Why massed practice doesn't work
Introducing "reward prediction error"
When I was writing books with Temple Grandin, she would observe, frequently, that the key to reading scientific literature was to know when different researchers were using different terms for the same thing.
She was right.
So before getting into the reason massed practice doesn’t work, I’m posting a list of terms researchers use when they’re talking about information-integration learning, the subject of Katie’s & my May talk at researchEd.1
information-integration learning
reinforcement learning
trial-and-error learning
implicit learning
statistical learning
probabilistic learning
procedural learning
operant conditioning
instrumental conditioning
intuitive learning (folk term)
These are the terms I’m aware of. There may be others.
Why do we have so many terms for the same fundamental mechanism?
One reason is likely that the researchers who use these terms are working in different fields (e.g. behaviorism versus cognitive neuroscience) and have never met. Nor have they read each other’s publications, in many cases.
Another reason is that they’re looking at different learning goals and outcomes. Some of these terms are used by researchers investigating category learning (e.g. information-integration learning); others are used by researchers investigating action learning (e.g. operant conditioning).
But all depend on the same fundamental mechanism of trial and error.
Trial and error learning is what it sounds like: you make a choice (that’s the trial), then you find out — immediately (timing is important) — whether your choice was right or wrong. This feedback stage is the error part. If you’re wrong, you make another choice, and learning proceeds as you reduce and finally eliminate error.
Inside the brain, though, learning isn’t a simple matter of being wrong until you’re right. Instead, the difference between what you expect to happen on each trial and what actually does happen is the learning signal inside your brain.
Here’s the counterintuitive part.
If you expect to be right and you are right, you don’t learn.
That’s why massed practice doesn’t work: you already know you’re right when you choose your answer. When we stop making mistakes, we stop learning.2
If you expect to be wrong and you are wrong, you don’t learn.
But if you expect to be right and you’re wrong3 — or if you expect to be wrong and you’re right — you learn.
The term for this phenomenon is reward prediction error. 4
More soon.
Why struggle is (usually) bad and right answers are good: “information-integration learning” and reward prediction error
There is probably a very good evolutionary reason for this design — which we can see when we consider the fact that addictions happen when the learning process does not stop, a subject for a future substack
There is actually some disagreement over whether we learn from negative prediction errors — from wrong answers we thought were right — but there’s no question we learn from positive “RPEs.” UPDATE 3/22/2026: see Retrieval Attempts Enhance Learning, but Retrieval Success (Versus Failure) Does Not Matter UPDATE 3/27/2026: the phenomenon of hypercorrection is more evidence that we learn from wrong answers as well as right answers. The key in both cases is going to be the degree of confidence we had in our answer, right or wrong.
It looks like reward prediction errors underlie observational learning, too.
I think one of the fallacies is to conflate not making mistakes with doing things automatically, and/or to conflate practicing doing things automatically with achieving automaticity.
As you point out, automaticity isn't achieved by getting things right automatically.