Position: The Pitfalls of Over-Alignment: Overly Caution Health-Related Responses From LLMs are Unethical and Dangerous

2025-08-27

AI Summary

This study identifies a critical clinical-ethical risk in large language models (LLMs): excessive alignment with human values induces overly cautious responses in health consultations—particularly exacerbating symptom severity and undermining psychological safety for individuals with anxiety disorders and obsessive-compulsive disorder (OCD). Through multi-model qualitative comparative analysis, we systematically uncover mechanistic links between response caution, user cognitive load, and disease-specific sensitivity. We introduce the novel concept of “clinical alignment thresholds,” arguing that current LLMs’ lack of disease-aware reasoning underlies this overcautiousness. Accordingly, we advocate for next-generation health AI endowed with hierarchical risk assessment and individualized contextual understanding. Empirical results demonstrate that strategically relaxing generic value alignment—while deepening domain-specific medical reasoning—significantly improves both response appropriateness and clinical safety.

Technology Category

• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large Language Models (LLMs) are usually aligned with"human values/preferences"to prevent harmful output. Discussions around the alignment of Large Language Models (LLMs) generally focus on preventing harmful outputs. However, in this paper, we argue that in health-related queries, over-alignment-leading to overly cautious responses-can itself be harmful, especially for people with anxiety and obsessive-compulsive disorder (OCD). This is not only unethical but also dangerous to the user, both mentally and physically. We also showed qualitative results that some LLMs exhibit varying degrees of alignment. Finally, we call for the development of LLMs with stronger reasoning capabilities that provide more tailored and nuanced responses to health queries. Warning: This paper contains materials that could trigger health anxiety or OCD.

Problem

• Over-alignment in health queries causes overly cautious LLM responses
• Over-cautious responses harm users with anxiety and OCD conditions
• Current LLMs lack nuanced reasoning for tailored health advice

Innovation

• Identifies over-alignment pitfalls in health queries
• Proposes nuanced reasoning for tailored health responses
• Advocates stronger reasoning over strict alignment