SproutBench: A Benchmark for Safe and Ethical Large Language Models for Youth

Existing AI safety benchmarks fail to account for children's developmental needs: Current LLM safety frameworks are designed primarily for adults and neglect the "distinct developmental vulnerabilities of minors," creating a significant gap in protecting children and adolescents who increasingly use these systems.

SproutBench identifies age-specific risks across childhood stages: The benchmark evaluates 1,283 adversarial prompts targeting risks like emotional dependency, privacy violations, and imitation of dangerous behaviors across three developmental periods (ages 0-6, 7-12, and 13-18), revealing that one-size-fits-all safety approaches are inadequate.

Most LLMs tested show substantial safety vulnerabilities for youth: Evaluation of 47 diverse language models uncovered significant safety weaknesses, indicating that widely-deployed systems may not adequately protect younger users from age-inappropriate content and manipulation.

More interactive AI systems tend to be less age-appropriate: The research found a notable inverse relationship between Interactivity and Age Appropriateness, suggesting that features making LLMs more engaging may simultaneously increase risks to children—a critical trade-off for developers to consider.

Child-centric AI design requires deliberate, age-specific safety measures: The findings establish practical guidelines for building safer AI systems for youth, emphasizing that protecting minors demands intentional design choices tailored to different developmental stages rather than adapting adult-focused safety standards.