AI in schools / AI preparedness
Get ready to work with AI
Four modules, no math or coding needed: how AI works, how companies use it, how to direct AI agents, and how to use AI authentically.
Module 1
How AI works
Every AI system, from a spam filter to ChatGPT, is built from one simple idea: find the line that fits the data.
Try it: the simplest machine learning models find trendlines
Seven short steps. Click “Next” to go through them.
Draw a line through the dots
Drag across the plot from left to right, where you think the line belongs.
When you drew your line, you were doing in your head what a computer does with arithmetic: picking the numbers that make the line miss the dots by as little as possible. For a straight line that’s two numbers. For a curve, you add pieces: each one is a line that only switches on past a certain point.
That’s a neural network. Each “neuron” multiplies, adds, and bends. Stack enough of them and the network can bend to fit almost any shape. “Training” just means nudging all those numbers, over and over, until the misses get small.
From curves to letters
Now swap the dots for text. Instead of “given x, where is y?”, the question becomes “given the letters so far, which letter comes next?” The network gives every letter a score, and the highest score wins.
This model learned from one short paragraph about teaching. It looks at the last few letters and picks the letter that most often came next in what it read. Try “the teach” or “a good q”. It can only echo what it read, and it has no idea what any word means.
Learning without an answer key
The dots in the curve lab came with answers: every x had its y, and the computer learned to match them. That’s called supervised learning, and it needs someone to supply the answers. Chatbots learn mostly without one. Hide the next word of any sentence, and the sentence itself is the answer key. Every sentence ever written becomes practice, with no one grading.
To get good at guessing the next word, the model has to pick up what words mean. Nobody tells it that “teacher” and “student” are related; it notices they show up in the same kinds of sentences. Inside the model, each word becomes a list of numbers, like a location on a map, and words used the same way end up near each other.
From autocomplete to assistant
A model that has only read the internet is a very well-read autocomplete. Ask it a question and it might continue with more questions, because that’s what the text it read often did. Two more stages of training turn it into an assistant.
Why it makes things up: every stage rewards answers that look right. When the model doesn’t know something, the most likely-looking answer is still a confident one. That’s why it can invent a quote, a source, or a date, and why checking its work matters.
Why early AI models made things up
Early chatbots were built to produce the next likely word, not to look anything up. Ask one for sources and it wrote text shaped like a citation: a plausible author, a plausible title, a plausible year. Nothing in training checked whether the paper existed. In one 2023 test, most of the research references Bard gave were made up or badly wrong, and so were more than a quarter of GPT-4’s (Chelli et al., 2024).
Labs have since changed what they reward. Assistants now search the web and quote the pages they found, and training increasingly penalizes confident wrong answers more than an honest “I don’t know.” On one independent test of questions models often can’t answer, OpenAI’s GPT-6 Astra guessed wrong instead of admitting it didn’t know 51% of the time, down from 92% for the model before it (Artificial Analysis, 2026).
Better, not solved. Models still make things up, especially about niche topics or without search turned on. Check any source before you rely on it.
Now make it enormous
Chatbots work the same way: predict the next piece of text, add it, repeat. The difference is size. Trained on a huge share of everything ever written, with billions of numbers to tune, the network has to model grammar, facts, and lines of reasoning to make good guesses. That’s where the abilities come from: not from rules anyone wrote, but from these same building blocks at a scale that’s hard to picture.
The curve you just fit
The letter predictor above
GPT-2 (2019)
GPT-3 (2020)
Today’s largest models
Each step on this chart is many times the last: the bars use a log scale, or the first two would be invisible.
Module 2
How AI is changing work
Here’s what AI is actually doing in companies right now, and what’s still left for people.
What AI does at work today
Most of it is busywork: summarizing, data entry, first drafts. The bigger shift is that teams with no engineers, like a legal team that used to copy answers from one form to another by hand, can now build their own software. Complex expert work is next, and still experimental.
- 1
Busywork
Summaries, data entry, first drafts
Working now
- 2
Software for everyone
Non-engineers build their own tools
Working now
- 3
Complex expert tasks
Drafting a patent, a legal brief
Early
- 4
Judgment and taste
Knowing what’s worth doing
Not yet
- 5
The physical world
Hands, robots, real objects
Hardest
The hard parts are still ahead. One is judgment: the intuition and taste to know what’s worth doing, which people still supply. Intelligence is almost solved, judgment is not.
The other is the physical world. Robots can plan what to do; the hard part is control. A human hand has 27 bones and makes constant tiny corrections to catch a falling ball. Copying that turns out to be extremely hard.
Will AI take jobs or create them?
In the 1860s, economist William Stanley Jevons noticed that more efficient steam engines made Britain burn more coal, not less: when something gets cheaper, people use more of it (Jevons paradox). Automation often works the same way.
