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

Current AI safety tools don't adequately protect children: The paper reveals that existing AI ethics research focuses on general audiences, leaving significant gaps in how LLMs handle content-specific risks unique to minors—a critical oversight as these systems increasingly enter schools and children's daily lives.

Real-world deployment exposed serious safety gaps: A case study of an LLM chatbot in a middle school showed students could misuse the system in ways current safety measures didn't anticipate, demonstrating that theoretical safety testing doesn't capture how children actually interact with AI.

Major LLMs show inconsistent child-safety performance: Testing six prominent LLMs revealed "substantial variability" in their ability to refuse unsafe or inappropriate requests from children, meaning some models are significantly less protective than others depending on how they're configured.

A new framework is needed to evaluate child safety: The paper introduces MinorBench, an open-source benchmark with a taxonomy of content-based risks specific to minors, providing the first standardized tool to measure and compare how well LLMs protect young users—addressing a previously unmeasured gap.

System prompts significantly affect child protection: The research shows that how LLMs are instructed (their "system prompts") substantially impacts their willingness to refuse harmful requests, suggesting that better configuration and guardrails could meaningfully improve child safety without requiring complete model redesigns.