Fig. 4: Impact of bot behavior and decoy landscape on similarity of answers to the target noun and to the decoy noun.

a Cosine similarity between nouns shared by bots and the target noun in each game. A total of 125 groups went through three games under different bot conditions for a total of 375 data points. The regression model’s independent variables included fixed effects for the three bot conditions (random, least, most), the landscape variables, and their interactions, with the reference variables being the most-similar bot condition and the tall/wide landscape. The solo and no-bot conditions were excluded because these conditions did not have bots. We see that the most similar bot (red) helped participants identify high-value nouns by propagating them through the network. b Cosine similarity between nouns guessed by participants and the decoy noun. The data from the no-decoy landscape was excluded, resulting in 500 data points. The regression model’s independent variables included fixed effects of bot conditions, landscape variables, and their interactions, with the reference variables being the most-similar bot condition and short/wide landscapes. The most-similar bot (red) did not pull the group toward the decoy local optimum in the same way that it moved the group toward the global maximum at the target. The plots show summary statistics of the raw data presented with box-and-whisker plots. The box represents the interquartile range (IQR). The line within the box represents the median value. The upper (lower) whisker extends from the hinge to the largest (smallest) value no further than 1.5 * IQR from the hinge. Data outside the whiskers are plotted individually.