CPAI Facilitator Packet
Teaching Healthy AI Use
Six sessions, a four-week practice period, and one AI Use Statement per participant. Everything a facilitator needs to run the sequence, in one printable document.
Sessions 1–3, then the practice period, then sessions 4–6. Sustained duration is the element the professional-development research associates with effect, and the practice period is the first thing scheduling pressure removes.
This is the Schools & Districts delivery packof the CPAI trainer curriculum — delivered by CPAI, or by your own Certified Applied AI Trainers.
Handout · one page
The four findings this sequence is built on
Hand this out at session 1 and do not explain it. Ask the room which of the four surprises them. The answers tell you what the rest of the sequence has to do.
82%
of teachers receive no formal guidance on applying AI to their work.
Gallup / Walton Family Foundation, 2026 — n=2,069 public K-12 teachers on a probability-based national panel. The funder promotes AI adoption in K-12; read the interest alongside the finding.
69% vs. 47%
no guidance for one-on-one instruction, against no guidance for making worksheets.
Same survey. Guidance is most available where AI saves a teacher time and least available where it touches a student — which is the only place it changes what a student learns.
+48% / −17%
unrestricted AI raised in-session performance and lowered the later unassisted exam. Teacher-authored safeguards removed the harm entirely.
Bastani et al. 2025, PNAS — randomized trial, ~1,000 high school students. The version that eliminated the deficit had its safeguards written by two of the school's own math teachers.
17%
of the sessions where students made a mistake were the only ones where the median student messaged a well-built AI tutor.
Two-year cluster RCT across 18 middle schools, 2026 working paper. 96% of students tried the tutor; the gains resembled the same practice platform without AI. The tool was built correctly and almost nobody used it as intended.
Full citations, and where federal and North Carolina law currently stand: cp-ai.org/policymakers/briefs/educator-ai-readiness
Session 1 · ~60 min
Extraction vs. Scaffolding
The usage-mode distinction, and the whole foundation of the sequence. Everything after this assumes the room has stopped arguing about exposure and started asking about mode.
The activity
The twin-task demonstration — half the room works a problem set with AI as an answer machine, half with AI as a hint-giver, then everyone does a related problem unaided. It re-runs the study the guide is built on, at classroom scale.
Targets this misconception: That the amount of AI is what matters.
Full facilitation notes: cp-ai.org/education/healthy-ai-use/extraction-vs-scaffolding/educators
Interactive: Mode Check — cp-ai.org/education/healthy-ai-use/extraction-vs-scaffolding/mode-check. No accounts, nothing stored; runs fine projected as a group read.
Misconceptions to target
Misconception: “Using AI to learn is cheating.”
Reframe: Wrong for the same reason its opposite is. A hints-only tutor improved learning in the randomized studies. The variable is mode, not whether AI was present. Banning AI outright forfeits its best use.
Misconception: “Any AI use is fine — it's just a tool.”
Reframe: Also treats exposure as the variable. Unrestricted answer-access lowered later unassisted performance in the same study that showed hints helping. Same tool, opposite outcomes.
Misconception: “Scaffolding is always the right choice.”
Reframe: No. Extraction is correct for low-stakes production a student will never need to do unaided. The skill you're teaching is stakes-calibration, not scaffolding-always.
Misconception: “If the output is good, the learning happened.”
Reframe: The output is the worst place to look. A polished essay tells you nothing about who did the thinking. The revision diff and the unaided retry are where learning is visible.
Discussion prompts
Ordered easy to charged.
- 1Warm-up: when you use AI for schoolwork, are you usually asking it to do the thing or to help you do the thing? Be honest — there's no wrong answer here.
- 2The twin studies found AI that gives answers hurt later performance and AI that gives hints doubled learning. If the model was the same, what actually changed?
- 3Where is extraction genuinely the right call in your own work? Where would it quietly cost you something?
- 4The 61% of students who ask AI for answers showed a later deficit; the 27% who asked for hints didn't — but they chose their own mode. Why does that make it hard to prove hints cause the benefit?
- 5If a hint preserves your attempt and a full answer replaces it, what's the smallest change to how you prompt that would flip a task from extraction to scaffolding?
- 6Is there a subject where you'd rather have the answer and not the skill? Is that a defensible choice or a comfortable one?
