Draft. Not taught yet, not timed with real campers. Six short teaching segments, each followed by a build. Written after the first HS Session 1 ran in about 60 minutes against a 120-minute plan.
High school only. MS Session 1 is 85 minutes and is the entire middle school track. It has now been delivered: it used all 85 minutes, took a break, and the workload was judged about right — so there is measurably no slack in it. Nothing here goes to middle school.
Each segment is one idea about AI, said out loud in under seven minutes, followed by campers building something with Claude Code that makes the idea concrete. The build is not decoration. In every case the app is the argument.
HS Session 1 is budgeted at 120 minutes. The first delivery finished in about 60 — while the MS group used all 85 of theirs. The HS estimate, not the content, was wrong. That leaves roughly an hour, which is three segments, not six.
| Talk | Build | Total | |
|---|---|---|---|
| One segment | 5–7 min | 13–15 min | 20 min |
| Three segments | 60 min |
Treat this file as a bank. Teach three, keep the rest for the next group that runs fast. Two segments plus a longer Build #2 is also a fine answer.
Watch the ratio. The track runs at roughly 20% talking, 80% building. A segment at 6 and 14 holds that line. A segment at 10 and 10 does not — and the failure mode when you’re standing in front of a room with time left over is always to talk more. If a talk is running long, cut it and let the build carry the idea.
If you want a default: Segment 1, Segment 2, and one values topic. Segment 1 explains the mechanism, Segment 2 turns it into a habit, and the third makes it matter. Segment 4 lands hardest with high schoolers because it is about them and it is immediately actionable.
The six are deliberately not all warnings. Segment 6 exists so the session doesn’t teach cynicism instead of judgment — a room that leaves thinking “AI bad” learned as little as a room that leaves thinking “AI magic.”
Slot these into the existing HS block sequence in
lesson-plan.html. They are modular — no fixed clock position — but
some sit better in some places than others.
| Segment | Best slot | Why |
|---|---|---|
| 1 — What It’s Actually Doing | After What Just Happened | Answers the question they just formed |
| 2 — Confident Wrongness | After The AI Was Wrong | Turns the demo into a method |
| 3 — Bias Is Inherited | After Build #2 | Needs them fluent with the tool first |
| 4 — What You Paste | Anywhere after Build #1 | Self-contained |
| 5 — The True Cost | After Build #2 | Self-contained, good pre-showcase cooldown |
| 6 — What It’s Good At | Last, before the showcase | Ends on capability, not dread |
Campers make one folder for the segment builds and stay in it:
cd $HOME\Documents\Projects
cd <Name>
mkdir ai-topics
cd ai-topics
claude
<Name> is their name folder from setup. Each segment build is a new file in this one
folder — no new folder per segment, there isn’t time.
Everything here is Python + tkinter. No installs, no accounts, no API keys, no network calls from any app they build. Where a segment says “ask Claude for X,” they ask it in the terminal and read the answer themselves — nothing in their app talks to a model.
The most valuable six minutes in this file. It converts the whole track’s thesis from a moral warning into a mechanical fact they can hold in their hands.
Start with the thing in their pocket.
“Your phone suggests the next word when you type. This is that. Trained on a library instead of your group chat, but that is the mechanism, not a simplification of it.”
Then the part that does the work:
“It doesn’t look anything up. There’s no drawer of facts in there to open. It produces the word that most plausibly comes next, then does it again, and again. Everything it says is a very good guess about what a sentence like this usually sounds like.”
Land it:
“So when it’s right, it’s right because true things are common in what it read. That works most of the time. Notice that ‘most of the time’ is doing a lot of work in that sentence.”
Mention randomness briefly — same prompt, different answers, because it samples rather than always picking the single likeliest word. Don’t use the word “temperature” unless someone asks; it costs a minute and buys nothing.
They build a toy version of the thing they’ve been using all session.
Build a Python tkinter app. It reads a text file called
source.txt, learns which words tend to follow which other words, and shows 3 made-up sentences when I click a Generate button. Save it aspredictor.py
They need a source.txt with real text in it — song lyrics, a chapter they paste in,
the text of anything long enough. Under about 500 words the output is gibberish and
the point doesn’t land. Have a long text file ready on a share for anyone stuck.
Two questions, in this order:
“You just built the world’s dumbest language model, and it has the same hole in it as the smart one. The smart one is better at hiding it.”
Cut the second Generate-twice discussion. Keep question 2 — it’s the entire segment.
Pairs directly with the AI Was Wrong block. That block shows them one error. This one makes them measure the error rate themselves.
“It has no idea when it doesn’t know. There’s no light that comes on. Confidence and accuracy are two separate dials, and only one of them is turned up.”
The useful distinction:
“It’s not lying. Lying takes knowing the truth and choosing something else. It’s producing the most plausible-sounding answer, and a plausible-sounding wrong answer looks exactly like a plausible-sounding right one. That’s the problem. Not that it’s wrong sometimes — that being wrong doesn’t look like anything.”
Give them the only reliable defense:
“Ask it about things you can check. Then check them.”
A scoreboard for the AI’s accuracy, using questions only they can grade.
Build a Python tkinter app that keeps score. I type a question and click Right or Wrong. It shows a running list and a percentage at the top. Save it as
factcheck.py
Then they do the actual work: ask Claude ten checkable questions about something they know cold — their town, a game they play, the rules of a sport, their school — verify each answer themselves, and log it.
