A total of 96 subjects in 8 12-person sessions, split across two treatments, interacted as avatars in 35 90-s periods (representing days; 75 s of day (including 5 s of dusk) and 15 s of night). Their goal was to earn as many points as possible, which were converted into US Dollars (at a 1:1 ratio) at the end of the experiment. Avatars were numbered and color coded so that individuals could identify one another. During the day, avatars could earn points that were directly converted into cash earnings by foraging for one of two types of food (“fruit” and “grass”; see below for details) and participating in a generic social interaction that was a proxy for beneficial social engagement. Fruit was high value but replenished slowly (never within the same day), and was always scarce, whereas grass was low value but infinitely renewable, that is, it was continuously available at the site it appeared at each day. The social interaction was labeled “health” for the participants but hereafter we refer to it as “grooming”, for it represents all directional social interactions that provide a direct benefit to one other avatar at a time. Because grooming was equally important to earning points in both conditions, it was not useful for measuring differences in sociality between the two. At night, remaining stationary in nests (all extant apes exhibit such nesting behavior) increased points. (See the supplementary online material for our precise language. For example, we did not use the words grooming, chimpanzee, or bonobo.)
In both conditions, the world was a rectangle with two “groves” of trees, one in the north and one in the south, which was designed to make it costly for avatars to congregate around a single supply of fruit, as apes in the naturally occurring world must search out fruit from dispersed groves. The amount of fruit was equally distributed between northern and southern trees, and grass was randomly distributed throughout the world so that there was no caloric incentive to prefer one area of the world over another. Fruit trees remained in the same location, but flowered and bore fruit in a cyclical pattern. Fruit was thus not available on each tree each period, but avatars could predict that it would be available in a day or two based on the flowering. Moreover, once a fruit was eaten in a given period, it was no longer available. Avatars could not guard fruit or exclude others from a tree. The location of grass changed each day as well, so subjects could not obtain enough food without moving, but within a day the grass continuously renewed and multiple individuals could feed on the same patch at the same time. The aggregate amount of food was held constant between Chimpanzee and Bonobo conditions. There was three times as much fruit per day (120 vs 40 pieces) in the Chimpanzee treatment vs the Bonobo treatment, but it took three times as long to forage on grass in the Chimpanzee treatment. Note that this was not meant to reflect naturally occurring handling times, but provided a way to incentivize different food choices while keeping the rate of food consumption the same across conditions.
Randomizing the location of the grass around the world and having trees fruit at different times made the problem of forming and maintaining groups nontrivial. In other words, before conducting the experiment we did not know if our design choices would induce any grouping behavior. The virtual environment was sufficiently large relative to avatar speed that it took 22% of the day to walk between the two groves of trees. Consistent with foraging in a forested environment, subjects could not see the entire world, but only a limited range around them. A map in the upper left corner of the screen displayed their location as well as the location of the trees (but not whether they were fruiting), which was designed to be a proxy of the mental maps apes have of their environment40. Subjects could call to one another over a greater distance and tell from what direction others’ calls emanated.
Finally, subjects, at a severe potential cost to themselves, could also individually attack a lone outsider, explicitly termed a “pirate”, who ate the fruit, but not less valuable grass. If one avatar attacked the pirate, the avatar incurred a significant cost and the pirate continued eating fruit. Subsequently, any avatar within the viewing window received a message indicating the outcome of two simultaneous attacks. If two avatars attacked the pirate, neither incurred a cost, and the outsider would leave for the rest of the day only to return the next day. Likewise, nearby avatars then received a message explaining three simultaneous attacks: if three or more group members attacked the pirate, it was “killed” and did not return in future days, although unannounced to the participants, there were a total of three pirates in each world; if all three were killed, no additional pirates appeared. Note that we intentionally made a solo attack extremely costly because solo attacks are not reported in the wild. However, we did not disallow solo attacks because one of our goals was to see whether such behavior emerged endogenously. In addition, this approach required minimal instruction and no explicit rules restricting behavior. This latter point was extremely important, as our goal was to see how people would explore the space and what decisions they would make without instruction, which could bias their subsequent decisions. An online video (https://www.youtube.com/watch?v=i0o_9nf2wwc) illustrates the subjects’ tasks in the virtual world and provides the experimental context.
