Biology of User Friction

Users don’t navigate your platform pixel by pixel; they predict what should be there based on years of digital experience. If your software follows their mental models, navigation is effortless. When you introduce innovative (AKA unpredictable) mechanics, their brain flags a prediction error. Cognitive load spikes, engagement tanks. The rule is simple: keep core mechanics predictable, and use deviations to capture attention.

Brains are built to take shortcuts and conserve energy. That is why so many teams rely on vanity metrics, metrics they don’t understand, or freeze when problems fall outside their immediate domain of expertise. Deep root-cause analysis is metabolically expensive. We are satisficers - we choose “good enough” instead of optimal. Because of this, we invented divorce. 🥁

Predictive processing

I earned my PhD alongside a team investigating a hypothesis called predictive processing.

Early neuroscience treated the visual system like a digital camera, taking in reality pixel by pixel. This is false. Under the predictive processing model, the visual cortex operates as a six-layer prediction engine. The brain predicts reality and uses raw sensory data to update its internal model.^1, 2

In canonical predictive coding models, deep cortical layers maintain an internal map of the world - the “what should happen”. Raw sensory data from your eyes enters via Layer 4 and moves to Layers 2 and 3.² These layers are thought to act as error-detectors, firing when there is a mismatch between reality and prediction.³ To conserve energy, the brain focuses on the mismatch, not each pixel. The mismatch is used to update the internal map, but also to start immediate action to prevent adverse outcomes.

Imagine walking down the street. Your motor cortex commands your legs to move forward and simultaneously sends a copy (efference copy) of that command to your visual cortex. The visual cortex predicts that the scenery should move backward as you step forward. This is what you’ve learned should happen. If reality matches the prediction - good game. ✨

If your visual flow suddenly shifts sideways (as it would if you moved your head to the side), mismatch neurons fire. The change in the visual flow without your conscious effort might be dangerous. The other times this happens is when you are actively falling. So your body executes a rapid adjustment of your limbs to prevent that from happening (finally, the years of training your proprioception with yoga have paid off!). It also updates your internal map - so you can prevent it the next time.

If you were walking down the same path and a bunny hops across, a mismatch would also happen. However, that one might not update your internal map. A fall threatens survival; a bunny does not (unless it’s the Monty Python one). The brain handles this through precision weighting, dialing up attention for high-stakes errors while filtering out background noise ^1, 2.

AI Can't Match Human Brain (Yet)

If you follow artificial intelligence, this "predict and update" loop should sound familiar. Self-supervised machine learning models rely on the same core logic: predict the next state, calculate the error against reality, and update the weights.

But brains do it better. And I am not saying this just because I am a neurobiologist.

Standard deep learning architectures typically rely on global backpropagation updates: when an error occurs, the entire network rewires itself. In the mammalian brain, nearly simultaneous activation of most neurons is rare. Activating all neurons is not efficient. Evidence suggests that biological brains use localized, event-driven learning: only the specific microcircuits involved in a mismatch update their local connections ^4, 5.

AI researchers are trying to replicate the energy savings of localized learning by combining local layer-wise goals (in software) with parameter-efficient adapters (like LoRAs) on neuromorphic hardware.^5, 6 However, running asynchronous, event-driven logic on Von Neumann chips is a classic square-peg - round-hole engineering problem.

While software scaling dominates current headlines, I believe that the next improvement in AI won’t be a better model. The real paradigm shift will be hardware capable of replicating biological precision weighting and local predictive coding.⁶

Products for humans

Predictive processing is a hypothesis on how we experience life.

In product, every strange layout, navigation, or interaction creates a mismatch. If the user needs more energy to figure out the workflow than the value of the product, the user closes the tab. The learning is always the same - keep core mechanics predictable, and use mismatches to capture attention.

Predictive processing is an explanation on how we perceive the world. It means that somewhere in your brain there is an internal map of all of the sunsets you’ve ever seen. When you stand on a beach and look at the evening sky, you are quietly seeing all of them at once.