The best products I’ve used make me curious. Not just efficient. And when I look at what’s actually producing that feeling, it’s not novelty. It’s proximity.
A recommendation that nails your taste is satisfying. But a recommendation that gets it 80% right, close enough that you can feel the algorithm’s model of you, is more interesting. You notice the gap between what the system thinks you want and what you actually reach for. That near-miss creates a specific cognitive state: you almost recognize the thing, but not quite. And that’s where curiosity lives. Not in the unknown. In the almost-known.
In a diary study I followed during a streaming product redesign, listeners described the same pattern consistently. A feed that was entirely familiar felt stale. Pure novelty felt random and alienating. The sweet spot was roughly 80% recognition, 20% stretch. But the reason that ratio works isn’t because 80/20 is some magic number. It works because the 20% stretch sits right at the edge of what the listener already knows. Close enough to feel like “this knows me,” far enough to open a question about your own taste that you didn’t have before.
Robert Bjork’s concept of “desirable difficulty” in learning research describes the same mechanism. Retention improves when the task is just hard enough to require effort, just unfamiliar enough to force engagement, but not so far that the learner gives up. The productive zone is the near-miss. And the near-miss is what generates the impulse to keep going.
This matters because most AI products right now optimize hard for accuracy. Get the answer right. Predict the next word. Match the intent. That works for search. But for anything involving taste, learning, or creative exploration, a perfect match kills the very state you want to create. A system that’s always exactly right teaches you nothing about yourself. A system that’s right in a way that slightly surprises you teaches you what you didn’t know you were looking for.
That’s the design constraint worth taking seriously. Not “how do we get this right?” but “how do we get this right in a way that keeps the person thinking?” If you optimize purely for accuracy, you optimize curiosity out of the product. The near-miss isn’t a flaw in the recommendation. It’s the feature.