The Waymo effect. Self-driving cars were supposed to remove drivers. In the short run they created new human work: cleaning cars, charging them, helping remotely when one gets stuck. Waymo even pays DoorDash drivers to close doors that riders leave open, one offer was $11.25 (Bloomberg, 2026). Jobs that look easy to automate often aren’t yet.
But some jobs are shrinking. Call centers were the obvious job for AI to replace, and for years they kept growing. Now, finally, hiring has slowed: new hires per 100 workers fell from 25.1 to 17.2 (Revelio Labs).
Who wins right now
Models are trained on examples, so they handle common work well and the rarest, best work poorly. Work in the top 1% is out of reach for them, and the people who do it are becoming more valuable: they can now manage dozens of AI agents doing the routine parts.
The open question
If AI eventually handles all knowledge work, and then physical work, what is left for people? Nobody knows whether that day will come. But even with judgment solved, governance isn’t: people still have to take part in a democracy, build agreement, and make the big decisions together. Teaching students to do that well may be the most valuable thing a school can do.
Module 3
Orchestrating AI agents
The next skill isn’t asking a chatbot a question. It’s directing a team of AI agents on a real project, and knowing when to step in.
This lesson is still being built. Here’s the outline. For the full lesson materials, email ben@fix.school.
A model that can ask for tools
On its own, a language model can only write text. Give it tools and it can do more: a program around the model watches for requests it writes, like “search for this” or “open that page,” runs them, and pastes the results back in. When the steps are fixed in advance, that’s a workflow. When the model decides each next step itself, it’s an agent (Anthropic).
The loop
An agent works in a loop: think about the next step, act, read what came back, and check whether the goal is met. If not, it goes around again.
When it stops
The goal is met
It answers without asking for another tool. That ends the loop.
It needs you
It’s stuck, or it’s about to do something it can’t take back, like sending an email. A good agent stops and asks.
It hits a limit
Builders cap how many steps an agent can take, so a confused one can’t run forever.
Watch one work
Step through a real-looking run, one move at a time. Watch for the two loops, the moment it decides to stop, and the checkpoint at the end.
You give the agent a goal, not step-by-step instructions.
- You
When is the next school board meeting in Maple Grove, and is AI policy on the agenda?
The definition many builders use: “An LLM agent runs tools in a loop to achieve a goal” (Simon Willison).
From one agent to a team
For bigger projects, one agent becomes the orchestrator. It splits the work into pieces, hands each to a worker agent, then checks and combines what comes back (Anthropic). Your job is to manage the orchestrator: give it a clear brief, and review what it hands you.
How much to let it do alone
Andrej Karpathy calls this the autonomy slider: from autocomplete that suggests your next word, to an agent that works for hours on its own. The skill is choosing where to set it for each task, and the riskier the task, the more often you check in.
They can work longer every year
Researchers at METR measure how long a task AI agents can finish, timed by how long it takes a skilled person. Since 2019, that length has doubled about every seven months (METR, 2025). The longer agents can work alone, the more managing them matters.
The lessons
Write a brief an agent can follow
Say the goal, the context, what “done” looks like, and what not to do. Then watch where the agent goes wrong, and fix the brief, not the output.
Split a project into pieces
Decide which parts can run at the same time, which depend on each other, and what each agent needs to know to do its part.
Check the work
Open the sources, run the code, compare two agents’ answers to the same question. An agent sounds just as sure when it’s wrong.
Run a team on a real project
Research a problem at your school and draft a proposal with a small team of agents, then present it, and what you’d do differently next time.
Want to run these with your class or club? Email ben@fix.school for the materials.
Module 4
AI and trust
Work made with AI often reads as low-effort or inauthentic. As AI shows up in more of what we make, that reaction only gets more likely.
Why people push back
People value the same work less once they learn AI made it. In one study, people rated identical paintings as less likeable, less beautiful, less profound and worth less when they were labeled AI-made. In another set of 13 experiments, people who said they had used AI were trusted less than people who said nothing, and being found out was worse still.
How to keep people’s trust
Use human art for posters
Even when AI is faster. Draw it, photograph it, or ask a friend who draws. People notice the difference, and someone gets a chance to make something real.
Don’t hand down one right way to use AI
The technology changes every few months. Share your opinions about how you use it, but don’t lock people into rules or views that are hard to understand, or that will be out of date by spring.
Say when you used it
If AI helped, say so up front. Disclosing costs a little trust; getting caught hiding it costs much more (Schilke & Reimann, 2025).
Bring it to your school
Run this at your school
Use it in a class, a club or an after-school workshop. Want help, or the agent lesson materials? Email us. For teachers, there’s a staff version.