Homophily in Human-AI Interaction

This repository accompanies the manuscript “Homophily in Human–AI Interaction” (Winter & Gërxhani). It contains de-identified, analysis-ready data and reproducibility materials for two studies examining whether demographic homophily shapes partner selection when potential partners are human versus AI agents.

Study 1 (online experiment; Prolific; February 2023). Participants were randomly assigned to a Human–Human (HH) condition (partners were other Prolific participants represented by avatars) or a Human–AI (HA) condition (partners were bot agents represented by avatars). Participants selected (i) a competitor in an individual competition round and (ii) a teammate in a team competition round (order randomized) from balanced avatar pools (50% male/female; 50% White/Black). Participants then completed an incentivized arithmetic reasoning task in each round, producing measures of partner choice, task performance (number correct), and competition outcomes (win/tie/loss). The Study 1 dataset includes participant demographics, treatment assignment, partner-choice variables and demographic cues, derived indicators of same-gender and same-race selection, and performance/outcome variables.

Study 2 (field onboarding data; U.S. users; July 2025). New users of a mobile mental-health companion application selected between two AI companion voice options (one masculine-presenting, one feminine-presenting) during onboarding. The Study 2 dataset includes de-identified user demographics and the selected voice option, enabling tests of gender-based homophily in AI companion selection.

The repository also includes a master R reproducibility script that reads the deposited CSV files and regenerates the main analyses, tables, and figures, writing outputs to a structured outputs/ folder. For replication, it additionally includes the oTree implementation used to run Study 1.