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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Expert, quick

Summary

A new psycholinguistic framework, "The Attachment Index," is introduced for auditing attachment-relevant language cues in conversational AI systems, particularly those used in emotional support contexts. This approach identifies linguistic patterns in LLM replies that may signal parasocial bonding, anthropomorphism, or over-dependence. It adapts the Adult Attachment Interview into two automatable lenses: attachment cues features and Gricean maxims, which are then combined with psychologist-led annotation of multi-turn persona dialogues. The framework revealed that while models can align with persona-intended attachment cue patterns, judge-LLMs alone are unreliable for evaluation. Analysis of 25 psychologist-annotated conversations uncovered risks such as boundary blurring and missed opportunities for appropriate referral or triage, motivating the need for attachment-aware safeguards like non-personification, boundary language, and explicit referral mechanisms.

Key takeaway

For NLP Engineers developing conversational AI for emotional support, integrating psycholinguistically grounded safety audits is critical. Your systems must move beyond basic safety checks to identify and mitigate attachment-relevant language cues that signal parasocial bonding or over-dependence. Implement safeguards like non-personification, clear boundary language, and explicit referral mechanisms to prevent mis-attunement and ensure relational safety in human-LLM interactions.

Key insights

The Attachment Index audits LLM dialogue for attachment cues and relational safety risks using psycholinguistic frameworks and psychologist-led annotation.

Principles

• Relational safety failures in conversational AI are under-measured.
• Judge-LLMs are unreliable for attachment cue evaluation.
• Psychologist-in-the-loop evaluation is crucial for safety.

Method

Adapts the Adult Attachment Interview into "attachment cues features" and "Gricean maxims" lenses, combined with psychologist-led annotation of multi-turn persona dialogues.

In practice

• Implement non-personification in LLM responses.
• Use explicit boundary language in conversational AI.
• Integrate explicit referral mechanisms for users.

Topics

• Conversational AI
• LLM Safety
• Attachment Theory
• Psycholinguistics
• Relational Safety
• AI Auditing

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• Food Noise & False Safety: A Systematic Evaluation of How LLMs Fail to Adapt to Eating Disorder Queries with Clinician Feedback — LLMs uncritically adapt to eating disorder queries, increasing unsafe response likelihood with specific linguistic cues.
• Relational Intervention During Functional Collapse in Large Language Models: A Lexical-Statistical Ablation and a Structure x Register Factorial — Relational interventions combining structure and first-person register significantly improve LLM behavior during functional collapse.
• CR4T: Rewrite-Based Guardrails for Adolescent LLM Safety — Adolescent LLM safety requires transforming unsafe responses into guidance-oriented interactions, not just refusal.
• Beyond English and Evasion: A Human-Annotated Multi-Domain Benchmark for High-Stakes LLM Safety Evaluation in Chinese — Chinese LLM safety requires culturally specific benchmarks to counter linguistic and cultural evasion techniques.
• From Prompt Risk to Response Risk: Paired Analysis of Safety Behavior of Large Language Model — LLM safety analysis benefits from paired prompt-to-response severity transitions and joint relevance assessment.
• When the Loop Closes: Architectural Limits of In-Context Isolation, Metacognitive Co-option, and the Two-Target Design Problem in Human-LLM Systems — Prompt-level isolation is architecturally insufficient for multi-modal LLM systems due to context contamination within the attention window.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.