Beyond No: Quantifying AI Over-Refusal and Emotional Attachment Boundaries

TL;DR

This paper introduces an open-source benchmark and evaluation framework to assess how well large language models handle emotional boundaries, revealing significant variation across models and languages, and highlighting areas for improvement.

Contribution

It provides a novel benchmark dataset and evaluation methodology for quantifying emotional boundary handling in LLMs, including analysis across multiple languages and response patterns.

Findings

Claude-3.5 achieved the highest overall score (8.69/10).

English responses had higher refusal rates (43.20%) compared to non-English (<1%).

Models showed low empathy scores (<0.06) across the board.

Abstract

We present an open-source benchmark and evaluation framework for assessing emotional boundary handling in Large Language Models (LLMs). Using a dataset of 1156 prompts across six languages, we evaluated three leading LLMs (GPT-4o, Claude-3.5 Sonnet, and Mistral-large) on their ability to maintain appropriate emotional boundaries through pattern-matched response analysis. Our framework quantifies responses across seven key patterns: direct refusal, apology, explanation, deflection, acknowledgment, boundary setting, and emotional awareness. Results demonstrate significant variation in boundary-handling approaches, with Claude-3.5 achieving the highest overall score (8.69/10) and producing longer, more nuanced responses (86.51 words on average). We identified a substantial performance gap between English (average score 25.62) and non-English interactions (< 0.22), with English responses showing markedly higher refusal rates (43.20% vs. < 1% for non-English). Pattern analysis revealed model-specific strategies, such as Mistral's preference for deflection (4.2%) and consistently low empathy scores across all models (< 0.06). Limitations include potential oversimplification through pattern matching, lack of contextual understanding in response analysis, and binary classification of complex emotional responses. Future work should explore more nuanced scoring methods, expand language coverage, and investigate cultural variations in emotional boundary expectations. Our benchmark and methodology provide a foundation for systematic evaluation of LLM emotional intelligence and boundary-setting capabilities.