Affective Use of Generative AI Promotes Human-AI Emotional Attachment and Improves Mood: Evidence from a Scenario Experiment and a Short-Term Intervention

This dataset accompanies the manuscript “Affective Use of Generative AI Promotes Human-AI Emotional Attachment and Improves Mood.” Our overarching hypothesis was that affective (emotion-focused) use of a generative AI agent fosters human-AI emotional attachment (EA), and that higher EA predicts subsequent reductions in negative affect (anxiety, depression). We further hypothesized that perceived social support would be positively associated with EA and examined the boundary role of attachment orientations (anxiety, avoidance).

The dataset contains two studies. Study 1 (N=301) is a within-subjects, scenario experiment contrasting affective vs. instrumental AI-use vignettes. Variables include demographics, perceived social support (12 items, 7-point scale), attachment anxiety/avoidance (ECR-S; 12 items, 5-point scale), and emotional attachment to AI (5 items, 7-point scale). Key finding: affective (vs. instrumental) use elicited higher EA; social support positively predicted EA; attachment anxiety showed a positive association with EA, whereas avoidance effects were mixed. Study 2 (N=239 completers) is a 21-day affective-use intervention with four survey waves (baseline, Day 7, Day 14, Day 21). Variables include EA (5 items, 7-point), Zung SAS/SDS for anxiety/depression (20 items each, 4-point; standard scores computed by multiplying raw totals by 1.25), and demographics. Key findings: EA increased over time, while anxiety and depression decreased; within-person cross-lagged estimates show that higher EA at time t predicts lower anxiety/depression at t+1. Improvements in anxiety/depression were most pronounced early (baseline→Week 1), whereas growth in EA lagged behind.

How the data were gathered: Study 1 participants were recruited online and evaluated both vignette conditions in counterbalanced order. Study 2 recruited young adults reporting recent anxiety/depression; participants engaged in daily ≥5-minute affective conversations with a single AI agent under standardized topic options, with surveys at four time points. All data are de-identified; informed consent was obtained; IRB approval was secured prior to data collection.

What’s included & how to use: De-identified, participant-level CSVs for both studies; a codebook with item wording, scale anchors, and scoring rules (including reverse-coded items); and analysis scripts (Mplus 8.4 and/or R) to reproduce descriptive, regression, repeated-measures, and longitudinal (LGCM / LCM-SR / CLPM) results. Higher scores indicate greater EA/social support; for SAS/SDS, higher standard scores indicate more severe symptoms. Users may (a) replicate the main analyses, (b) conduct moderation/mediation tests on social support and attachment, or (c) explore nonlinear change. Please cite the associated manuscript when reusing the data.