Whether it’s a surprise birthday cake or an underdog team suddenly stealing the ball, moments that violate our expectations tend to feel startling—and according to new research, the brain may respond to them in a surprisingly consistent way.
Scientists at the University of Chicago built a whole‑brain “surprise network” that tracked how much a person’s expectations were violated and showed that the same pattern predicted surprise across three quite different scenarios: a laboratory learning game, the nail‑biting final minutes of college basketball, and short animated films that either followed or broke everyday rules. Their findings have been published in Nature Human Behaviour.
The project began with a simple puzzle that has long preoccupied psychologists and neuroscientists. Surprise is one of the most vivid human feelings, but past studies rarely agreed on whether different kinds of surprise—social, physical, or purely informational—share a single neural signature. Study authors Ziwei Zhang and Monica Rosenberg reasoned that the best way to answer that question was not to look at one brain region at a time but to examine how hundreds of regions pulse together, moment by moment, whenever expectations are upended.
To build their model, the researchers first revisited an open dataset collected at the University of Pennsylvania. Thirty‑two young adults lay in a brain scanner while playing an “adaptive learning” game. On each of 120 trials a cartoon helicopter hid behind the top edge of the screen and dropped a bag somewhere along a horizontal line. Before the drop the volunteers had to slide a bucket to the predicted landing spot.
Most of the time the helicopter stayed in roughly the same place, but now and then it jumped to a new hidden position. The size of that jump, combined with uncertainty about whether a jump had just happened, provided an objective yardstick for how surprising each outcome should be to an ideal observer. While the game unfolded, the scanner recorded blood‑flow changes across 268 predefined brain parcels.
Rather than calculate average connections over long intervals—a popular but slow‑moving approach—the team computed the “co‑fluctuation” between every pair of parcels at every single image frame. Each co‑fluctuation value captured whether two parcels’ signals rose and fell together in that instant. Feeding these high‑frequency edge traces into a leave‑one‑person‑out learning routine, the scientists identified two sets of connections. One set grew stronger when the modelled surprise level was high; the other did the opposite. Subtracting the average strength of the two sets yielded a single number each moment, the surprise network score, for each participant.
Having trained the network on the game data, Zhang and Rosenberg next asked whether it would forecast surprise when the context changed completely. A separate group of twenty volunteers watched the last five minutes of nine National Collegiate Athletic Association tournament games while undergoing functional magnetic resonance imaging at Princeton University.
For every possession change in each game an established sports‑analytics algorithm updated the home team’s win probability based on score difference, time remaining, possession, and team strength. Surprise was operationalised as the absolute change in that probability, but only when the change contradicted the prevailing belief about which team was likely to win.
When the researchers aligned these probability swings with the surprise network score, the two rose and fell together even after accounting for visual and auditory features of the broadcasts, subtle head movements, court position, and remaining game time. In other words, the connection pattern forged in a joystick‑based learning task signalled belief‑inconsistent moments during real sports viewing, despite the switch from an interactive setting to passive spectatorship.
To guard against the possibility that any large network might show the same property, the authors ran several control tests. Networks built from slower “sliding‑window” connectivity did not generalise. Networks made by averaging activity within established systems such as the default mode network or a well‑known sustained‑attention model failed to predict surprise in both directions.
Models based only on activity of individual parcels, without considering how they interact, also stumbled. Even edge patterns tuned to the players’ motor predictions or to the reward signal in the learning game did not capture the basketball surprise metric. These comparisons suggest that moment‑to‑moment coordination across widely distributed regions contains information that simpler summaries miss.
The team then flipped the analysis. They trained a new edge‑based network on the basketball data and tested it on the learning game. The flipped network again tracked surprise, confirming that the relationship was not tied to any one dataset.
Interestingly, the edges that overlapped between the two independently trained networks concentrated in similar anatomical territories. Connections linking visual and parietal areas, the medial and lateral frontal cortices, and limbic zones repeatedly showed stronger co‑fluctuations when expectations were broken. In contrast, connections within primary sensory and motor systems tended to decrease in strength at those moments.
A computational “lesioning” analysis—removing all edges tied to one functional system and rerunning the predictions—showed that eliminating links involving the frontoparietal control system or the default mode network markedly weakened performance, highlighting their importance for monitoring belief violations across contexts.
For a final challenge the investigators examined a third open dataset collected at the Massachusetts Institute of Technology. Twenty‑nine adults watched brief videos in which cartoon agents either behaved logically or violated everyday psychology, such as walking straight through a solid wall, alongside clips that violated simple physics with no agents present, like objects passing through each other.
The overlapping surprise network was significantly stronger when agents acted in unexpected ways compared with their expected counterparts, yet it did not distinguish between expected and unexpected physics clips. This outcome aligns with prior evidence that physical and social violations may rely on partly distinct brain mechanisms, and it hints that the surprise network is especially sensitive to belief shifts about intentions and actions.
While the findings knit together three very different experimental worlds, the authors acknowledge several caveats. The network explains only a modest share of the moment‑to‑moment variance in surprise, meaning that other unmeasured factors also play large roles. Sample sizes, especially in the sports‑viewing dataset, were small by population standards. The magnetic resonance technique used is not optimised for deep brainstem nuclei that are thought to broadcast surprise signals, so those contributions remain blurry. Finally, the study depended on computational definitions of surprise rather than on each participant’s subjective reports, an issue future work could address with real‑time ratings or eye‑tracking.
The study, “Brain network dynamics predict moments of surprise across contexts,” was published December 2024.