LLM Content Moderation and User Satisfaction: Evidence from Response Refusals in Chatbot Arena

Jun 9, 2026

Upload by Anda

LLM safety and ethical alignment remain widely discussed, while the effect of content moderation on user satisfaction is still underexplored. Analyzing nearly 50,000 model comparisons from Chatbot Arena, this work studies how users react when models refuse prompts. A RoBERTa-based refusal classifier separates ethical refusals from technical limitations. Results show a substantial refusal penalty: ethical refusals receive lower win rates than technical refusals and standard responses. The penalty varies with prompt sensitivity and refusal phrasing, highlighting a safety-versus-expectations tension.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages29

Reading Time1 h 13 min

Introduction

Background and motivation

Research questions

Approach and dataset

FAQs

What gap does the paper address about LLM safety and moderation?

The paper targets how content moderation via refusal behavior affects user satisfaction, which has received less direct study than system-level refusal optimization.

How does the study distinguish ethical refusals from technical limitations?

It fine-tunes a RoBERTa-based refusal classifier on a hand-labeled dataset to classify different refusal causes.

What overall effect do ethical refusals have on user preferences?

Ethical refusals incur a substantial penalty, producing significantly lower win rates than technical refusals and standard responses.

Jun 9, 2026

Upload by Anda