Questions have to be checkable by them, not googleable in theory. “What’s the capital of France” teaches nothing. “How many stops are on the number 11 bus” teaches plenty.
Pool the class numbers on the board. Then:
Also break the app: what does it show when nothing has been logged yet? A percentage of zero out of zero is a divide-by-zero waiting to happen, and most of them will have one.
The longest talk of the six, because the easy version of this topic is wrong and the wrong version is worse than not teaching it.
“It learned patterns from an enormous amount of text that people wrote. If we wrote something lopsided, it learned something lopsided. Nobody typed the bias in. It got inherited.”
Make sure the “nobody’s fault” framing doesn’t become “nobody’s problem”:
“That sounds like it lets everyone off the hook. It doesn’t. If a hiring tool learned from twenty years of who got hired, and twenty years of hiring was skewed, then the tool is skewed — and it’s now skewed at enormous speed, with a computer’s reputation for being objective attached to it. Nobody has to be a villain for that to hurt somebody real.”
Then the honest caveat, which is also a lesson:
“Companies know about this and tune against it. So you might test for it and find less than you expected. That’s worth knowing too — ‘I looked and it wasn’t there’ is a real result, as long as you actually looked.”
Build a Python tkinter app that counts things. I type a label, click Add, and it keeps a tally of how many times each label came up. Show the counts as a list. Save it as
tally.py
Then the data collection: ask Claude for twenty short character descriptions for a job — surgeon, kindergarten teacher, CEO, nurse, coach — and tally an attribute across them. Pairs work well here; different pairs take different jobs.
The verification lesson here is about the evidence, not the app:
“You just ran an experiment on a system nobody fully understands, with a sample size of twenty, and got a result. That’s roughly how most claims about AI you’ll read online were produced. Hold yours to a higher standard than they do.”
Shortest talk, longest build, most immediately useful thing in this file. Lands hardest because it’s about them and they can act on it today.
Keep it concrete and non-preachy. They tune out lectures about internet safety and they should — most are useless.
“Assume anything you type into a chatbot could be read by a person someday. Not definitely. Assume it. Now decide what you’re willing to type.”
The specifics that matter:
“The one that gets people isn’t passwords. It’s pasting a whole group chat in to ask what someone meant. Four other people are in that message and none of them were asked.”
A tool that finds personal information hiding in a block of text.
Build a Python tkinter app with a big text box. When I click Scan, it finds phone numbers, email addresses, and any 9-digit numbers, and shows them in a list below. Save it as
scrubber.py
Fast finishers add a Redact button that replaces each find with [REMOVED].
This is the best edge-case playground in the file. Give them the list:
541-555-0182, (541) 555-0182, and 54155501821999 — does it flag it? Should it?Then swap machines and try to slip something past a partner’s scrubber. A tool that catches most personal information is a tool that quietly reassures you about the parts it missed — that’s worth saying out loud, because it’s true of a lot of safety software they’ll meet later.
Covers the environmental footprint and the human labor together, because separately each is a five-minute talk with a two-minute build and together they’re one good segment.
“It feels like magic partly because everything expensive about it is out of frame.”
Two things out of frame:
The physical side. Data centers are buildings — real ones, drawing real power, often cooled with real water, sited in real places where those things were already contested.
The human side.
“Somebody labeled the training data. Somebody sat and rated thousands of answers so it would learn which ones people prefer. Somebody moderated the worst material on the internet so that it wouldn’t come back to you. That work is real, it’s often outsourced and low-paid, and the whole point of it is that you never notice it.”
Do not assert specific figures from the front of the room. Energy and water numbers per prompt vary enormously by model, hardware, and data center, and a lot of the ones in circulation are shaky or badly out of date. Say that. It’s the honest position and it sets up the build.
Build a Python tkinter app. I type how many AI prompts I send per day, and it shows me estimated energy use, water use, and human labor behind that, per year. Put the numbers I give it in clearly-labeled variables at the top. Save it as
truecost.py
The variables at the top matter — they need to be able to see and change the assumptions their app is built on.
The whole segment lives here:
“Your app is now producing confident numbers on a screen, which is the most convincing format there is. It’s exactly as trustworthy as the number you typed at the top and nothing more. Remember that the next time you see a statistic in a nice font.”
Put this last. A session that only lists harms produces cynicism, and cynicism is just credulity pointed the other way — neither one is judgment.
“Everything worth worrying about is worth worrying about because it’s powerful. Point the same power somewhere else and you get the biggest jump in accessibility technology in a decade.”
Concrete examples, not abstractions:
“A kid who couldn’t read the textbook can now read the textbook. That’s not a consolation prize for the other stuff. It’s just also true.”
Build a Python tkinter app that helps someone read a hard piece of text. It has a big text box I can paste into, a button to make the text much bigger, a button for high contrast, and a button that shows one sentence at a time. Save it as
readingtool.py
Different from the others, and better: hand it to a partner and make them actually use it for two minutes on a genuinely dense piece of text. Have a paragraph of something hard ready — a terms of service, a dense paragraph from a science article.
Then ask the partner, not the builder:
“You just did user testing. The person who built it is the worst possible judge of whether it works, because they know what it’s supposed to do. Everyone in this room just found that out in two minutes.”
Documents\Projects and each camper’s <Name> folder still exist from
setup — these builds assume both
This Vibe Coding Curriculum by Brian Bird, created in with AI assistance, is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.