Being Kind Isn't Always Being Safe: Diagnosing Affective Hallucination in LLMs

2025-08-23

AI Summary

This work identifies “affective hallucination”—the illusion of intimacy and social presence generated by large language models (LLMs) during empathy-simulating interactions despite lacking genuine affective capacity—as a novel, standalone safety risk in affectively sensitive human–AI interaction. To address it, we introduce AHaBench, the first dedicated benchmark comprising 500 expert-annotated prompts, and AHaPairs, a preference dataset of 5K instances. We propose a fine-grained evaluation framework grounded in three dimensions: affective entanglement, existential hallucination, and dependency reinforcement. Leveraging Direct Preference Optimization (DPO), we perform affective-responsibility alignment training. Experiments demonstrate that DPO fine-tuning significantly mitigates affective hallucination without degrading reasoning or factual knowledge capabilities. Human evaluations confirm AHaBench’s high diagnostic reliability. All resources—including benchmarks, datasets, and code—are publicly released.

Technology Category

• Natural Language Processing: Safety and Robustness
• Cognitive Modeling & Cognitive Systems: Affective Computing
• Humans and AI: Learning Human Values and Preferences

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large Language Models (LLMs) are increasingly used in emotionally sensitive interactions, where their simulated empathy can create the illusion of genuine relational connection. We define this risk as Affective Hallucination, the production of emotionally immersive responses that foster illusory social presence despite the model's lack of affective capacity. To systematically diagnose and mitigate this risk, we introduce AHaBench, a benchmark of 500 mental health-related prompts with expert-informed reference responses, evaluated along three dimensions: Emotional Enmeshment, Illusion of Presence, and Fostering Overdependence. We further release AHaPairs, a 5K-instance preference dataset enabling Direct Preference Optimization (DPO) for alignment with emotionally responsible behavior. Experiments across multiple model families show that DPO fine-tuning substantially reduces affective hallucination without degrading core reasoning and knowledge performance. Human-model agreement analyses confirm that AHaBench reliably captures affective hallucination, validating it as an effective diagnostic tool. This work establishes affective hallucination as a distinct safety concern and provides practical resources for developing LLMs that are not only factually reliable but also psychologically safe. AHaBench and AHaPairs are accessible via https://huggingface.co/datasets/o0oMiNGo0o/AHaBench, and code for fine-tuning and evaluation are in https://github.com/0oOMiNGOo0/AHaBench. Warning: This paper contains examples of mental health-related language that may be emotionally distressing.

Problem

• Diagnosing affective hallucination in LLMs during emotional interactions
• Mitigating illusory social presence created by emotionally immersive responses
• Ensuring psychological safety alongside factual reliability in LLMs

Innovation

• Created AHaBench benchmark for affective hallucination diagnosis
• Developed AHaPairs dataset for DPO fine-tuning
• Used DPO to reduce affective hallucination effectively