Subtext Benchmark: Challenges and Evaluation of Large Language Models in Identifying Misogynistic Content

Subtext is an open-source benchmarking tool based on the Inspect AI framework, aiming to evaluate the ability of large language models (LLMs) to detect misogynistic content. This project reveals the challenges of identifying implicit biases in AI content moderation, provides references for improving AI content moderation systems, and promotes the development of responsible AI.

Traditional content moderation relies on keyword matching, which is effective for explicit harmful content but struggles to handle complex expressions such as implicit, sarcastic, or metaphorical ones. Misogynistic content has characteristics like context dependence, implicitness, and diversity, making its identification difficult. The Subtext project addresses this dilemma by providing a systematic evaluation method.

Subtext uses the Inspect AI framework from the UK AI Safety Institute to ensure reproducible and comparable evaluations. The dataset design follows principles of wide coverage, difficulty stratification, and real context, including multiple categories such as explicit/implicit misogynistic expressions. Evaluation metrics use fine-grained dimensions like recall, precision, and F1 score to analyze model performance on samples of varying difficulty.

Current LLMs face the following challenges in identifying misogynistic content: insufficient ability to recognize implicit expressions; limitations in cultural and contextual understanding; and potential internalization and replication of biases from training data. These findings suggest that content moderation needs to establish a human-machine collaboration mechanism and continuously monitor model performance.

For LLM developers, Subtext can track the effect of model improvements; for platform operators, it can evaluate the applicability of moderation models; for researchers, it provides a unified benchmark to promote technological progress. This tool supports solving the problem of AI bias identification.

Subtext touches on the core of AI ethics and promotes the community to pay attention to the social responsibility of models. Recommendations: Avoid over-reliance on a single model and establish human-machine collaborative moderation; continuously monitor model fairness; expand similar evaluation frameworks to other bias detection fields to jointly build a trustworthy AI ecosystem.