A few years ago someone asked me to teach a sales force to use AI. Not a room of engineers. People spread across a continent who sell industrial wood coatings for a living, whose real skill is reading a customer and knowing a product cold, and who mostly wanted to know whether this thing was about to take their jobs. I stood at the front of that room as the person who was supposed to have answers, and I want to be honest about what I noticed first: the two reactions in the audience were the same two reactions I now see in almost every school, training department, and faculty meeting I walk into.
Some people had already decided the tool was a cheat, and their instinct was to keep it out. Others had already decided it was a shortcut, and their instinct was to hand it the work. Both groups thought they were making a decision about technology. Neither of them was, and that gap is what this whole book is about.
This is a plain field guide to bringing AI into your teaching without faking the actual learning. If you deliver learning for a living, in a classroom, a training room, an L&D team, a lecture hall, you can read this chapter on its own and walk away with something usable. Read in order, it is chapter one of thirteen. Let me start where most of us start, which is by getting it wrong in one of two ways.
The two moves that feel like answers
The first move is to ban it. And I want to give that its due, because the people reaching for a ban are usually the people who care most about real learning. Their worry is legitimate. If a learner can hand in an essay they did not write, or an analysis they cannot explain, then the credential stops meaning anything and so does the effort of the learners sitting next to them who did the work. That is a real integrity problem, and anyone who waves it away has not spent enough time responsible for someone else’s understanding. So the ban deserves real respect. It is a serious answer to a serious worry.
The second move is to bolt it on. This is the friendly version, the one that shows up in the enthusiastic staff meeting. Use it to generate worksheets. Use it to draft the quiz. Use it to summarize the reading, grade the first pass, spin up the lesson plan you did not have time to write. Treat it as a very fast assistant and get more done. And I want to give that its due as well, because a lot of teaching labour is genuinely repetitive, and anyone offering you time back is not your enemy.
Here is what caught my attention in that sales training, and what has held it ever since. Both moves feel like a position. You can defend either one in a meeting. But both of them answer the same question, which is what does this tool do, and quietly skip the question that actually determines whether anybody learns anything. The banner and the bolter are arguing about the tool. They are on the same side of a line they cannot see.
It was never a technology question
The question underneath all of this is a teaching question. How do people actually learn when this thing is in the room with them? The real subject is what happens inside the learner while the tool does its work, which makes it a question about people. That is pedagogy, and it is the discipline most of us in the room were never trained in, myself included until fairly recently. I am partway through a master’s in learning and technology, and I am doing it out of plain curiosity rather than any career plan. The question of how adults actually learn turned out to be far more interesting, and far less settled, than I assumed from thirty years of building software and thinking I understood teaching because I could explain things.
Once you put the pedagogy question first, the two moves both look like ways of avoiding it. Banning the tool avoids asking how your learners will meet AI the moment they leave your care, which they will, on the first day of the job you prepared them for. Bolting it on avoids asking whether the learner’s mind was ever engaged, or whether you just made the finished object appear faster. Both are tempting precisely because the real question is harder and does not have a settings menu.
So here is the one move I want you to take from this chapter, and I will spend the rest of the book earning it. You do not teach the machine. The instinct that it is a thing to be commanded, out-prompted, corrected, wrestled into obedience, is the trap, and it is a comfortable trap for those of us who like being the smartest one in the room. The move is to learn how to learn with it, honestly, in your own work first, and then teach your learners to do the same. That is a shift from managing a tool to practising a skill, and the skill is learning itself.
I keep a public page of how I use AI and how I disclose it, at my stances, and I keep it because modelling honest use is the first thing I owe the people I teach.
What the machine is actually doing
You cannot teach with something you have not looked at honestly, so let me give you just enough of the machine to make the next point land. I will spend all of chapter two on the proper mental model, so treat this as the sketch, not the blueprint. A large language model is a prediction engine. It has read a staggering amount of human text and learned, in fine statistical detail, which word tends to follow which. When you ask it something, it is not looking up an answer or reasoning toward a fact. It is generating the most likely next stretch of words given everything before it. Most of the time that lands on something true, because true things are well represented in what people have written. Sometimes it lands on something false with exactly the same confidence, because confidence is not something it measures. Fluency is the only setting it has.
