# Probing Association Biases in LLM Moderation Over-Sensitivity

arXiv: https://arxiv.org/abs/2505.23914 · HTML: https://mlsanity.com/paper/2505.23914/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Builds on 19 of its 36 references that are in the map.
- First posted on arXiv 29 May 2025.

## Builds on (most cited first, 19 of 19)

- [SimCSE: Simple Contrastive Learning of Sentence Embeddings](https://mlsanity.com/paper/2104.08821/) (2021)
- [Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations](https://mlsanity.com/paper/2312.06674/) (2023)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [Beyond Accuracy: Behavioral Testing of NLP models with CheckList](https://mlsanity.com/paper/2005.04118/) (2020)
- [Nuanced Metrics for Measuring Unintended Bias with Real Data for Text Classification](https://mlsanity.com/paper/1903.04561/) (2019)
- [Social Bias Frames: Reasoning about Social and Power Implications of Language](https://mlsanity.com/paper/1911.03891/) (2019)
- [Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection](https://mlsanity.com/paper/2111.07997/) (2021)
- [Large Language Model Alignment: A Survey](https://mlsanity.com/paper/2309.15025/) (2023)
- ["Kelly is a Warm Person, Joseph is a Role Model": Gender Biases in LLM-Generated Reference Letters](https://mlsanity.com/paper/2310.09219/) (2023)
- [FalseReject: A Resource for Improving Contextual Safety and Mitigating Over-Refusals in LLMs via Structured Reasoning](https://mlsanity.com/paper/2505.08054/) (2025)
- [SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering](https://mlsanity.com/paper/2408.11491/) (2024)
- [FASTopic: Pretrained Transformer is a Fast, Adaptive, Stable, and Transferable Topic Model](https://mlsanity.com/paper/2405.17978/) (2024)
- [Fewer Errors, but More Stereotypes? The Effect of Model Size on Gender Bias](https://mlsanity.com/paper/2206.09860/) (2022)
- [Uncovering Implicit Gender Bias in Narratives through Commonsense Inference](https://mlsanity.com/paper/2109.06437/) (2021)
- [Investigating Implicit Bias in Large Language Models: A Large-Scale Study of Over 50 LLMs](https://mlsanity.com/paper/2410.12864/) (2024)
- [MentalManip: A Dataset For Fine-grained Analysis of Mental Manipulation in Conversations](https://mlsanity.com/paper/2405.16584/) (2024)
- [Lost in Moderation: How Commercial Content Moderation APIs Over- and Under-Moderate Group-Targeted Hate Speech and Linguistic Variations](https://mlsanity.com/paper/2503.01623/) (2025)
- [Implicit Bias in LLMs: A Survey](https://mlsanity.com/paper/2503.02776/) (2025)
- [Large Language Models for Automatic Detection of Sensitive Topics](https://mlsanity.com/paper/2409.00940/) (2024)

## Built on by (most cited first, 3 of 3)

- [CoPE: A Small Language Model for Steerable and Scalable Content Labeling](https://mlsanity.com/paper/2512.18027/) (2025)
- [Evaluating Second-Order Bias of LLMs Through Epistemic Entitlement](https://mlsanity.com/paper/2606.17506/) (2026)
- [Whitewashing Hate, Smearing Harmless Content: Annotator-Style Rebuttal Attacks on LLM-Based Moderation](https://mlsanity.com/paper/2608.22230/) (2026)

## Abstract (arXiv)

Large Language Models are widely used for content moderation but often present certain over-sensitivity, leading to misclassification of benign content and rejecting safe user commands. While previous research attributes this issue primarily to the presence of explicit offensive triggers, we statistically reveal a deeper connection beyond token level: When behaving over-sensitively, particularly on decontextualized statements, LLMs exhibit systematic topic-toxicity association patterns that go beyond explicit offensive triggers. To characterize these patterns, we propose Topic Association Analysis, a behavior-based probe that elicits short contextual scenarios for benign inputs and quantifies topic amplification between the scenario and the original comment. Across multiple LLMs and large-scale data, we find that more advanced models (e.g., GPT-4 Turbo) show stronger topic-association skew in false-positive cases despite lower overall false-positive rates. Moreover, via controlled prefix interventions, we show that topic cues can measurably shift false-positive rates, indicating that topic framing is decision-relevant. These results suggest that mitigating over-sensitivity may require addressing learned topic associations in addition to keyword-based filtering.