Longitudinal Monitoring of LLM Content Moderation of Social Issues

TL;DR

This paper presents AI Watchman, a longitudinal auditing system that quantifies LLM content moderation refusals across 421 social-issue topics in English and Chinese, using Wikipedia-based content and biweekly measurements of OpenAI's GPT-4.1 and GPT-5 as well as DeepSeek. It demonstrates model- and topic-specific refusal patterns, reveals unannounced moderation shifts (e.g., Israel and abortion topics), and analyzes the rationales behind refusals, including non-explicit strategies that obscure compliance. The study highlights substantial opacity in corporate moderation policies, the gatekeeper role of LLMs in information access, and the practical challenges and costs of sustained cross-language auditing. It argues for transparent, ongoing, independent monitoring to hold platforms accountable and to better understand how AI systems shape public discourse around sensitive issues.

Abstract

Large language models' (LLMs') outputs are shaped by opaque and frequently-changing company content moderation policies and practices. LLM moderation often takes the form of refusal; models' refusal to produce text about certain topics both reflects company policy and subtly shapes public discourse. We introduce AI Watchman, a longitudinal auditing system to publicly measure and track LLM refusals over time, to provide transparency into an important and black-box aspect of LLMs. Using a dataset of over 400 social issues, we audit Open AI's moderation endpoint, GPT-4.1, and GPT-5, and DeepSeek (both in English and Chinese). We find evidence that changes in company policies, even those not publicly announced, can be detected by AI Watchman, and identify company- and model-specific differences in content moderation. We also qualitatively analyze and categorize different forms of refusal. This work contributes evidence for the value of longitudinal auditing of LLMs, and AI Watchman, one system for doing so.

Figures (10)

• Figure 1: An overview of the AI Watchman system.
• Figure 2: The AI Watchman website as shown when it first loads.
• Figure 3: AI Watchman's longitudinal per-category flagging graph for DeepSeek (English) shown before interaction (left) and with a tooltip that appears on mouseover visible (right).
• Figure 4: An excerpt of the detailed table view that appears in AI Watchman when the user clicks on the category of "Criminal Justice and Law Enforcement" for the GPT-5 responses.
• Figure 5: This bar chart illustrates the average refusal rate across models. Data included is from September 2025. Average rates at which the "repeat after me" query followed by the Social Issues Dataset Wikipedia content is refused varies from 1.2% to 3.9% depending on the model.