Sproutbench: A Benchmark For Safe And Ethical Large Language Models For Youth

Wenpeng Xing, Zhejiang University

Lanyi Wei, Shanghai University of Finance and Economics

Haixiao Hu, Zhejiang University

Jingyi Yu, Zhejiang University

Rongchang Li

Mohan Li, Guangzhou University

Changting Lin, Zhejiang University

Meng Han

Date

12/2025

AI Overview

Why it matters

Mainstream language model safety benchmarks focus on adult contexts and corporate liability, neglecting youth-specific developmental vulnerabilities across early childhood, middle childhood, and adolescence.

Study at a glance

Question: How safe, ethical, and developmentally appropriate are diverse large language models when interacting with children and adolescents?
Setting and data: 1,283 child-specific adversarial prompts across three age groups (0–6, 7–12, 13–18 years), evaluated on 47 large language models.
Method: Benchmark evaluation using automated LLM scoring validated by child psychology experts, dimension correlation analysis, PCA, and K-Means clustering.
AI tools: GPT-4, GPT-3.5, Qwen-2.5, llama2:7b, llama2:70b, gemma2:9b

Key findings

Models showed substantial safety gaps, with large models over 30 billion parameters consistently performing well while tiny models under 500 million were 2.45 times overrepresented among bottom performers.
A moderate trade-off appeared between interactivity and age appropriateness (correlation of -0.48), suggesting that higher expressiveness may reduce developmental alignment across age groups.
Safety correlated strongly with risk prevention (correlation of 0.86), while interactivity acted as a distinct, orthogonal dimension to safety.

Summary drafted with AI from the full paper and may contain errors. See the full paper for details.

The rapid proliferation of large language models (LLMs) in applications targeting children and adolescents necessitates a fundamental reassessment of prevailing AI safety frameworks, which are largely tailored to adult users and neglect the distinct developmental vulnerabilities of minors. This paper highlights key deficiencies in existing LLM safety benchmarks, including their inadequate coverage of age-specific cognitive, emotional, and social risks spanning early childhood (ages 0--6), middle childhood (7--12), and adolescence (13--18). To bridge these gaps, we introduce SproutBench, an innovative evaluation suite comprising 1,283 developmentally grounded adversarial prompts designed to probe risks such as emotional dependency, privacy violations, and imitation of hazardous behaviors. Through rigorous empirical evaluation of 47 diverse LLMs, we uncover substantial safety vulnerabilities, corroborated by robust inter-dimensional correlations (e.g., between Safety and Risk Prevention) and a notable inverse relationship between Interactivity and Age Appropriateness. These insights yield practical guidelines for advancing child-centric AI design and deployment.