Scientists Find Why AI and Brains Need 'Messy' Extra Wiring

A team of neuroscientists trained small artificial neural networks — computer systems loosely modelled on the brain — to find rewards. The networks had to track a continuously changing speed, a bit like a rat chasing a moving food source. Some connections were 'low-rank', meaning simple, tidy and few. Others were 'high-rank', meaning many separate, crisscrossing lines of communication between the artificial neurons.

The tidy connections did most of the visible work. They carried a rough map of where the reward was, and the researchers could nudge those signals to change what the network 'decided' to do. But when they broke the network apart, they found something surprising: the messy, high-rank connections were quietly correcting small errors in that rough map. Remove them, and accuracy fell apart as soon as conditions changed.

That matters because our own brains are full of exactly this kind of apparently wasteful wiring, and neuroscientists have long wondered why biology bothers. This study suggests the extra connections work as a safety net — a scaffold that keeps a fast, rough system precise and lets it cope with unfamiliar situations. The same lesson may apply to AI: models built only from clean, simple patterns can look fine in testing and crack in the real world.

The catch is scale. This is a computer simulation of a small network solving a simple navigation task — not a scan of a living brain, and not a test of a chatbot. It's a clue about a principle, not a recipe. Still, it adds to a growing argument that the mess in biological brains isn't a flaw, and that AI designers might want more of it, not less.