Groups are often assumed to perform better because they can combine the knowledge of several people.
But simply placing people together does not automatically make a group intelligent.
New research suggests that one of the most important ingredients may be knowing which group members are actually doing well.
A study published in Nature Communications on 23 September 2026 tested 621 adults in an immersive 3D environment where they had to search for and track a hidden moving resource to maximise a monetary reward.
Participants completed the task either alone or in groups of five. In some group conditions, participants could see only where the others were moving. In another, they could also see how well each person was currently performing.
The groups outperformed solitary participants when payoff information was fully visible.
Without that information, being surrounded by other people was not reliably beneficial and could even reduce performance when the hidden resource moved quickly.
The researchers argue that visible success allowed participants to practise what they call payoff selectivity: following social information more strongly when another person was currently performing better than they were.
The challenge was to find something that never stopped moving
The experiment was designed to make social learning more dynamic than a conventional laboratory choice task.
Participants entered a first-person 3D environment and controlled an astronaut-like avatar inside a large circular arena.
Somewhere in the arena was an invisible moving resource.
The closer a participant moved to its centre, the more points they collected.
Each avatar had a detector showing whether the participant was moving closer to or farther from the resource.
The task therefore required a continuous balance between using private information from the detector and social information from the movements of other people.
The resource itself did not remain still.
Participants were assigned either to a slower resource moving at 40% of their avatar’s speed or a faster resource moving at 80% of their speed.
The faster condition made previously useful information become outdated more quickly.
The researchers created four different social conditions
The final sample contained 621 adults aged 18 to 65.
Each participant completed 15 minutes of the task in one of four conditions.
In the Alone condition, participants searched for the resource by themselves.
In the No Payoff-Sharing condition, participants worked in groups of five and could see the locations and movements of the other players, but not how many points they were currently earning.
In the Full Payoff-Sharing condition, participants could also see a visual indicator showing the current payoff of other group members.
A fourth, exploratory condition allowed participants to decide voluntarily whether to share their current payoff, at a small cost to themselves.
Combining these four social conditions with the slow and fast resources created eight experimental conditions.
Seeing where people were was not enough
The most important contrast was between seeing another person’s behaviour and seeing whether that behaviour was successful.
A participant could observe another player moving in a particular direction.
But movement alone does not reveal whether that person has discovered the resource, is following a bad guess or is simply exploring.
When payoff information was hidden, participants had access to social cues without a reliable way to judge their quality.
This could make social information actively misleading.
The problem became especially important when the resource moved quickly because another person’s location could become obsolete within a short period.
Under those conditions, following others simply because they were visible sometimes reduced performance compared with searching alone.
Full payoff information produced the strongest group performance
When the resource moved slowly, the groups with full payoff information collected substantially more points than solitary participants.
The researchers estimated an average difference of about 5,236 points between the Full Payoff-Sharing and Alone conditions in the slow-resource environment.
Full payoff sharing also outperformed the group condition in which participants could see one another but could not see current performance.
The voluntary-sharing condition generally fell between full information and no payoff information, although it did not consistently differ from the weaker conditions.
The result suggests that the benefit was not simply having more people searching the environment.
The advantage appeared when people had enough information to distinguish useful social cues from unhelpful ones.
Participants learned whom to follow from current success
The researchers analysed movement at a second-by-second level.
This allowed them to see not only where participants eventually ended up, but how their decisions changed as new information became available.
When payoffs were visible, participants responded more strongly to group members who were currently closer to the resource than they were.
They used both the successful person’s location and their movement direction.
When another participant was performing poorly, their behaviour was given less weight.
This selective use of social information is what the authors describe as payoff selectivity.
The group reorganised its attention when someone found the resource
The first-person environment also meant participants could not see every other group member at all times.
The researchers therefore constructed dynamic visibility networks showing who could see whom at each moment.
When someone discovered the resource and their payoff was visible, other participants were more likely to orient toward that successful player.
The network of attention effectively reorganised itself around the person who currently had the best information.
This helped the group converge on the resource more quickly during the search phase.
Without payoff information, discoveries were much easier for the rest of the group to miss.
The same process helped once the resource had been found
Finding the resource was only half of the problem.
Because it continued moving, participants also needed to remain close to it.
Visible payoff information allowed participants to use other players as real-time indicators of where the resource might be moving.
If another player was currently earning more points, following their direction could be more informative than relying only on one’s own recent detector readings.
Participants with access to full payoff information therefore achieved higher tracking efficiency than solitary participants.
The advantage came from dynamically deciding when social information was better than private information rather than simply copying the crowd.
Blindly following a crowd can make performance worse
The study illustrates why social learning can be both useful and dangerous.
Other people provide information, but their behaviour is not automatically correct.
If several people follow the same mistaken person, a group can amplify an error rather than correct it.
This is one mechanism behind information cascades and herd behaviour.
In the experiment, the No Payoff-Sharing condition created exactly this risk.
Participants could see where others were going but had no direct indicator of whether those people were successful.
The faster the environment changed, the more damaging low-quality social information could become.
A changing environment makes old social information less useful
Environmental volatility played an important role in the results.
When the hidden resource moved quickly, the position of another participant was a weaker guide to where the resource would be a few moments later.
In a stable environment, copying someone’s location can remain useful for longer.
In a rapidly changing environment, successful decision-making requires more up-to-date information.
Visible current payoff provided that quality signal.
This is why the distinction between simply seeing others and seeing whether they are succeeding became especially important in the fast-resource condition.
The researchers also built a computational model of movement decisions
The behavioural results were supported by computational modelling.
