Abstract
People may experience large language models (LLMs) as therapeutically allied, emotionally close, or socially substitutive, but whether these relationships become visible in conversations is unknown. We distinguished what users discussed, what unfolded relationally, and how they experienced the AI relationship. We linked self-reports from 94 adults to donated, personally meaningful ChatGPT conversations using 21 outcome-blind frozen topics, blinded triple annotation of 500 sequences, and participant-grouped machine learning frozen before outcome access. Digital Therapeutic Alliance (DTA) and AI Attachment were related but showed different conversational correlates. Validation (R²_CV = .423) and uptake (R²_CV = .260) were recoverable out of sample, yet, neither formed a generalizable DTA signature. Conversely, a post-registration multivariate analysis identified a distributed Social Substitution fingerprint beyond conversation length and 21 topics (ΔR²_CV = .070) and improved prediction in 98% of repetitions. Ruptures were rare and consistently repaired; their association with DTA was fragile. Thus, different human–LLM relationships leave different—and sometimes no readily observable—conversational traces.