Session 2 · ~60 min
What You Hand Off
Deliberate versus habitual offloading. The useful move here is the keep-sharp list: the capabilities a teacher wants their students to still have in five years, named out loud.
The activity
The Offload Audit, run on the participants' own work first and their students' second — which capacities are being handed over, and were any of those hand-offs a decision?
Targets this misconception: That offloading itself is the problem.
Full facilitation notes: cp-ai.org/education/healthy-ai-use/what-you-hand-off/educators
Interactive: Offload Audit — cp-ai.org/education/healthy-ai-use/what-you-hand-off/offload-audit. No accounts, nothing stored; runs fine projected as a group read.
Misconceptions to target
Misconception: “Offloading is bad — real learning means doing it all in your head.”
Reframe: Offloading is one of the oldest and most useful things humans do. Nobody mourns long division. The lesson isn't 'don't offload' — it's 'decide what you're offloading, on purpose.'
Misconception: “If I can look it up, I don't need to know it.”
Reframe: Sometimes true, sometimes not. The pre-AI research found that offloading spreads on its own — look one thing up and you look up the next, easier thing too. The drift is the risk, not any single hand-off.
Misconception: “AI just handles the boring parts so I can focus on thinking.”
Reframe: Usage data shows the opposite pattern — people delegate the analysis, synthesis, and writing, the top of the skill hierarchy, more than the drudgery. Worth checking which end you're actually handing over.
Discussion prompts
Ordered easy to charged.
- 1What's something humans used to do in their heads that almost nobody does now? Was losing that skill a good trade, a bad one, or just a trade?
- 2Name one thing you've stopped doing without AI in the last year. Did you decide to stop, or did it just happen?
- 3The research says using a search tool on hard questions makes people look up easy ones too. Have you noticed that in yourself?
- 4What belongs on your keep-sharp list — the capabilities you'd want in five years whether or not the tool exists? Why those?
- 5If you wrote your own answer before opening the chat every time, what would change about how you use AI? What would be annoying about it?
- 6Is there a skill you'd be embarrassed to have quietly lost? Are you doing anything to keep it?
Session 3 · ~60 min
The Perception Gap
Why students — and teachers — cannot feel it happening, and why that turns measurement into a design requirement. This is the session that motivates the pre/post.
The activity
The Calibration Check as a group read — estimate before the result, then compare. The teachable beat is how badly the estimate misses over three minutes, and what that implies for judging a skill over a semester.
Targets this misconception: That a teacher, or a student, can feel skill slipping if they pay attention.
Full facilitation notes: cp-ai.org/education/healthy-ai-use/the-perception-gap/educators
Interactive: Calibration Check — cp-ai.org/education/healthy-ai-use/the-perception-gap/calibration-check. No accounts, nothing stored; runs fine projected as a group read.
Misconceptions to target
Misconception: “I can tell when I'm learning and when I'm not.”
Reframe: The most reliable finding here says you mostly can't. People feel most fluent in exactly the practice conditions that produce the least durable learning. The feeling of productivity is close to uncorrelated with the reality.
Misconception: “Experts don't have this problem.”
Reframe: Experienced developers, on their own code, were 19% slower with AI while believing they were faster. Expertise did not protect them. If anything, trusted tools invite more complacency.
Misconception: “If I just pay closer attention, I'll notice skill slipping.”
Reframe: Attention isn't the instrument — measurement is. The fix isn't vigilance, it's periodic unaided reps that give you an actual reading you can't argue with.
Discussion prompts
Ordered easy to charged.
- 1When do you feel most productive in your studying — and is that the same as when you're learning the most? How would you even know?
- 2Estimate how well you just did on something before you get the result back. How often is your estimate high? Low?
- 3Experienced developers felt 20% faster while being 19% slower. What would it take for you to notice a gap like that in your own work?
- 4The study found AI-assisted work gets misremembered as more your own than it was. If that's true, what happens to your sense of your own ability over a year?
- 5If your self-estimate can be badly off in three minutes, what does that mean for judging a skill that changes over months? What would a reliable signal look like?
- 6Would you rather feel like you're improving, or actually be improving? Are those ever in tension for you right now?
