QR
(paste here)
90 seconds. Anonymous.�Your votes seed Lexington's AI governance framework — built ground-up, starting from this room.��Your group's report is on your phone now — same QR, tap your group number. Live for 48 hours.
Think First · Partner with AI, Don't Outsource Your Thinking
Russ Wilcox · with gratitude to Dr. Julie Hackett and the LPS Joint Council