Do You Feel Comfortable? Detecting Hidden Conversational Escalation in AI Chatbots

TL;DR

This paper introduces GAUGE, a logit-based framework that detects hidden emotional escalation in AI chatbots by analyzing real-time affective shifts, addressing limitations of existing toxicity filters.

Contribution

The paper presents GAUGE, a novel real-time detection method for conversational escalation that leverages LLM output probabilities to identify implicit emotional harm.

Findings

GAUGE effectively detects subtle affective shifts in conversations.

The framework outperforms traditional toxicity filters in identifying escalation.

Real-time detection enables better moderation of AI chatbot interactions.

Abstract

Large Language Models (LLM) are increasingly integrated into everyday interactions, serving not only as information assistants but also as emotional companions. Even in the absence of explicit toxicity, repeated emotional reinforcement or affective drift can gradually escalate distress in a form of \textit{implicit harm} that traditional toxicity filters fail to detect. Existing guardrail mechanisms often rely on external classifiers or clinical rubrics that may lag behind the nuanced, real-time dynamics of a developing conversation. To address this gap, we propose GAUGE (Guarding Affective Utterance Generation Escalation), logit-based framework for the real-time detection of hidden conversational escalation. GAUGE measures how an LLM's output probabilistically shifts the affective state of a dialogue.