AI makes learning feel too easy

I've spent a lot of hours learning things with AI chatbots, and I've noticed a pattern I don't love.

The hour goes great. It explains the thing clearly, answers every follow-up, and untangles each knot the moment it shows up. By the end the material feels obvious. Then a few days later I go to actually use it, and I'm reaching for the chat again almost immediately. The hour felt like learning. Not much of it stuck.

The easy explanation is that I'm being lazy, or writing bad prompts. I don't buy it. It happens even when I'm paying close attention, asking specific questions, and chasing down every follow-up I have. I think the real issue is more annoying than that: the feeling of understanding and the fact of understanding are two different things, and AI is the first tool I've used that can max out the first while starving the second. It doesn't feel broken when it happens. It feels like the best teacher you've ever had. That's exactly why it gets you.

How we decide we've learned something

We don't have an internal exam that tells us whether we've learned something. What we have instead is a feeling: how smooth and obvious the material feels as it goes past. Call it ease.

Ease is a bad witness. Re-reading your notes feels productive, because the second pass runs smoother than the first, and it does almost nothing for retention. Trying to recall something cold feels slow and inefficient, and it's one of the most effective things you can do to actually remember it. Psychologists have a name for this: desirable difficulties, from Robert Bjork's lab.¹ The conditions that make learning feel harder tend to make it last. The ones that make it feel easy tend to let it fade.

So the feeling and the result aren't just loosely related. A lot of the time they point in opposite directions.

What AI does to the difficulty

Now look at what AI does to each of those difficulties.

The struggle to pull an answer out of memory? It hands you the answer before you've started reaching. The work of putting a fuzzy idea into your own words? It produces better words than yours, instantly. The discomfort of sitting in a confusion until it resolves? It clears the confusion on contact, because that's the whole job of a helpful assistant.

So AI isn't refusing to teach you. It's doing something sneakier. It strips out the exact friction the learning was hidden inside, and it does it while producing a stronger sense of ease than anything else I've used. Everything is smooth. Little of it is earned. And here's the part that should bother you: the smoother the session feels, the more of the actual work the machine probably did instead of you.

This has actually been measured

I know this sounds like the kind of thing you can argue either way, so it's worth knowing it's been measured.

In a randomized trial² with about a thousand high-school students, one group practiced math with a plain chatbot. While they had the bot sitting next to them, they scored 48% higher than students with no AI. Then the researchers took the bot away for the exam, and the same students scored 17% worse than the ones who'd never used it. The crutch had felt like a leg.

The part that gets me is what the students believed. The chatbot users thought they'd learned just as much as everyone else. From the inside, the tool that quietly cost them the material was indistinguishable from one that taught it. Perceived learning went up, real learning went down, and there was no signal in between to tell them apart. That's the illusion, sitting right there in the data.

But AI isn't the problem

Here's where I want to be careful, because the obvious conclusion — "so AI is bad for learning" — is wrong.

There's a second trial³ that points the other way. College students taught by an AI tutor beat a live, active-learning class by a wide margin. Same general technology. But this tutor wasn't a plain chatbot. The instructors had built it to make students retrieve, generate, and work through problems themselves, with the difficulty deliberately put back in. Same model, opposite design, opposite result.

So the thing that predicted learning was never whether AI was involved. It was whether the difficulty survived contact with it.

What I actually do about it

The fix isn't to use AI less. It's to stop trusting the feeling and put the friction back by hand. A few things that work for me:

• After it explains something, I close the tab and rebuild the idea from nothing. Whatever I can't rebuild, I never actually learned — no matter how clear the explanation was.
• I make it ask me the questions instead of answering mine.
• I make it withhold the solution and just mark my attempt.
• I treat a suspiciously smooth session as a warning, not a result. The effortless stretches are exactly where nothing got laid down.

The uncomfortable bit is that the friction is the part that's easiest to hand off, and it's also the part that is the learning. When you give away the struggle, you're not offloading the overhead around learning. You're offloading the learning.

The thing to remember

The danger with these tools was never that they give wrong answers. It's that they deliver the feeling of having learned, on demand, and that feeling is the gauge we use to decide we can stop. AI breaks the gauge. You finish an easy hour completely sure the thing is yours, right up until you reach for it and find the gap where it should be.

The explanation was perfect. You just never did the work.