Safe-Child-LLM: A Developmental Benchmark for Evaluating LLM Safety in Child-LLM Interactions

Current AI safety evaluations ignore children's unique vulnerabilities: The paper reveals that existing LLM safety assessments focus almost exclusively on adult users, leaving a critical gap in protecting minors who increasingly interact with these systems through educational apps, homework helpers, and other child-facing applications.

Major AI systems show significant safety failures with children: Testing leading models including ChatGPT, Claude, Gemini, and others against 200 adversarial prompts designed for children (ages 7-12) and adolescents (13-17) uncovered "critical safety deficiencies in child-facing scenarios," indicating that popular AI systems are not adequately safeguarded for younger users.

Developmental stage matters for AI safety: The benchmark distinguishes between two age groups (7-12 and 13-17) because children and adolescents have different cognitive abilities and vulnerabilities—a nuance that current one-size-fits-all safety approaches fail to address.

Transparent benchmarking can drive industry accountability: By publicly releasing their benchmark dataset and evaluation code, the researchers enable independent testing and comparison of how different AI systems perform with child users, creating pressure for companies to improve safety rather than relying on self-reported claims.

Jailbreak attacks pose particular risks to minors: The paper uses adversarial prompts from red-teaming corpora to test whether AI systems can be manipulated into unsafe responses, revealing that children may be especially susceptible to these attacks due to their developmental stage and trust in technology.