Given that events during the day occurred in real time at the discretion of the participants, and may depend on idiosyncratic social temperaments, a daily pattern of the events was ex ante unpredictable. Our first result establishes the consistency of behaviors across four different sessions of a treatment in response to the biological imperatives we induced in the experiment. In Figs. 1 and 2, we report the total number of grooming, grass foraging, and fruit foraging events over the course of a day (summed over all 35 days) for each session in the Chimpanzee and Bonobo treatments, respectively.
The results in Figs. 1 and 2 indicate a consistency with which the sessions replicated a daily pattern in the two different ecological environments. Such consistency in an experiment with a relatively unstructured decision space indicates that we have created an environment in which the participants responded to the incentives we presented. In other words, we appear to have designed an experiment such that rewards of the experiment (the money they earned for their choices) were high enough to maintain the attention of the participants, i.e., “the reward structure dominates any subjective costs (or values) associated with participation in the activities of an experiment” (41, p. 934).
One key design goal of our virtual environment was that the virtual worlds contained the same amount of total food even though the treatment conditions varied the amount of the fruit and processing time for grass. This goal was achieved; the average, maximum, and minimum earnings for all participants were very similar for the Bonobo and Chimpanzee treatments—respectively, US$15.98 (s.d. = $9.23) vs. US$16.29 (s.d. = $8.00), US$27.87 vs. $27.53, US$3.00 vs. US$2.85—indicating that the environments, by design, were indeed equally challenging for the participants. There was no significant difference in average session earnings (Mann–Whitney U4,4 = 8 > critical value = 0, α = 0.05, two-tailed test). Nonetheless, we observed differences between the treatments (see Figs. 1 and 2). The hominoids in the Chimpanzee treatment spent the earliest part of the day (15 s) foraging for fruit, followed by a slow sustained increase in grass foraging and a variable, but a flat rate of grooming. In the Bonobo treatment, hominoids quickly increased their grass foraging over the first half of the day (40 s) and then spent the rest of the daylight time (35 s) grooming. Consistent with the different ecological inducements, Bonobo hominoids spent very little time foraging for fruit as compared to their Chimpanzee counterparts, and Chimpanzee hominoids spent much less time foraging for grass. While there were subtle differences in the patterns of daily events within a treatment (some social groups groomed more than others as compared to other sessions in the same treatment condition), the data in Figs. 1 and 2 visually indicate that Chimpanzee sessions were more similar to each other than they were to Bonobo sessions and vice versa.
The nesting locations of the avatars indicated with whom the avatars concluded their day’s activities and with whom they began the next day; this was our measure of social affiliation since it earned no points for social partners (like grooming did) and was therefore a measure of subjects’ endogenous affiliation choices. If all 12 avatars decided to nest, there were 12C2 = 66 combinations of unique distances between the avatars. To quantify the avatars’ proximity to one another at the end of a day, we summed the unique distances between all avatars who chose to nest. As some avatars occasionally decided not to nest (and instead stood in place or walked around), we divided the sum by the actual number of nest combinations for that day to ensure the distance measure was comparable across days. (For example, if only 10 avatars nested in a day, there are only 10C2 = 45 distances between 10 avatars that day). Figure 3 illustrates the nesting proximity of avatars by day, with sessions represented by dashed lines and treatment averages across all sessions represented by solid lines (orange for Chimpanzee, blue for Bonobo). Lower numbers indicate closer nesting proximity within the session. The trendline for the Bonobo average is decreasing (− 40.6 pixels/day) at a statistically significant rate (p-value < 0.0000). The trendline for the Chimpanzee average is slightly increasing (8.4 pixels/day) but is statistically insignificant (p-value = 0.0770).
Changing the distribution, availability, and processing time of food resulted in distinct patterns of association in these equally challenging environments. For the first third of a session, the hominoids in the two treatment conditions were roughly equally close to (or evenly dispersed from) each other in the virtual world. Beginning on day 15, the Bonobo hominoids began to associate in one larger group of 9 to 12 avatars while the Chimpanzee hominoids separated into northern and southern subgroups of 3 to 6 avatars (a group was considered any avatars who were in same overlapping range of vision). The average measure of proximity by session for all 35 days was significantly lower in Bonobos than Chimpanzees, (Fig. 3: Mann–Whitney U4,4 = 0, p-value = 0.05, two-tailed test). Because the amount of fruit was equally distributed between the northern and southern trees, splitting into two subgroups was the most efficient way for hominoid avatars to forage in the Chimpanzee treatment. The low variance of proximity in the Chimpanzee sessions indicated that half of the avatars spend their days and nights by the north trees and the other half by the southern trees.