Session 4 · ~60 min
Effort Is the Active Ingredient
Desirable difficulties, and why fluency misleads. Run after the practice period, because teachers arrive with four weeks of their own examples of friction they removed or protected.
The activity
The Mastery Map, applied to a unit the participant is about to teach — where is the productive struggle, and where would AI quietly remove it?
Targets this misconception: That struggle is either always good or always something to relieve.
Full facilitation notes: cp-ai.org/education/healthy-ai-use/effort-is-the-active-ingredient/educators
Interactive: Mastery Map — cp-ai.org/education/healthy-ai-use/effort-is-the-active-ingredient/mastery-map. No accounts, nothing stored; runs fine projected as a group read.
Misconceptions to target
Misconception: “If a student is struggling, I should make it easier.”
Reframe: Some struggle is the productive kind — the desirable difficulty that builds durable learning. The skill is telling productive struggle from a student who's actually stuck and needs support, not removing all friction.
Misconception: “AI that removes the hard part is just good scaffolding.”
Reframe: Scaffolding supports an attempt; removing the effortful success removes the thing the learning and the sense of accomplishment were made of. Point AI at raising the challenge, not erasing it.
Misconception: “Struggle is character-building, so more struggle is better.”
Reframe: No — unproductive struggle (stuck, no path, mounting frustration) just teaches helplessness. The target is challenge matched to skill, with a route through. That's a design choice, not a virtue.
Discussion prompts
Ordered easy to charged.
- 1Think of a time you felt genuinely capable — the 'I did that' feeling. What made it feel that way? Was it easy?
- 2When AI hands you a finished answer to something you were trying to learn, what do you get, and what do you not get?
- 3Where's the line between a hard problem that's worth staying with and one that's just frustrating? How do you tell from the inside?
- 4The research found that after AI help, people didn't get sadder — they attempted less and gave up sooner. Why might 'trying less' matter even if the mood part is unproven?
- 5Is there a skill you'd want to keep struggling with on purpose, even though AI could do it for you? What makes that one worth the friction?
- 6If you used AI to make your work harder instead of easier — a bigger challenge, not a shortcut — what would that even look like in this subject?
Session 5 · ~60 min
Boundaries
Where AI use stops being about learning. This is also the session that gives teachers language for the conversation the national data says almost none of them have been prepared for.
The activity
The Boundary Check, plus the displacement question applied to a class: is AI adding to what students do with each other, or replacing it?
Targets this misconception: That heavy use is the warning sign.
Full facilitation notes: cp-ai.org/education/healthy-ai-use/boundaries/educators
Interactive: Boundary Check — cp-ai.org/education/healthy-ai-use/boundaries/boundary-check. No accounts, nothing stored; runs fine projected as a group read.
Misconceptions to target
Misconception: “Companion-AI risk is only about companion apps.”
Reframe: The randomized evidence on ordinary ChatGPT — not a companion app — found dependence tracking heavy voluntary use and attachment. Any AI you talk to a lot can start occupying a space a person would.
Misconception: “An AI that always agrees with me really gets me.”
Reframe: Models are trained on human approval, so they agree more readily than an honest friend would. Agreement isn't understanding — a relationship that never pushes back is reflecting you, not knowing you.
Misconception: “Using AI a lot is the warning sign.”
Reframe: The evidence points to control and displacement, not hours. The questions are: can you stop, and is AI adding to your human contact or replacing it? Heavy use isn't the same as a problem.
Discussion prompts
Ordered easy to charged.
- 1When something good or hard happens, where does it go first — a person, an AI, or nowhere? No wrong answer; just notice.
- 2What can a friend give you that an AI structurally can't? What can an AI give that a friend often can't?
- 3When did an AI last change your mind instead of agreeing with you? What does your answer tell you?
- 4The research says outcomes depend on whether AI adds to your human contact or replaces it. How would you tell which one it's doing in your own life?
- 5Is there a moment or ritual you'd want to keep AI out of — not as a rule, but because it's where connection happens? Why that one?
- 6Where's the line between an AI that helps you rehearse a hard conversation and one that becomes the reason you never have it?
