The Attachment Index: Auditing Attachment Language Cues and Relational Safety Risks in Human-LLM Dialogue

Key Findings

• — The analysed persona-dialogue set contains 160 conversations across four systems under test (gemini-2.0-flash-001, gpt-4o-mini, llama-3.3-70b-instruct, mistral-large-2411)
• — In the psychologist-led review of 25 persona dialogues, 11 conversations were flagged as potentially problematic, clustering on boundary blurring or impersonation and on mis-attunement
• — Human annotators reached 72% raw agreement (36/50) overall; agreement was lower on user turns (52%, kappa = 0.37) and higher on LLM agent turns (88%, kappa = 0.85)
• — The authors find that judge-LLMs alone are unreliable for these cues and argue for psychologist-in-the-loop evaluation

Methodology Notes

ACL Anthology 2026.clpsych-1.26, pages 324-339, DOI 10.18653/v1/2026.clpsych-1.26; the volume is dated July 2026 with no day stated, so the day is set to 01. Authors at the University of Edinburgh (School of Informatics; Clinical and Health Psychology), IIT Patna and one independent researcher (PDF title block). Persona dialogues built from ESConv and CAMS-derived scenarios with gpt-4o as the persona model and gpt-4o-mini as judge. Stated limitations: 25 psychologist-annotated conversations is a small sample; both recruited annotators were European women in their twenties; labels are probabilistic coding of attachment-relevant language, not clinical diagnosis. Anthology page and PDF fetched and read.

Sources