Desirable difficulty: the hidden cost of delegating to AI

The work looks better and better with AI. But your capacity, not so much.

That's the uncomfortable part of the recent research on AI and cognition. And almost no one sees it coming, because there's no alarm that goes off. And no one teaching it either.

Three separate works, an academic synthesis by Singh et al., the University of Technology Sydney report by Lodge and Loble, and a series of Harvard interviews, arrive at the same place by separate paths. The problem is never AI. It's the moment you decide to use it.

AI presents itself as the perfect temptation. It simplifies, it speeds things up, it agrees with you. And on top of that it takes the tedious tasks, the ones no one wants to do. And that's the trap, because very often the tedious task was exactly where the learning was hiding. What's easy isn't always what's desirable.

Learning science has been warning about this since the 90s.

Robert Bjork, at UCLA, called it "desirable difficulty." His finding is counterintuitive but forceful: the effort it takes you to process something isn't an obstacle to learning it, it's the mechanism that locks the learning in. When something takes a bit of effort, your brain works, and that work is what makes the concept stick. Almost the same as at the gym: if you take the resistance off the weight, you stop building muscle. The brain is no different.

That's why the output can shine today while capacity quietly empties out inside. Sydney calls it "false mastery." The report comes out flawless, the email sounds professional, the presentation lands. Everything seems to improve. That's why no one stops to ask what's happening to their own head. But not all that glitters is gold. The gap only shows up the day you have to think without the tool, and that's when you realize the muscle, those skills you'd been building for a long time, is gone. Everything feels harder, more complex.

An honest caveat: in people who are just building their knowledge base, the evidence of this decline is fairly clear. In experts with already-formed judgment, it's still being studied. But in neither case does the problem warn you in time.

When I teach the responsible-use module, I suggest to my students that they go back to pen and paper every so often. The look on their faces is surprise, if not annoyance (sometimes something worse). "That's impossible today," they tell me.

And they're right about one thing: the point isn't the pencil.

The point is to preserve the step that costs effort. AI erases the desirable difficulty without you noticing, and sometimes you have to put it back on purpose. Think through the argument before you prompt. Solve the problem before asking for a check. Write the summary in your own words before generating one. The paper is just a trigger. The tool is optional. The friction isn't.

There are plenty of tips and pieces of advice that usually come with that module. But we can boil them down to one simple, practical rule:

• If you produce first and AI reacts after, it amplifies you.
• If AI produces first and you react after, it replaces you.

It's almost the same prompt. The only thing that changes is who makes the first effort. And that order changes everything.

Using AI well isn't measured by how much you use it. It's measured by where you choose to keep making the effort.

You first, or the tool first?