I’ve now run four AI workshops — two in Mumbai, two in Hyderabad — for somewhere around two hundred students. They’re forty-five to fifty minutes each, though the good ones run over. I’m in grade ten, and the rooms are mostly eighth to tenth graders, sometimes older. If I’ve learned one thing from doing this, it’s that the single most useful sentence I can leave a student with isn’t about how a neural network is built. It’s this: AI predicts, it doesn’t know.

That’s the idea I spend the first twenty minutes building toward. We talk about what a neural network actually is, how a model processes what you give it, and then the part that matters most — that the model isn’t retrieving a correct answer from somewhere, it’s predicting the most likely next thing to say. Once that clicks, a lot follows for free. Hallucination stops being a mysterious glitch and becomes an obvious consequence: a system built to produce plausible output will confidently produce plausible output even when it’s wrong, because “wrong” and “plausible” are not the same axis. A student who understands that treats the tool completely differently. They check it. They use it where being roughly right fast is valuable and avoid it where being exactly right matters. The mechanics aren’t the point on their own — they’re what makes that judgment possible.

I care about getting there honestly because the alternative is teaching AI as magic, and magic is useless to a student. If you think the machine knows, you defer to it. If you understand it’s guessing well, you stay in charge of it. That distinction is the whole game, and it’s why I refuse to skip the twenty minutes even when a room clearly wants to jump straight to “so what do I type.”

The first two workshops were with my robotics team, which is based in Mumbai. The Hyderabad ones I did on my own — not for the team’s numbers, just because I wanted to keep doing it. I’ll be honest that this distinction mattered to me. It’s easy to do something once when there’s a scoreboard attached. The second time you do it with nothing to show for it is when you find out whether you actually believed in it or were just collecting the credit. I found out I believed in it.

The second half is where it gets specific, and where I stopped giving the same talk twice. Most of the schools I go to are low-income schools, and the useful application of AI turns out to depend heavily on who’s in the room. For a lot of these students, AI isn’t a productivity toy or a homework shortcut — it’s leverage. So instead of a generic “here’s how to write better prompts,” I tailor it. For some groups the most valuable thing I can show is how to use these tools to earn on the side, because that’s the constraint that actually shapes their week. For others it’s how to compress learning: pick a skill, spend fifteen focused minutes a day for thirty days with the model as a tutor that never gets tired of your questions, and you come out the other side with something real. The tool is the same. What it’s for changes with what the person needs from it, and pretending otherwise would waste their time.

At the end I ask questions — actual questions, back to the room — because a workshop where everyone nods and nobody understood is a failure I can’t see from the front. This is the part that most often blows past the time limit, and I’ve stopped minding. It’s also, selfishly, the part that’s sharpened my own understanding the most. You don’t really know whether you understand how a model predicts until a fourteen-year-old asks you a question that assumes it thinks, and you have to take the idea apart and rebuild it in words that land. Every time that happens, my own version of the explanation gets a little cleaner. Teaching it is the best debugging tool I’ve found for my own understanding of it.

Two hundred students is the number I could put on a form, but it isn’t the thing I care about. The thing I care about is narrower and harder to count: how many of them walked out able to use the tool instead of being impressed by it. I don’t get to measure that directly. But every time a student stops mid-session and says some version of “wait, so it’s just guessing?” — a little indignant, like they’ve been let in on something — I know at least that one landed. That’s the sentence I’m actually there to deliver. Everything else is scaffolding around it.