A friend of mine is building what he calls “Duolingo for everything.” The idea: use AI to generate personalized learning content for any subject, with images, diagrams, and structured materials, all on demand. The generation part works. You pick a topic and the app produces a clear, tailored lesson in seconds.
The problem is that clear and tailored isn’t the same as learned. And the gap between those two things is where most AI learning tools quietly fail.
There’s a framework from the MIT Media Lab’s AHA (Advancing Humans with AI) research on human flourishing that sharpens this: offload and unload timing. Sometimes a tool should offload cognitive work, handle the complexity, reduce the burden. But sometimes it should unload, step back and force the person to struggle. The question is when to do which.
The book Rewire by Nichole Vignola makes the neuroscience concrete. There are two kinds of dopamine release. One comes from effort: pushing through friction, solving something hard, completing a challenge that required real engagement. That dopamine is associated with durable motivation and lasting behavioral change. The other comes from instant reward: scrolling, swiping, getting a perfectly packaged answer without working for it. That’s the dopamine loop behind social media. It feels good and builds nothing.
Most AI learning tools are accidentally optimizing for the second kind. The content is clear. The explanations are perfect. The user feels smart. But they haven’t built the mental model themselves. The struggle was where the learning would have happened, and the tool removed it.
I challenged my friend on this. His app can generate a great cocktail lesson on any topic in seconds. But if the user never has to recall, apply, or wrestle with the material, what actually sticks? Duolingo is effective partly because it delays hints, forces recall, and spaces repetition. It adds friction by design. The question is whether his AI-native tool can do the same thing adaptively: adjusting per user and per moment, knowing when ease helps and when ease hurts.
The harder product problem underneath is that you can’t easily change human behavior. Most people aren’t opening a learning app because they want to be challenged. They’re on their phone. They’re tired. They want the easy version. During our conversation we kept coming back to the same conclusion: don’t try to make lazy users disciplined. Design for the narrow window when someone is genuinely ready to push, and make that window count. Meet the user where they already are, and maximize the learning that happens inside the time they’re willing to give.
That tension, between what users want in the moment and what actually produces learning, is the core design problem for this entire category. No amount of generated content solves it.