What a Two-Year-Old Taught Me About AI

Genevieve Bosma Martinez offers this commentary on what a two-year-old taught her about AI. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

My nephew is two and a half. Lately his favorite four words are “I can do it myself.”

Sometimes he can. He wrestles with snapping his carseat belt with real determination, insists on picking out his own snack, waves off my hand on the stairs like it personally offended him. Sometimes he genuinely cannot. The zipper wins. The plate tips. He needs help whether he’s ready to admit it or not.

Watching him, I keep noticing how hard it is, in any given moment, to know which situation I’m looking at. Step in too fast and he never learns the zipper. Step back too far and he melts down at the bottom of the stairs, furious and stuck. Psychologists have a name for the space he’s standing in: the zone between what he can do alone and what he can only do with the right kind of help, at the right moment. That zone moves every time he grows, which might be why this judgment call never really gets easier. Just more familiar.

I’ve started to notice this isn’t really a developmental question. It’s the same one adults face constantly, we’ve just traded shoes and jackets for lesson plans, budget models, and now, increasingly, AI.

Most of what I read about AI circles one question: what can this tool do for me? Lately I’ve been wondering if the better first question is simpler: what is this task actually for?

That question changes everything downstream of it. A task designed to produce a document and a task designed to build a skill can look nearly identical from the outside. Hand AI the first one, and you’ve saved yourself an afternoon. Hand AI the second, and you may have quietly canceled the reason the task existed in the first place.

Every Task is Doing Two Things at Once

Nearly every meaningful task produces two things: the output, and the person doing it. The memo gets written, but so does the writer. The lesson gets planned, but so does the teacher’s feel for pacing a room. We tend to judge tasks only by the first thing. Did the deck get built? Did the report read well? That’s fine for administrative work. It’s a strange standard for anything meant to develop a person, because the output can be excellent while the person walks away having learned almost nothing from producing it.

Memory researchers have shown something similar: testing yourself on material, forcing your mind to produce an answer instead of just rereading it, builds retention that lasts far longer. The value was never really in the answer. It was in reaching for it.

3 Ways to Respond, Not 2

Once you start asking what a task is for, you notice you have three real options, not the two we usually assume. You can do it yourself, because the friction is the point. Some tasks only teach you anything if you stay in them long enough to get stuck.

You can get help that develops you, which is easy to confuse with outsourcing your thinking but isn’t the same thing. A mentor asking a better question instead of handing you the answer is help. AI used as a genuine thinking partner, pressure-testing a plan instead of writing it for you, is help too. Both leave you better at the thing, not just finished with it, echoing what researchers found decades ago studying expertise: it comes from focused practice with real feedback, not raw hours.

Or you can delegate the whole thing. If a task is repetitive or purely administrative, that’s exactly where AI should be doing the work. There’s no hidden virtue in manually reformatting a spreadsheet.

An EdTech leader evaluating a new AI tutoring platform runs into all three at once. Automating scheduling and grading multiple-choice questions is friction worth removing. But a feature that instantly answers a homework question only looks like help. It quietly does the thinking a student was supposed to be building. The real question was never “does this tool work,” but which of these three things is actually happening, feature by feature.

Going Slow to Go Fast

We live in a culture that treats speed as its own virtue. But speed only matters if it serves what the task was for. Rushing through something meant to build judgment doesn’t save time, it just moves the cost downstream, usually right when that judgment turns out to be missing. There’s a name for this too, a desirable difficulty: learning conditions that feel slower and harder but produce understanding that lasts. That idea deserves more room than I can give it here.

For now, it’s enough to say some effort is the whole point, and AI’s real opportunity may be helping us tell which effort that is.

A Few Questions Worth Carrying

I don’t think this calls for a rulebook. What holds up, for me, is pausing to ask:

• What is this task actually designed to accomplish?
• What capability is it trying to build, in me or in someone else?
• What kind of help would genuinely help me grow, instead of just getting this done?
• Is this use of AI extending my thinking, or quietly standing in for it?

These questions work the same whether you’re a teacher scaffolding an assignment, a workforce leader deciding whether AI should draft a curriculum from scratch, or an aunt deciding whether to help with a zipper.

What My Nephew is Teaching Me

The lesson isn’t that independence is always right. He needs help constantly. It isn’t that help is always right either. Some of what he’s learning, he can only learn by failing at that zipper a dozen more times. The wisdom was never in the doing, or the helping, on their own. It was in noticing, task by task, which one the moment was actually asking for.

So maybe that’s the question worth carrying with you, the next time you reach for a tool of any kind. What is this moment actually asking of me? And am I about to help, or about to get in the way?