AI Moderation: Inconsistencies in Hate Speech Detection Across LLM-Based Systems (aclanthology.org)

A recent study reveals significant inconsistencies in hate speech detection across seven leading AI moderation systems, including OpenAI’s Moderation Endpoints, Claude 3.5 Sonnet, GPT-4o, and Google’s Perspective API. By analyzing over 1.3 million synthetic sentences designed to cover a wide range of expressions and demographics, researchers found that identical content is often classified differently depending on the system used. These discrepancies are especially pronounced for specific demographic groups, raising concerns about fairness and reliability in automated hate speech moderation. This work is crucial for the AI/ML community as it highlights the lack of a standardized approach to defining harmful content, which undermines the predictability and perceived fairness of moderation outcomes. Technically, the study demonstrates that differences in model architectures and training data lead to varied decision boundaries when flagging hate speech. This variation can cause moderation policies to appear arbitrary, potentially impacting user trust and the effectiveness of automated content controls. The findings call for more transparency and alignment in content moderation models to ensure consistent and equitable AI-driven content governance.