The model estimated how private information and several social features influenced the direction participants chose from one second to the next.
The analysis showed that visible payoff allowed people to tune their responses selectively toward better-performing peers.
Participants also adjusted how they used information when several people were visible at once.
For example, the direction of a cohesive group could become more informative than the group’s exact position.
The result suggests that human social learning is not simply a switch between “copy others” and “think independently”.
People continuously reweight different private and social cues according to how useful those cues currently appear to be.
Simulations tested whether payoff selectivity itself was driving the advantage
The researchers then created agent-based simulations based on the mechanisms identified in the experiment.
Simulated agents could rely on private information or social information depending on the quality of the cues available.
High payoff selectivity meant an agent would rely strongly on social information mainly when another agent was currently doing better.
The simulations reproduced the central pattern: selective attention to successful peers improved collective outcomes under the right environmental conditions.
They also showed that the advantage is not universal.
Whether payoff-selective social learning helps depends on how quickly the environment changes and how difficult it is for individuals to integrate multiple streams of information.
The experiment did not show groups beating the very best individual
The phrase “collective intelligence” can be used in different ways.
In this study, groups with full payoff information performed better on average than people working alone.
However, the experiment did not provide evidence that the groups exceeded the performance capacity of the most skilled individual participants.
The authors explicitly describe this as a stricter definition of collective intelligence that their behavioural experiment did not meet.
They suggest that the demanding task required participants to monitor their own detector, navigate the environment and process multiple social cues at the same time.
When some of those cognitive constraints were relaxed in simulations, the same mechanism could produce group performance beyond even the strongest individuals.
That simulation result is useful for understanding the mechanism, but it should not be confused with something directly demonstrated by the human participants.
The voluntary sharing condition was exploratory
The condition in which participants chose whether to reveal their own payoff deserves separate caution.
It was added as an exploratory condition and was not part of the researchers’ original preregistered predictions.
Sharing cost one point per second, creating a small incentive to keep information private.
Performance in this condition tended to sit between full payoff sharing and no payoff sharing.
However, the differences were not consistently reliable.
The researchers also showed that simply adding the signalling costs back to participants’ scores did not eliminate the overall pattern.
This suggests that the weaker performance was not only caused by the small financial cost of sharing.
The study used an artificial task, even though it was more realistic than many lab experiments
The experiment was intentionally designed to be more naturalistic than a flat-screen decision task.
Participants had a limited first-person field of view, moved through a 3D space and continuously balanced exploration with social learning.
But it was still a virtual game conducted online.
Searching for a hidden digital resource is not the same as deciding which colleague to trust during a workplace crisis, following investment advice or coordinating during an emergency.
Real social groups also involve status, reputation, friendship, communication and unequal expertise.
The experiment deliberately controlled many of those factors to isolate how visible performance changes social learning.
The findings may still apply to many real group decisions
The basic problem studied in the experiment appears in many real settings.
People often observe what others are doing without knowing whether those people have better information.
An employee can see which strategy colleagues are following but may not know which strategy is actually producing better results.
An investor can observe what is popular without knowing whether the crowd is informed.
A team can copy the most visible person rather than the most successful one.
The new study suggests that groups become more adaptive when they can distinguish social activity from social success.
Transparency about outcomes may therefore matter as much as transparency about behaviour.
More social information is not always better
The findings also challenge the assumption that giving people more access to what others are doing will automatically improve group decisions.
Additional social information can increase noise.
If people cannot distinguish good signals from bad ones, they may overreact to visible behaviour and neglect their own useful information.
The value of social information therefore depends partly on whether its quality can be evaluated.
This is especially important in fast-changing environments, where yesterday’s successful behaviour may no longer be useful today.
The most adaptive groups may not be those that copy the most.
They may be the ones that are best at deciding when and whom to copy.
The bigger lesson is that collective intelligence needs a quality filter
Groups contain more information than individuals, but they also contain more opportunities for distraction and error.
The new experiment shows that the difference between useful coordination and maladaptive herding can depend on a simple piece of information: whether the people being observed are currently succeeding.
When participants could see that signal, they redirected attention toward better-performing peers, reorganised social information flow and improved both search and tracking.
When the signal was absent, social behaviour could become noise rather than guidance.
Collective intelligence therefore may depend less on how much people copy one another and more on whether they have a reliable way to identify which information deserves to spread.
Source Information
Study Title: Payoff selectivity drives collective intelligence during dynamic resource tracking in humans
Authors: Valerii Chirkov, Ralf H. J. M. Kurvers, Pawel Romanczuk and Dominik Deffner
Journal: Nature Communications
Published: 23 September 2026
Sample: 621 adults aged 18–65 in the final analysed sample, recruited online through Prolific. Participants completed the task alone or in groups of five across eight experimental conditions combining four payoff-information conditions with slow or fast resource movement.
Method: Participants completed a 15-minute first-person 3D resource-tracking task in which they searched for a hidden moving target to maximise points and monetary reward. The researchers compared solitary performance with group conditions offering no payoff information, full payoff information or voluntary payoff sharing, and analysed movement trajectories and dynamic visibility networks using hierarchical Bayesian models, computational movement models and agent-based simulations.
Main finding: Groups outperformed solitary participants when full payoff information was available because participants selectively relied on successful peers and dynamically reorganised social attention. Without payoff information, social cues could become maladaptive, particularly when the resource moved quickly. The human experiment did not show groups exceeding the very best individual performers under the strictest definition of collective intelligence, although simulations showed that the identified mechanism could produce such outcomes under less cognitively demanding conditions.
DOI: 10.1038/s41467-026-77762-z
