Treating what generative AI produces as a health risk in itself — a safeguard framework for young adults built on four thrusts

1Not dangerous information but dangerous generation

Debate about online safety has long assumed that harmful material exists somewhere and the risk is encountering it. Generative AI changes that. Content is not sitting there in advance; it is produced on the spot in response to what a user types. What comes out cannot be enumerated beforehand and differs person to person.

This project starts from the premise that the open-ended, prompt-driven nature itself can create risk to young people mental health.

2The subject is youth mental health

1. 1FramingGenerative AI platforms now mediate how young adults access information, express creativity and interact
2. 2Where the risk sitsThe open-ended, prompt-driven nature may generate content that negatively affects mental health
3. 3StructureA safeguard framework developed across four interconnected research thrusts
4. 4The first thrustA cognitively grounded framework for large language models
5. 5Intended outcomeSafer online interaction and improved well-being and mental health for adolescents and young adults

The phrase worth noting is cognitively grounded. Rather than handling the problem with a banned list of words and topics, the framework is to be built around how content acts on the reader cognition. The same words can act differently depending on the age and state of the person receiving them.

3A four-year period

The obligated amount is $1,200,000 over four years, from 2026-10-01 to 2030-09-30. Across the 120 NSF awards this site holds as of 2026-09-02, durations have a median of three years, a maximum of six and a minimum of one, so this sits above the median. Generative AI moves quickly, and models and patterns of use will change within four years.

Building a framework rather than countermeasures for a particular model reads as an acknowledgment of that pace.

4The trustworthy cyberspace program

The program is Secure and Trustworthy Cyberspace (SaTC 2.0). It includes classical security subjects such as intrusion and cryptography, but also awards like this one concerning the psychological safety of users. The shift in what is being protected — from systems toward people — shows in the composition of the program. Amounts are the obligated amount as of the check date and may change.