Two strangers arrive at the same graduate programme. They have not shared classes, late dinners, difficult assignments or the small disclosures from which friendship is usually built. Yet when each watches the same run of movie clips alone inside an MRI scanner, the timing of activity across parts of their brains already looks unusually alike.
Months later, they name each other as friends.
A 2025 study in Nature Human Behaviour followed that sequence through a real social network. Incoming students were scanned near the start of term, then their cohort’s friendships were mapped two and eight months later. Similarity in the brain responses recorded near arrival carried information about who would eventually be socially close.
It is an arresting result because it places the neural resemblance before the shared history. It is also easy to oversimplify. The final imaging sample contained 41 people, not the full cohort. The broad friends-versus-all-non-friends comparison was not significant. One future-friendship effect weakened after demographic controls, while the clearest and most extensive result came from comparing pairs who grew closer with pairs who drifted apart.
The paper does not reveal a friendship code or show that brain similarity causes two people to connect. What it offers is a carefully timed clue: some of the compatibility friends eventually discover may be detectable while they are still strangers.
Friendship has a selection problem and an influence problem
People resemble their friends in age, education, interests, habits, attitudes and countless other ways. Social scientists call the tendency for similar people to connect homophily. The old phrase “birds of a feather flock together” captures the selection side of it.
But friendship also creates similarity. Friends share information, places, jokes and social circles. They influence one another and accumulate common experiences. If two established friends respond similarly to a film, a one-time measurement cannot show whether they were alike before meeting or became alike afterward.
An influential 2018 Nature Communications study found that friends’ neural responses to naturalistic video clips were more similar than those of people farther apart in the same social network. The paper could identify friendship from neural response patterns above demographic similarities, but its cross-sectional design left direction unresolved.
The new project was built around that missing timeline. Measure neural response when a community is forming, wait while relationships develop, and only then ask whether early similarity anticipates the network that emerges.
The study caught a social network before it settled
Yixuan Lisa Shen, Ryan Hyon, Adam Kleinbaum, Christopher Welker, Carolyn Parkinson and Thalia Wheatley recruited 43 people from an incoming MBA cohort of 288 at a private university in the United States. The work was reviewed locally at Dartmouth College.
Most participants entered the scanner almost immediately after arriving on campus. The median interval between arrival and scanning was three days; the most common interval was one day. One participant did not complete the scan and another did not complete the final network survey, leaving 41 people, aged 25 to 34, in the main analysis.
Forty-one people may sound tiny beside a cohort of 288. For the pairwise analysis, however, they produced 820 unique dyads. Every scanned participant could be paired with every other scanned participant, and each pair had both a neural-similarity value and a later social distance.
“Before they met” was nearly, but not perfectly, literal. Eighteen of those 820 pairs reported some interaction before scanning, such as briefly encountering one another at a happy hour. The researchers reran key analyses without those pairs and observed a similar pattern. Two people scanned more than a month after arrival had been enrolled in different earlier graduate programmes at the university and reported not knowing each other.
Fourteen clips turned the scanner into a shared test
The participants did not answer friendship questions while being scanned. They watched videos.
Fourteen clips with sound were presented across six functional runs. They lasted from 88 to 305 seconds and spanned different styles and subjects. Everyone saw the same material in the same order, an experience the researchers compared to watching television while someone else channel-surfed.
The clips were selected to be relatively unfamiliar, engaging and capable of producing meaningful differences between viewers. A straightforward scene can make most people attend to the same event at the same moment. More ambiguous material creates room for viewers to notice different details, draw different inferences or react with different emotional timing.
Functional MRI does not record neurons firing directly. It tracks the blood-oxygen-level-dependent, or BOLD, signal associated with changes in local brain activity. The team divided each brain into 200 cortical and 14 subcortical regions. For every pair of participants, it correlated the response time series in each region across the clips.
If the signal in a region rose and fell in a more similar rhythm for two viewers, their inter-subject correlation was higher. This does not mean the scanner recovered a shared sentence from their minds. It means that, while receiving identical audiovisual input, the regional responses unfolded more similarly over time.
After scanning, participants separately rated how interesting and enjoyable they found each clip. Those simple preference ratings later helped test whether the brain result was merely a complicated way of saying that future friends liked the same videos.
Friendship required two people to name it
The social side of the study was unusually complete.
About two months after scanning, all 288 members of the cohort completed an online network survey. At eight months, 287 did so, a 99.6 per cent response rate. Students saw a roster and selected classmates with whom they most often spent free time on informal activities such as meals, drinks, films or visits to one another’s homes. They could select as many names as they wished.
The analysis counted a direct friendship only when both people nominated one another. One-sided nominations were not treated as edges in the network. Direct friends had a social distance of one. Two non-friends linked through a mutual friend had a distance of two. Pairs connected through two intermediaries had a distance of three.
This is stricter than asking one participant whether someone feels like a friend. It also means the headline’s “who would be friends” refers to reciprocal reported ties inside this particular cohort, not every meaningful relationship in each person’s life.
By the eight-month survey, the 820 pairs in the imaging sample included 93 direct-friend dyads, 544 pairs at distance two and 183 at distance three. The network was not a loose collection of occasional classmates. Students lived near one another in a rural setting, took classes together, ate together and had many opportunities to connect.
