Your Brain Learns by Being Wrong

In the 1990s, Wolfram Schultz at the University of Cambridge stuck electrodes into the midbrain of monkeys and watched their dopamine neurons fire. What he found broke the most basic assumption about reward.

Dopamine neurons don't fire when you get a reward.

They fire when you get a reward you didn't expect. If the reward is fully predicted, the neurons go silent. If a predicted reward fails to show up, they suppress below baseline. The signal isn't "that was good." It's "that was different than I thought."

Schultz published this in Science in 1997. It became one of the most cited findings in modern neuroscience. Not because it told us something interesting about monkeys. Because it told us something fundamental about how brains learn.

They learn from being wrong.

The Error Is the Signal

If your brain is a prediction machine (and the first article in this series laid out the evidence that it is), the most important signal isn't the prediction itself. It's the mismatch. The moment reality deviates from what you expected.

That mismatch is called a prediction error. It's the only thing that forces your brain to update its model of the world.

Andy Clark, a philosopher at the University of Sussex, described the brain as a system constantly trying to minimize prediction error. In his 2013 paper "Whatever Next?" in Behavioral and Brain Sciences, he argued the brain has two options when it encounters a mismatch. Update its predictions to match reality. Or act on reality to make it match its predictions.

Think about that second one. When you feel cold and put on a jacket, your brain didn't passively receive the signal "cold." It predicted a body temperature, detected an error, and moved your body to eliminate the mismatch. Perception and action both serve the same goal. Minimize the error.

Why You Didn't See the Gorilla

Prediction errors explain one of the strangest findings in psychology.

In 1999, Daniel Simons and Christopher Chabris at Harvard had participants watch a video of people passing a basketball and count the passes. Halfway through, a person in a gorilla suit walked into the frame, beat their chest, and walked off.

Roughly 50% didn't see the gorilla.

Not "didn't notice right away." Didn't see it at all. They refused to believe it happened until they rewatched the video.

Their eyes received the gorilla data fine. But their brain was modeling a scene of people throwing a ball. A gorilla didn't fit. The sensory data didn't generate a strong enough prediction error to override the model, so it never reached conscious awareness.

Your brain doesn't show you reality. It shows you predictions. Only errors large enough, surprising enough, break through.

Errors Flow Up, Predictions Flow Down

Karl Friston at University College London formalized this in a 2005 paper in the Philosophical Transactions of the Royal Society B. The brain is organized as a hierarchy of prediction machines. Higher levels send predictions downward. Lower levels send errors upward.

Each level tries to explain away the errors from below it. Only errors that can't be resolved at lower levels propagate upward to update higher-level beliefs.

Picture layers of management. A front-line sensor detects something unexpected. The first level checks whether it can explain the error with a minor adjustment. If it can, the error dies there. If not, it gets escalated. Only the really surprising stuff makes it to the top.

This is why strong beliefs are so hard to change. Not because people are stupid or stubborn. The architecture is designed to absorb small contradictions before they reach the belief level. Jakob Hohwy laid this out in The Predictive Mind (2013). Lower-level errors get explained away. Reinterpreted. Absorbed. The belief never sees them.

The hollow mask illusion shows this in action. Look at the concave inside of a face mask, and your brain flips it so it looks convex. A normal face poking outward, even though it's caved inward. Your prior that faces are convex is so strong it overrides the visual data.

Danai Dima and colleagues showed in a 2009 NeuroImage study that people with schizophrenia don't fall for the illusion. They see the mask as concave, as it actually is. Their prediction machinery weights sensory data more heavily relative to priors. In this specific case, they're more right. But the cost of that weakened prediction system shows up everywhere else.

What This Means for Changing Anything

To update a deeply held prediction, whether it's a belief, a habit, or an emotional pattern, you need prediction errors large enough and persistent enough to propagate through the entire hierarchy.

Small contradictions don't work. Your brain absorbs them. One good experience doesn't override years of expecting failure. The lower levels explain it away. Lucky break. Exception. Fluke.

Michelle Craske at UCLA has studied this in anxiety treatment. In a 2014 paper in Behaviour Research and Therapy, she argued the most effective exposure therapy isn't about "getting used to" the feared thing. It's about maximizing prediction error. The bigger the mismatch between what you predicted (panic, disaster, rejection) and what actually happened (nothing, safety, acceptance), the more the brain is forced to update.

This is why easing gently into change often fails. A tiny deviation generates a tiny error. Absorbed. Explained away. The experiences that actually rewired me weren't the small incremental shifts. They were times reality was so different from my prediction that my brain couldn't ignore it. Moving to a new place. Losing someone. Building something that worked when I was sure it wouldn't. Those errors were too big to explain away.

Dopamine Isn't What You Think

Back to Schultz's finding, because the implications go further than learning.

The reward prediction error signal became the foundation of modern reinforcement learning in AI. DeepMind's algorithms, the ones behind AlphaGo, are directly inspired by this biological mechanism. The artificial system learns the same way your neurons do. Not from reward. From unexpected reward.

In humans, predictable pleasures lose their punch. The first time you try a great restaurant, dopamine fires. By the tenth visit, silence. The food didn't get worse. Your prediction caught up. No error, no signal.

This is why novelty feels alive and routine feels flat. Not because new is better. Because new generates prediction errors. Your brain pays attention when it's wrong. When it's right, it coasts.

The Brain Doesn't Want Truth

Your brain's job isn't to show you reality accurately. Its job is to minimize prediction error. Two ways to do that. Update the model to match reality. Or filter reality to match the model.

It does both. Constantly. And it strongly prefers the second, because updating high-level beliefs is expensive. It cascades through the hierarchy. It's destabilizing.

So your brain would rather explain away contradictory evidence than change what it believes. Rather not show you the gorilla than rebuild its model of the scene. Rather reinterpret a compliment as sarcasm than update its prediction that people don't like you.

This isn't a bug you can patch with awareness. It's the architecture. Prediction errors are the only signal that forces an update. The system is specifically designed to suppress them.

The next article in this series explores what happens when this turns inward, when your brain starts predicting your own body's signals and constructing what you experience as emotion. The feelings you think are reactions to the world might be predictions too.

Sources

• A Neural Substrate of Prediction and Reward (Schultz, 1997, Science) (opens in new tab)
• Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science (Clark, 2013, Behavioral and Brain Sciences) (opens in new tab)
• A Theory of Cortical Responses (Friston, 2005, Philosophical Transactions of the Royal Society B) (opens in new tab)
• Gorillas in Our Midst: Sustained Inattentional Blindness for Dynamic Events (Simons & Chabris, 1999, Perception) (opens in new tab)
• The Predictive Mind (Hohwy, 2013, Oxford University Press) (opens in new tab)
• Understanding Why Patients with Schizophrenia Do Not Perceive the Hollow-Mask Illusion Using Dynamic Causal Modelling (Dima et al., 2009, NeuroImage) (opens in new tab)
• Maximizing Exposure Therapy: An Inhibitory Learning Approach (Craske et al., 2014, Behaviour Research and Therapy) (opens in new tab)
• The Experience Machine: How Our Minds Predict and Shape Reality (Clark, 2023, Pantheon) (opens in new tab)

---

Part of the Prediction Machine series. Previous: Your Brain Is Hallucinating Right Now. Next: Your Brain Runs on Probability, Not Facts.