Fig. 5.

Explanatory results from the culminating user study. a The speech profiles of participants and S#, which shows the average number of times that people and S# used messages of each type (see Supplementary Note 7 for message classification) over the course of an interaction when paired with people across all games. Statistical tests confirm that S# sent significantly more Hate and Threat messages, while people sent more praise messages. b Results of three post-experiment questions for subjects that experienced the condition in which cheap talk was permitted. Participants rated the intelligence of their partner (Intelligence), the clarity of their partner’s intentions (Clear intent?), and the usefulness of the communication between them and their partner (Useful signals?). Answers were given on a 5-point Likert scale. Statistical analysis did not detect any significant differences in the way users rated S# and human partners. c The percentage of time that human participants and S# were thought to be human by their partner. Statistical analysis did not detect any difference in the rates at which S# and humans were judged to be human. Further details along with the statistical analysis for each of these results are provided in Supplementary Note 7