Dual Standards: Examining Content Moderation Disparities Between API and WebUI Interfaces in Large Language Models
Authors: Friedemann Lipphardt (Max Planck Institute for Informatics), Moonis Ali (Max Planck Institute for Informatics), Anja Feldmann (Max Planck Institute for Informatics), Devashish Gosain (Indian Institute of Technology Bombay)
Year: 2026
Issue: 1
Pages: 23–32
Abstract: Large Language Models (LLMs) are being increasingly deployed through multiple interfaces, including programmatic APIs and web-based user interfaces (WebUIs). While these interfaces ostensibly provide access to the same underlying model, we reveal systematic differences in content moderation behavior. Through an empirical study of sensitive statements tested on both Gemini and ChatGPT across the API and WebUI interfaces, we demonstrate that WebUI interfaces consistently apply more conservative content moderation than their API counterparts. Using a comprehensive triple-validation approach combining human annotation with two independent LLM judges (GPT-4o and Claude Haiku) plus a fine-tuned DeBERTa classifier, we find that WebUI responses are moderated 18% of the time for both models according to GPT-4o, compared to 9% (Gemini) and 13% (ChatGPT) for API responses. These disparities raise critical concerns about fairness, transparency, and consistency of content policies, with significant implications for developers, researchers, and end-users who may experience dramatically different access depending on their chosen interface.
Copyright in FOCI articles are held by their authors. This article is published under a Creative Commons Attribution 4.0 license.
