When Different Relational Experiences with AI Become Visible in Conversation: Linguistic and Dyadic Evidence from Naturalistic ChatGPT Interactions

Key Findings

• — 94 German-speaking adult ChatGPT users (mean age 24.87; 77.7% female; mostly students) contributed self-reports and a donated conversation; 664 eligible three-turn sequences from 85 conversations, 500 triple-annotated
• — Digital therapeutic alliance and AI attachment were related but had different conversational correlates
• — Validation (cross-validated R2 = .423) and user uptake (R2 = .260) were predictable out of sample, but neither formed a generalizable alliance signature
• — A post-registration multivariate analysis found a distributed social-substitution fingerprint beyond conversation length and topic (delta R2 = .070), improving prediction in 98% of repetitions
• — Ruptures were rare (12 rupture-positive sequences among 500 coded) and consistently repaired; their association with alliance was fragile

Methodology Notes

Recruitment in Germany via university advertising and social media; eligibility required current ChatGPT use, native-level German and at least one prior personally meaningful conversation (emotional, interpersonal, loneliness, stress, decisions). Preregistered analysis plan on OSF; the social-substitution finding is post-registration. Small convenience sample skewed young, female and student. Not peer reviewed. Verification: OSF API record hr7gc_v1 200 (title, contributors, abstract, published 2026-10-12T16:52Z) and the full PDF via osf.io/download/hr7gc_v1 (200, 3.8 MB; participants section and results read).

Sources