The eight-month result was real, but narrower than a brain-based friend detector
The first comparison bundled all non-friends together. In that test, future friends did not show significantly greater initial neural similarity than non-friends across the 214 regions after correction for multiple comparisons.
Future friends also did not differ significantly from friends-of-friends. The reliable contrast emerged at the widest distance tested: direct friends had shown greater pre-existing similarity than pairs three degrees apart in a portion of the left orbitofrontal cortex, or OFC.
The OFC is involved in representing subjective value. In this setting, a synchronized response there might reflect related tastes, valuation or affective reactions. That interpretation remains tentative. Brain regions do many things, and the study did not manipulate one psychological process at a time.
Similarity in reported enjoyment and interest did not explain away the OFC difference. Yet demographic adjustment changed the picture. When the researchers controlled for pairwise similarities in age, gender, nationality, hometown, undergraduate education, major and prior industry, the OFC result no longer survived false-discovery-rate correction across all regions. Gender similarity contributed significantly to that reduction.
That does not make the original contrast meaningless. It shows why “their brains matched, therefore they became friends” is too clean. Neural response itself can carry a history of background, culture and experience. The scan may summarize compatibilities produced by those forces rather than reveal an independent biological magnet.
Growing closer versus drifting apart produced the broader result
The study’s second analysis asked a different question. Instead of using only the final eight-month map, it examined how each pair’s social distance changed between month two and month eight.
Of the 820 pairs, 279 moved closer in the network, 445 stayed at the same distance and 96 moved farther apart. These changes could reflect a friendship forming or dissolving, but also ties changing elsewhere in the network. A pair can move closer because one person becomes friends with someone in the other’s circle.
When people who grew closer were compared with everyone who did not, no brain region showed a significant difference. The closer group also did not differ significantly from pairs whose distance remained unchanged.
The sharp contrast was between opposite trajectories. Compared with the 96 pairs who drifted apart, the 279 pairs who grew closer had entered the programme with greater neural similarity in 40 cortical regions, both thalami and the left amygdala. The cortical regions extended across visual and ventral temporal areas, superior parietal cortex, angular gyrus, medial frontal cortex and lateral prefrontal cortex.
Those areas span systems associated with visual processing, attention, social interpretation and cognitive control. Similarity was therefore not confined to whether two people appeared to enjoy one clip. It involved the timing of how they allocated attention and integrated a changing narrative.
The closer-versus-apart pattern remained broadly similar after controlling for the measured sociodemographic variables and after removing pairs with pre-scan interaction. Adjusting for similarities in enjoyment or interest ratings also did not generally erase it, although enjoyment accounted for part of the effect in one superior parietal region.
A synchronized movie response is not a shared thought
“Neural similarity” can sound more mystical than the measurement warrants.
Every participant received the same light, sound and scene changes, so some synchronization is expected. The analysis looked at relative differences: which pairs tracked one another more closely than other pairs in each region. Those differences may reflect shared attention, emotional cadence, prior knowledge, interpretation or expectations.
They may also reflect factors the researchers did not measure. Language, politics, religion, personality, family environment, genes and past media exposure could shape both a person’s response to a film and the people they later find easy to know. The public data and analysis repository supports scrutiny of the processed results, but no dataset can adjust for every influence on a life.
The study also did not train a tool that could scan a new individual and name their future friends. Its inferences came from group-level comparisons among pairs. The final sample was 41 self-selected imaging volunteers from one immersive MBA programme, and its smallest trajectory group contained 96 interdependent dyads rather than 96 independent pairs of strangers.
Pairwise data are not equivalent to hundreds of unrelated experiments because each person appears in many dyads. The researchers used participant-level permutation tests and corrected results across 214 regions to address those dependencies and multiple comparisons. Those safeguards make the evidence more disciplined, not infinitely generalizable.
Friendship may select similarity and then deepen it
The setting provided a useful natural experiment without becoming a randomized friendship trial. Students were assigned to study groups through stratified random sampling, and much of the cohort’s housing was allocated by lottery. That reduced some freedom to choose who would be physically nearby.
Still, nobody was randomly assigned to become friends. The study was observational, and its authors explicitly stopped short of decisive causal claims. Similar neural processing may facilitate connection, but both may arise from shared demographic and cultural experience. Chance meetings and network position still matter.
The ideal next step would scan people again after their relationships matured. If friends began similar and then became even more aligned, researchers could observe selection and influence operating in sequence. The team planned another imaging wave two years later, but the COVID-19 pandemic prevented it.
So the study closes one part of the chicken-and-egg problem while leaving another open. It shows that at least some neural resemblance preceded later friendship and movement through the network. It does not show how much similarity was already present for a lifetime, how much the early months added, or whether the friendship later amplified it.
The result is most compelling when kept at human scale. Before two people exchange a joke, one may already be attending to the setup the way the other does. Before they discuss a difficult scene, their emotional responses may already be changing at a similar pace. Neither person can see that quiet alignment.
Conversation still matters. Proximity still matters. Trust, effort and accident still matter. The scan does not replace those ingredients. It suggests only that when friendship finally begins, the feeling of being on the same wavelength may sometimes be less metaphorical, and less newly made, than it appears.