I ran into this in a way I still find funny. I trained a small AI editor on my own writing, so it would catch when a draft drifted off my voice. I fed it an essay I had written entirely by hand, no machine involved, and it flagged the thing as very likely AI-generated, with eight specific tells. A human editor read the same essay and found three real issues. The machine was not lying to me. It was doing precisely what it does, producing a confident, fluent verdict that happened to be wrong, about the one topic where I could actually check. That is the whole character of the thing in one incident. It will sound just as sure when it is right as when it is not, and it has no idea which is which. Hold onto that, because it is the reason the next part matters so much.
The part nobody wants to say out loud
Here is the uncomfortable bit, and it is the reason this is a pedagogy problem and not an IT problem. A learner can sit down with one of these tools, produce work that is fluent, correct-looking, even genuinely good, and come away having learned close to nothing, while feeling like they nailed it. The output is real. The understanding behind it can be entirely absent. And the worst part is that the feeling of having understood is often strongest right at the moment the least learning happened, because the tool did the effortful part and left the learner the pleasant experience of watching it come together.
Think about the last time you drove somewhere new with the navigation voice guiding every turn. You arrived. You felt like you got there. Now picture doing that same drive a week later with your phone dead, and notice how little of the route you could actually reproduce. You never learned the way, because something else held the map for you, and arriving felt exactly like knowing. That gap, between the feeling of getting there and the ability to do it again on your own, is the single most important thing an educator has to see clearly in the age of AI. I am not going to unpack it fully here. It is the whole spine of chapter three, and it deserves its own room.
What I will tell you is that this trap is older than AI. Learning scientists have a name for a close cousin of it, the illusion of explanatory depth, the well-documented finding that people believe they understand how ordinary things work, a zipper, a toilet, a ballpoint pen, a door lock, in far more detail than they actually do, right up until you ask them to explain it step by step and the confidence collapses. Leonid Rozenblit and Frank Keil documented this across a dozen studies in 2002, in a paper on the limits of what they called folk science. AI pours fuel on that illusion, because now the fluent, detailed-looking explanation is always one sentence away, and the collapse never has to happen.
My honest disclosure is that this has fooled me, and it cost me. Not long ago I was working from a technical process well outside my depth, the correct order of steps to catalyse a chemical cross-linker, with the original instructions written in Italian on top of everything else. So I did the exact thing I would tell any of you to do: I asked an AI to translate it and explain the process, and it produced a clear, confident walkthrough that left me feeling like I had it. A few minutes later I was standing in an application lab looking at a cup of jelly that was supposed to be liquid. The AI had quietly swapped two steps, diluting the cross-linker with water and diluting the already cross-linked material with water, a small distinction I would have caught if I had actually understood the chemistry rather than a tidy summary of it. It read as competence right up until the moment it met a physical cup. The distance between feeling like I understood and being able to do the thing cost me a ruined batch, wasted material, and an afternoon I do not get back.
Try this before next week
Do not take my word for any of this. Feel it in your own hands first, because you cannot design your teaching around a trap you have only read about. So here is the one small thing to try this week, and it is aimed at your own learning, not your learners’, on purpose.
Pick something you genuinely do not know yet but want to, a real gap, not a topic you already teach. Make it mid-sized and checkable: not something you could already half-do, which proves nothing, and not something so far past you that failing only tells you it was hard. Sit down with an AI tool and learn it the easy way, the way a rushed learner would, asking it to explain and summarize until you feel like you get it. Right there, at the moment it feels like it has clicked and before you close anything, write one line predicting how much of it you could reproduce on your own. Then close the laptop and wait an hour, or a day.
Now, from memory and with nothing open in front of you, try to reproduce it. Explain it out loud, or better, actually do the task the knowledge was for. Compare what you managed against the prediction you wrote, and watch the gap open between how much you felt you understood and how much you can produce on your own. Write down what fell out of your head the moment the machine was gone. That list, and the distance between your prediction and your performance, is the most useful teaching document you will make all quarter, because it is a map of exactly where the feeling of learning and the fact of learning came apart, in your own mind, where you can see the seams.
What comes next
Banning it or bolting it on both dodge the real work, which starts with you learning to learn with this thing honestly, so you can eventually teach your people to do the same. That is the ground the rest of this book is built on. Before we can talk about how learners work with AI, though, we need a clear-eyed picture of what they are actually working with. So next week, in chapter two, I want to give you the proper mental model, what a large language model really is for someone who has to teach with one, why it is so confidently wrong, and what that changes about where you can trust it in a learning setting and where you cannot. Bring the list you made. We are going to need it.
Leave a Reply