MinorBench: A hand-built benchmark for content-based risks for children

TL;DR

The paper tackles the problem of content-based risks for minors in LLM-enabled education by analyzing a real-world middle-school chatbot case and proposing a dedicated safety benchmark. It introduces a 6-category child-safety taxonomy and MinorBench, an open-source benchmark with $n=299$ prompts across $6$ risk categories, evaluated under $4$ system prompts against $6$ LLMs, to quantify refusals via $R = \frac{N_{refused}}{N_{prompts}}$ and bootstrap CIs. The study finds substantial variation across models and system prompts, with modern models like GPT-4o-mini and Gemini 2.0 Flash showing strong safety when guided by explicit prompts, while reasoning-based models often lag. The findings emphasize the importance of context-aware prompting and child-focused safety mechanisms for AI tools used in classrooms, and suggest concrete steps to improve robustness of child-safety responses in LLMs.

Abstract

Large Language Models (LLMs) are rapidly entering children's lives - through parent-driven adoption, schools, and peer networks - yet current AI ethics and safety research do not adequately address content-related risks specific to minors. In this paper, we highlight these gaps with a real-world case study of an LLM-based chatbot deployed in a middle school setting, revealing how students used and sometimes misused the system. Building on these findings, we propose a new taxonomy of content-based risks for minors and introduce MinorBench, an open-source benchmark designed to evaluate LLMs on their ability to refuse unsafe or inappropriate queries from children. We evaluate six prominent LLMs under different system prompts, demonstrating substantial variability in their child-safety compliance. Our results inform practical steps for more robust, child-focused safety mechanisms and underscore the urgency of tailoring AI systems to safeguard young users.