The Brain Starts Making Decisions Earlier Than Anyone Realized. It Could Change How We Build AI.
Neuroscience has operated on a tidy assumption for decades: sensory information travels upward through the brain, gets processed into progressively more abstract forms, and eventually reaches the frontal cortex, where a decision is made. Early regions sense. Later regions decide. Clean, orderly, hierarchical.
A new study from the University of Illinois Urbana-Champaign says that picture is wrong, or at least seriously incomplete.
What the researchers actually found
The team, led by electrical and computer engineering professor Yurii Vlasov, recorded neural activity in mice navigating a tactile virtual reality setup. They left the animals with only a single matched pair of whiskers, the C2 whiskers, which forced all sensory input through an extremely narrow channel. This was deliberate: it created what the researchers call an information bottleneck, so they could track exactly where decision-related processing happened first.
What they found in the primary somatosensory cortex (S1), the brain region responsible for basic touch perception, surprised them. Decision-related activity was showing up in S1 well before information would have had time to travel up to the frontal cortex and back. More importantly, S1 was not generating this activity on its own. Higher brain regions were sending feedback signals back down into S1, actively shaping what it encoded during the decision window.
The paper, published in the Proceedings of the National Academy of Sciences, puts it directly: "S1 appeared to be dynamically modulated by top-down regulation, engaged by the higher-level brain regions via feedback loops, suggesting that decision-making is not solely relying on unidirectional feed-forward processes as previously thought."
Why this matters for artificial intelligence
Current AI systems, including the convolutional neural networks used in computer vision and many large language models, are built on a feed-forward architecture. Information goes in, moves through layers in one direction, and a prediction comes out at the end. That design was borrowed, explicitly, from how neuroscientists believed the brain worked.
If the brain actually runs on bidirectional feedback loops, with information flowing down as well as up throughout the decision process, then today's AI may be missing something fundamental. "We want to learn from a billion years of evolution," Vlasov said. "By looking at the fast temporal dynamics of neural activity, maybe we can understand better how these feedback loops are engaged in making decisions."
The practical implication he points to: AI architectures that incorporate similar feedback loops could be more capable and far more energy-efficient than current systems. Biological intelligence handles remarkably complex tasks at a fraction of the power cost of modern neural networks. Understanding exactly why is an open engineering question, and this study offers one concrete place to start.
The limits of this finding
The experiments were conducted in mice, and recordings were limited to a single cortical column. Whether the same dynamics apply to human touch, vision, or other senses has not been established. The feedback from higher brain regions was inferred from patterns in S1 activity rather than measured directly in those regions. Vlasov and his co-author, neuroscience doctoral student Alex G. Armstrong, are careful to note these constraints.
Still, the finding settles a running argument in the field. Choice-correlated signals in the primary sensory cortex had been observed before but often dismissed as movement noise or experimental artifact. The Illinois team's tight information bottleneck design makes those explanations hard to sustain. Decision-related processing appears to begin at the brain's earliest sensory stages, which means the hunt for where perception becomes choice needs to start much earlier than most models assumed.
For AI engineers, the question now is how much of this feedback architecture can actually be built into the next generation of systems, and whether doing so will close the enormous gap between biological and artificial intelligence on energy efficiency. The gap right now is not small: a human brain runs on roughly 20 watts. A frontier AI training run can consume as much as a small power plant.