In the two Bonobo sessions that showed the closest proximity in nesting by the end of the experiment, 83% to 100% of the avatars gathered to nest as a single group within the same viewable area (in the wild, bonobos form larger nesting aggregations42 and generally show closer proximity than chimpanzees43). When the patch of grass randomly moved the next day, this large group moved together to find a new patch. Such collective movement among avatars was surprising and unexpected given our incentive structure. Further, large aggregations are not necessary for the Bonobo hominoids to successfully feed themselves and socialize. Two or even three subgroups spread out over the virtual world would be sufficient to maximize one’s intake of grass and to be maximally groomed each day (for details on the benefits of grooming, see the Methods/Experimental Design section below). However, presumably because they do not need to compete for limited fruit, the avatars in the Bonobo condition tend to form one large group visible to everyone. As soon as they were satiated on grass, they continuously groomed each other, which is the only overt prosocial interaction that could be expressed in the world.
The other two Bonobo sessions also have stable cohesive groups of 6 and 9 avatars, which, by the end of the session, spend their entire days together. In the former session, the remaining 6 avatars form stable groups of 4 and 2 avatars. In the latter session, the remaining 3 avatars tend to wander about on other own in the same neighborhood, but they do not stay in continuous view of each other like the 11-, 12-, 6-, and 9-avatar groups above do.
Having established consistent foraging and grooming behaviors within both treatment conditions and having found distinct social grouping behaviors between the two treatment conditions, our final analysis tested whether avatars in the two treatments would respond differently to outgroup members. A single pirate would appear each period unless a total of three pirates had been killed, in which case pirates never appeared again for the remainder of the session. Indeed, Chimpanzee hominoids more frequently attacked the pirate avatar than Bonobo hominoids did. Figure 4 reports the number of single attacks, attacks as pairs, and attacks as triplets on the pirate avatar.
Hominoids in both treatment conditions attacked the outgroup pirate at the beginning of the session, but they rarely killed it with three attackers. Chimpanzee hominoids continued to attack the pirate throughout the second half of the session (76 total attacks in the first half and 70 attacks in the second half), and in three of the four sessions they killed at least one pirate with three joint attacks. Avatars in two of the Chimpanzee sessions killed all three pirates, permanently ridding themselves of the rival foragers. Bonobo hominoids, on the other hand, substantially reduced their attacks (43 total attacks in the first half and only 13 in the second half) and never jointly attacked the pirate in the second half of the session. There were no attacks in the fourth Bonobo session for the last half of the session. Using the total number of attacks by session for all 35 days, Chimpanzees attack significantly more frequently than Bonobos (Mann–Whitney U4,4 = 0, p-value = 0.05, two-tailed test).
These results demonstrate that differences in the value and distribution of food can drive differences in both within-group cohesion and out-group behavior, such that when food abundance is held constant, virtual hominids in an environment in which high-value food is centralized in ephemeral clumps form smaller social groups and attack outgroup members more than virtual hominids in an environment with widely distributed low-value food. Of course, model systems by design highlight specific features for hypothesis testing, but these results suggest that this difference could have led to the actual differences observed between chimpanzee and bonobo social structure. While ecology is not the only difference between the species, our results support one widely accepted hypothesis suggesting that the initial divergence was based on geographic separation rather than behavioral differentiation3, and that these ecological differences then led to the differing social patterns. Indeed, it is reasonable to hypothesize that this became self-reinforcing, with less need to defend food resources and increased opportunities for group cohesion working in tandem to select for decreased aggression and increased social tolerance (i.e.27). Moreover, the coherence and consistency of avatar behaviors across sessions, with very minimal instruction to the participants or experimental constraint, suggests that these virtual worlds are a robust method for testing questions about the role of ecology and, potentially, social structure (i.e., by changing the costs and benefits of social interaction) on primate decision-making. In particular, this method may allow for experiments that are impossible or unethical to conduct in wild primates and difficult to model due to unknown parameters, thereby providing new opportunities for advancing our understanding of the evolution of primate sociality.