SaTC 2.0: RES: Collaborative Research: Protecting Young Adults from Unsafe AI-Generated Content

Generative artificial intelligence platforms increasingly mediate how young adults access information, express creativity, and interact online. However, their open-ended, prompt-driven nature may make them susceptible to generating content that could negatively affect youth mental health. To address this challenge, this project develops a safeguard framework for protecting young adults from unsafe artificial intelligence-generated content. The project aims to support safer online interactions and to improve the well-being and mental health of adolescents and young adults. The project has four interconnected research thrusts. Thrust 1 develops a cognitively grounded framework for large language models that detects inconsistencies between the beliefs and behaviors of users. Thrust 2 investigates semantic-preserving mitigation of images in diffusion models by removing unsafe visual elements while preserving prompt meaning, image structure, and visual coherence. Thrust 3 develops a normative knowledge-guided safeguard layer based on retrieval-augmented generation and vision-language models to inspect generated images and videos using relevant policies, age-appropriate guidance, contextual norms, and curated precedents. Thrust 4 evaluates behavioral harm reduction through surveys and in-depth interviews with young adults and parents, examining whether reduced exposure to unsafe generated content leads to measurable changes in engagement patterns, perceived acceptability, and beliefs. The project outcomes include more robust protection against adversarial and ambiguous prompts, safer multimodal content generation, transparent and auditable safety decisions, and behavior-centered evaluation methods for trustworthy generative artificial intelligence. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.