There's a moment I keep having at my desk, somewhere around the third time I've rewritten the same prompt.
I stop and think: I've done this before. Not for instructions to a language model, but for instructions on a worksheet for kids. The feeling is identical: the subtle, slightly embarrassed realisation that the instructions weren't clear enough, and whatever came back is not wrong exactly, but just not what I meant. Children do the unexpected thing when you leave room for the unexpected thing. So does GenAI.
The more time I spend working with AI, the more I notice how much of what I thought I left behind in the classroom turns out to be exactly what's needed here.
Take prompting. A good prompt, like a good task description on a worksheet, specifies the output format, names the constraints, gives an example if the task is genuinely ambiguous, and defines what "done" looks like. Miss any of those and you get creative interpretations of your intent - interesting, sometimes, but rarely what you were after. The model does what any reasonable reader of your instructions would do: fill in the gaps with something plausible. Students behave the same. In both cases the solution is almost identical.
Role-play is another one. If you want a kid to learn how a scientist presents lab results, you don't explain it - you give them a lab coat and tell them they are a scientist now. Set the context, and the output shifts in ways that a list of instructions alone rarely achieves. The classroom technique that felt slightly theatrical turns out to have a precise analogue in the world of system prompts.
Evals are grading. Output criteria are rubrics. Positive reinforcement shapes behaviour over time, whether the feedback loop closes in milliseconds or in months.
Neural networks were modelled after real networks of neurons - that's not news, it's the origin story. The whole architecture was built by looking at how brains learn and asking: what if we did that, but with numbers? So of course training a model looks like training a brain. That's the point. That's where it came from.
That's not the surprising thing. The surprising thing is what follows from it:
If the model was built to learn the way children learn - through examples, feedback, reward signals, accumulated adjustment over time - then anyone who has spent years figuring out how to make that process work in rooms full of actual humans has been doing a version of prompt engineering or LLM-training all along. Not metaphorically. Structurally.
Teachers are, in extension, trained AI engineers. They just don't have the title.
Now think vice versa: what else can we learn from the classroom about how to be better at channelling the full potential of AI?