Model-Dependent Moderation: Inconsistencies in Hate Speech Detection Across LLM-based Systems (2025.findings-acl)
Copied to clipboard
| Challenge: | Content moderation systems powered by large language models are increasingly deployed to detect hate speech . if two systems produce different outcomes for the same content, it undermines consistency and predictability . |
| Approach: | They analyze 1.3+ million sentences from a factorial design to determine hate speech classification . they find identical content receives markedly different classification values across systems . |
| Outcome: | The proposed model finds that identical content receives markedly different classification values across systems. |
Similar Papers
Hate Personified: Investigating the role of LLMs in content moderation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Our work provides preliminary guidelines and highlights the nuances of applying Large Language models in culturally sensitive cases. |
| Approach: | They propose to use large language models to help with content moderation to assess how well the needs of diverse groups are reflected in annotated posts. |
| Outcome: | The proposed model is able to leverage community-based flagging efforts and exposure to adversaries. |
Recent Advances in Online Hate Speech Moderation: Multimodality and the Role of Large Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | HS is any communication demeaning a person or a group based on social or ethnic characteristics that undermines social harmony and individual safety . the recent Israel-Hamas conflict has escalated both anti-Muslim and anti-Semitic sentiments worldwide . |
| Approach: | They examine the role of large language models and large multimodal models in HS moderation . they examine how text, images, and audio interact to spread hate speech . |
| Outcome: | The findings highlight the need for solutions in low-resource settings and highlight the gaps in existing methods. |
Revealing Hidden Mechanisms of Cross-Country Content Moderation with Natural Language Processing (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing knowledge on how and why NLP methods make content moderation decisions is limited . authors examine how and when to use LLMs in content modeation . |
| Approach: | They use Shapley values and LLM-guided explanations to reverse-engineer content moderation decisions across countries. |
| Outcome: | The proposed methods show that they reverse-engineer content moderation decisions across countries and over time. |
LLM generated responses to mitigate the impact of hate speech (2024.findings-emnlp)
Copied to clipboard
Jakub Podolak, Szymon Łukasik, Paweł Balawender, Jan Ossowski, Jan Piotrowski, Katarzyna Bakowicz, Piotr Sankowski
| Challenge: | a study aims to determine the effectiveness of large language models to counteract hate speech . it is the first real-life A/B test evaluating the effectiveness . |
| Approach: | They conduct the first real-life A/B test assessing the effectiveness of LLM-generated counter-speech. |
| Outcome: | The proposed system reduces user engagement by over 20%, the study shows . the proposed metric is based on a simple metric and is scalable to other platforms . |
Probing LLMs for hate speech detection: strengths and vulnerabilities (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Recent efforts to detect hateful or toxic language using large language models have not used explanation, additional context and victim community information in the detection process. |
| Approach: | They use different prompt variations, input information and victim community information to evaluate large language models in zero shot setting without adding any in-context examples. |
| Outcome: | The proposed models perform significantly better when included in the pipeline than baseline models. |
Don’t Go To Extremes: Revealing the Excessive Sensitivity and Calibration Limitations of LLMs in Implicit Hate Speech Detection (2024.acl-long)
Copied to clipboard
| Challenge: | Several studies have examined whether large language models exhibit bias or discrimination against individuals or groups in terms of protected attributes like race, gender, or religion. |
| Approach: | They evaluate LLMs' ability to detect implicit hate speech and express confidence in their responses by considering prompt patterns and mainstream uncertainty estimation methods. |
| Outcome: | The proposed models exhibit two extremes: (1) excessive sensitivity towards groups or topics that may cause fairness issues, resulting in misclassifying benign statements as hate speech; (2) confidence scores for each method excessively concentrate on a fixed range, remaining unchanged regardless of the dataset’s complexity. |
SLM-Mod: Small Language Models Surpass LLMs at Content Moderation (2025.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) are expensive to query in real-time and do not allow for a community-specific approach to content moderation. |
| Approach: | They propose to use small language models for community-specific content moderation tasks by fine-tuning and evaluating their performance against larger open- and closed-sourced models. |
| Outcome: | The proposed models outperform zero-shot LLMs in content moderation tasks with 11.5% higher accuracy and 25.7% higher recall across all communities. |
Decoding Hate: Exploring Language Models’ Reactions to Hate Speech (2025.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are trained on vast amounts of unmoderated internet data, enabling them to generate text autonomously. |
| Approach: | They investigate the responses of seven state-of-the-art Large Language Models (LLMs) to hate speech by qualitative analysis. |
| Outcome: | The proposed models can handle hate speech inputs and mitigate it through fine-tuning and guideline guardrailing. |
Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection (2026.acl-long)
Copied to clipboard
| Challenge: | censored models outperform uncensoreed counterparts in accuracy and robustness, achieving 69.0% accuracy versus 64.1% strict accuracy. |
| Approach: | They examine how large language models with minimal safety alignment compare with more heavily aligned counterparts when deployed using political personas. |
| Outcome: | The proposed model outperforms uncensored models in accuracy and robustness, while uncensors are more malleable to ideological framing. |
Is LLM an Overconfident Judge? Unveiling the Capabilities of LLMs in Detecting Offensive Language with Annotation Disagreement (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing studies focus on evaluating large language models' ability to handle disagreement cases. |
| Approach: | They evaluate the performance of large language models in detecting offensive language at varying levels of agreement. |
| Outcome: | The proposed model improves detection accuracy and model alignment with human judgment by using disagreement samples in training. |