Homophily in Human-AI Interaction

Description

This repository accompanies the manuscript "Homophily in Human–AI Interaction" (Winter & Gërxhani). It contains de-identified 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 is the raw experiment export and includes participant demographics, treatment assignment, partner-choice variables and demographic cues, and performance/outcome variables; indicators of same-gender and same-race selection are derived from it by the analysis script. 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 includes a master R script (reproduce_all.R) that reads the deposited CSV files and regenerates every statistic reported in the article and its online appendix, writing a labelled reproduction log, tables and figures to a structured outputs/ folder. See README.md for details of the data and the analysis. For replication, the repository additionally includes the oTree implementation used to run Study 1.

Institutions

• European University Institute

  Tuscany, Florence
• Vrije Universiteit Amsterdam

  North Holland, Amsterdam

Categories

Licence