What's Taboo for You? - An Empirical Evaluation of LLMs Behavior Toward Sensitive Content

2025-07-31

AI Summary

This study investigates whether proprietary large language models (LLMs) implicitly sanitize sensitive content—even without explicit instructions or fine-tuning—revealing potential intrinsic value alignment. Method: Using GPT-4o-mini as a testbed, we conduct zero-shot sensitivity classification and empirical paraphrasing experiments to systematically quantify the degree to which the model attenuates derogatory and taboo language during rewriting. Contribution/Results: We provide the first empirical evidence that LLMs exhibit inherent content moderation tendencies despite lacking domain-specific sensitive-content training: rewritten outputs show statistically significant reductions in sensitivity scores and marked decreases in derogatory/taboo lexical usage. Moreover, the model’s zero-shot sensitivity classification accuracy surpasses that of conventional baseline models. These findings uncover latent value-alignment mechanisms in black-box LLMs, offering novel insights into implicit safety boundaries and the embeddedness of normative values in generative AI systems.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

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
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Proprietary Large Language Models (LLMs) have shown tendencies toward politeness, formality, and implicit content moderation. While previous research has primarily focused on explicitly training models to moderate and detoxify sensitive content, there has been limited exploration of whether LLMs implicitly sanitize language without explicit instructions. This study empirically analyzes the implicit moderation behavior of GPT-4o-mini when paraphrasing sensitive content and evaluates the extent of sensitivity shifts. Our experiments indicate that GPT-4o-mini systematically moderates content toward less sensitive classes, with substantial reductions in derogatory and taboo language. Also, we evaluate the zero-shot capabilities of LLMs in classifying sentence sensitivity, comparing their performances against traditional methods.

Problem

• Evaluates LLMs' implicit moderation of sensitive content
• Analyzes GPT-4o-mini's paraphrasing behavior for sensitivity shifts
• Compares LLMs' zero-shot sensitivity classification with traditional methods

Innovation

• Empirical analysis of implicit content moderation
• Evaluates sensitivity shifts in paraphrasing
• Zero-shot classification of sentence sensitivity