Simultaneously, the team recorded a marked shift in network architecture. Connections within individual brain networks weakened, while connections between separate networks grew demonstrably stronger. The overall modularity of the brain—a parameter reflecting how neatly groups of neurons remain segregated into specialized, dedicated teams—dropped significantly across all four environmental settings of the experiment.
Furthermore, the study highlighted a striking collapse in the brain’s baseline sensory distinctions. Under sober conditions, brain activity in a person with closed eyes looks fundamentally different from brain activity when watching a movie. Under psilocybin, however, that distinct functional boundary largely vanished. Within the visual network, the connectivity gap between having eyes open and eyes closed shrank by 85 percent. Independent data from the subsequent EEG sessions confirmed this finding: alpha-band activity, which normally functions as a neural filter or gatekeeper restricting visual input when eyes are closed, was reduced by nearly half.
"When somebody takes a psychedelic and they’re able to close their eyes and see complex imagery, there seems to be less boundary between the internal and external world than we ordinarily experience," Stoliker explains.
On the surface, these primary analytical sweeps seemed to confirm the prevailing academic narrative: psychedelics appeared to induce a chaotic state characterized by blurred network boundaries, decreased modularity, and scrambled sensory processing. Yet when the researchers changed their analytical approach, that apparent chaos transformed into a highly structured pattern.
Hidden order
The reason previous neuroimaging studies frequently concluded that psychedelics cause brain chaos stems from two standard data-averaging practices deeply embedded in neuroscientific research.
The first standard practice involves averaging brain data over time. An eight-minute fMRI brain scan produces several hundred sequential whole-brain images. To evaluate how two distinct brain regions interact, scientists traditionally compress all of those sequential snapshots into a single averaged numerical score. This aggregate score reflects the degree to which the two regions synchronized across the entire eight-minute period. While computationally convenient, this averaging method risks obscuring brief, transiently structured patterns that assemble, perform a function, and disassemble within short windows of time, washing them out into what appears to be random noise.
The second standard practice involves averaging data across subjects. After computing a single summary metric for each participant, conventional methods pool all participants together into a single group average. This rests on the assumption that individual variations represent statistically irrelevant noise that will safely cancel out across a large sample size. Because the PsiConnect study was specifically tailored to explore personal, individual experiences, Stoliker’s team sought an analytical method that preserved individual variations and temporal dynamics.
Instead of relying on conventional temporal and group averaging, the researchers fed the real-time, moment-by-moment neural activity of 332 distinct brain regions into CEBRA—an advanced machine-learning algorithm designed to compress high-dimensional neural data into its core underlying structure while strictly preserving the temporal sequence of events.
For every participant, CEBRA constructed a dynamic neural trajectory: a path tracing through a three-dimensional mathematical space, with every point representing a specific moment during the brain scan.
"When they’re not under psilocybin, the brain activity is less differentiated by context," Stoliker notes. "But under psilocybin, we see that the activity becomes organized across time and more clearly differentiated by context."
Under psilocybin, the machine-learning trajectories did not dissolve into a disordered cloud of data. Instead, they cleanly separated into four distinct, highly organized clusters, each corresponding precisely to one of the four experimental contexts: rest, guided meditation, music, and the visual cloud video. The neural patterns were so clearly structured that a machine-learning classifier could analyze a single isolated moment of brain activity and accurately identify which of the four contexts the participant was experiencing at that instant.
Significantly, the machine-learning classifier’s ability to distinguish these contexts was not uniform across all subjects; its accuracy scaled directly with how profound and meaningful the participants subjectively reported their experience to be.