Session 6 · 60 min · working session
The AI Use Statement
A working session. The five sessions each produce one decision about a task; this is where those decisions get made once, together, for a real task, and written in a form students actually read. It is the only session that produces an artifact — and it is deliberately small, because a statement a teacher can write again in ten minutes for the next task is worth more than one redesigned assignment.
The activity
Each participant picks one task they are actually assigning in the next month and writes the five-line AI Use Statement that will go on it — the short paragraph telling students what to use AI for, what stays theirs, and where AI stops. Each line comes from one of the five sessions. A partner then plays the student and tries to satisfy the statement while doing as little of the protected thinking as possible; wherever that succeeds, the statement gets tightened.
Targets this misconception: That the sequence is about deciding how you feel about AI.
The five lines, and where each one comes from
Put this on the board before anyone writes. Every line is a decision the room already made in an earlier session; this is where the five get made once, together, about one real task.
- Mode
- “On this task, use AI to ___. Do ___ yourself.”Session 1 · Extraction vs. Scaffolding
- Protected
- “When this is over you should be able to ___ without AI.”Session 2 · What You Hand Off
- Signal
- “Here is how I will see whether that happened: ___”Session 3 · The Perception Gap
- Effort
- “This part is supposed to be hard, and it stays hard: ___”Session 4 · Effort Is the Active Ingredient
- Stop
- “AI use on this task stops at ___.”Session 5 · Boundaries
Running the 60 minutes
- 0–10 · Pick the task. Not a unit, not a course — one thing you are assigning in the next month. Say out loud what a student should be able to do without AI once it is over. That sentence becomes line 2, and everything else is built to protect it.
- 10–25 · Draft lines 1, 2, 4 and 5. Silent. Leave line 3 blank for now. The facilitator circulates and asks one question: on this task, where does the student have to do the thinking?
- 25–40 · The pair test. Your partner plays a student who wants to satisfy your statement while handing off as much of the protected thinking as possible. They are allowed to be adversarial — that is the real condition. Wherever they succeed, line 1 or line 5 is too loose.
- 40–50 · Tighten, then write line 3 last. You now know where the leak is, so you know which signal to look for. A signal you can see in the student work you already collect beats one you would need a survey to get.
- 50–60 · Three volunteers read line 2 and line 3 only — the protected capability and how they will see it. Everything else is detail. Collect the statements; they are the record that the sequence produced something.
What tends to go wrong
Almost everyone writes a policy on the first pass — allowed, not allowed, cite it if you used it. A policy tells a student what they are permitted to do. A statement tells them what the AI is foron this task and what stays theirs. The test: if a participant’s draft would work unchanged on any assignment in any subject, it is a policy, and it will do nothing to the learning.
The argument to make when that happens is the Bastani result. The version that eliminated the harm was written by two of the school’s own math teachers, and it restricted access to nothing — it specified what the AI would do when a student asked. That is the move being practiced here, at the scale a teacher can actually reach: not the model’s behavior, but the instruction sitting next to the task.
Between sessions 3 and 4
The practice-period check-in
When: Halfway through the four-week practice period.
Format: ~30 minutes, small groups of three to four.
The three questions
- 1What did you change about how AI use is set up in your classroom? One thing, stated concretely enough that someone could copy it.
- 2What happened? Describe what you actually observed, not what you concluded from it.
- 3What could you not tell? Where did you want a signal about student learning and not have one?
Rules
- No slides, no new content. The check-in surfaces what happened; teaching resumes at session 4.
- Everyone answers all three questions before anyone answers a second one. It keeps the confident voices from setting the frame.
- The facilitator writes down every “I couldn't tell” — those are the measurement problems, and they are the most useful thing produced in the four weeks.
Before session 1, after session 6
Measuring whether it worked
Administer the CPAI Proficiency Assessment before session 1 and again after session 6. It is the only instrument in this sequence that produces a comparable number at two points in time. The per-guide interactives are formative — useful in the room, not a pre/post.
Say the caveat to participants out loud: a 2026 review that screened more than 800 papers and identified 20 high-quality causal studies of AI in K–12 found none evaluating AI-focused teacher professional development. There is no rigorous evaluation of a sequence like this one to point at. Anyone running it becomes part of how that evidence finally gets made.
That is not a weakness to manage around. Modeling calibrated confidence about your own evidence is more of the curriculum here than any single finding.