Federal Research Lab for Human Harms

Nearly two-thirds of American teenagers use AI chatbots, and about three in ten use them daily. Washington requests billions to advance AI, yet its national youth health survey asks nothing about it, and its research on technology and children has centered on social media. Congress should count AI use in existing surveys, follow people over time, test products in real settings, and make the work permanent in a federal lab.

The Problem

In fall 2025, 64% of American teens said they use AI chatbots, about three in ten of them daily.¹ Children are forming habits with these tools while their judgment is still developing.

The country has run this experiment before. Despite years of research, the National Academies concluded in 2023 that the evidence did not support the conclusion that social media causes changes in adolescent health at the population level,² and that rigorous study designs were "severely lacking."³ Evidence that arrives after the habits set cannot guide parents, schools, or lawmakers.

The federal effort to learn faster is small. The President's 2026 budget requests $2.8 billion for AI research and development;⁴ the program Congress created in 2022 to study technology's effects on children received $15 million a year in fiscal 2023 and 2024, according to its sponsors.⁵ Three gaps keep the country guessing:

1. Surveys skip AI. CDC's national survey of high school students first asked about social media in 2023, when 77% reported using it at least several times a day.⁶ Its 2025 questionnaire asks 106 questions, none about AI.⁷
2. Cohorts were built for screens. The largest long-term U.S. study of brain development and child health, following 11,880 children from 21 sites into adulthood, was funded by NIH in 2015 to examine experiences such as sports, video games, and social media.⁸
3. Funding follows social media. The 2022 law directing NIH research on media's effects on children names artificial intelligence but authorizes no dedicated funding,⁹ and every funding call on NIH's program page has expired, most of them on social media.¹⁰

Why legislation: The obstacle is mission and money: HHS, home to NIH and CDC, already has broad authority to do this research.¹¹ Congress has legislated here before: it enacted the children's media research law in 2022,⁹ and in June 2026 a House committee advanced the CREATE AI Act, which would give researchers shared computing and data, by a vote of 29 to 0.¹² The lesson of social media is to build independent evidence early. Parents deserve to know what these tools do to their children, and the answer should not depend on the companies that sell them.

The Solution

A four-step staircase: each step stands alone, and each step up adds public knowledge and commitment. The lab studies people, not models, and it evaluates products instead of building them; testing what AI systems can do and funding nonprofit builders are separate proposals.

Step 1 — Count AI in the surveys we already run. Add questions on AI use to CDC's youth survey, NIH cohorts, and other national surveys, with common measures across them. It is the cheapest way to see who uses what, how much, and with what associated outcomes.

Step 2 — Follow people over time. Fund consented, long-term cohorts of children and adults, secure data environments, and common outcome measures, open to qualified independent researchers. Children join only with a parent's consent, their own assent, and full human-subjects protections.

Step 3 — Test in real settings, and publish everything. Compare products, design choices, and non-AI alternatives in schools, workplaces, clinics, and public services, measuring what people can do unaided, lasting outcomes, harms, and costs. Keep the teams that design fixes, such as better tutoring or memory controls, separate from those who judge them, and publish methods, null results, replications, and the model versions tested.

Step 4 — Make it permanent. Charter a lead research center with a network of field partners, housed where HHS, NSF, and NIST jointly recommend within 12 months, with protected publication and evidence reviews for Congress but no enforcement power. If voluntary data access fails, Congress may require large AI companies to give approved studies de-identified data, as a 2026 House bill proposes.¹³

Where to start: Step 1 is the floor: it adds questions to surveys that already run. Step 2 is the heart of the proposal.

Administration and enforcement: HHS leads with NSF and NIST, delivering an institutional plan within 12 months and first awards within 18, subject to appropriation. Independent scientific review, inspector-general oversight, and a five-year outside evaluation keep the lab honest; any mandatory data access needs its own privacy safeguards and due process.

Risks and Mitigations

• Research as political validation: Preregister major studies, protect publication, test competing hypotheses, and allow independent replication. Science can estimate consequences; elected officials still make the value judgments.
• Privacy: Use consented cohorts, minimal collection, and secure enclaves, and build no repository of Americans' conversations. Any large dataset carries risk that design can reduce but never remove.
• Slow science, fast products: Pair quick survey measures with long cohorts, and label short-term findings as short-term. Some answers will still arrive after the products have changed.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal

State

What this adds: Congress has authorized research on media's effects on children and is advancing shared computing for researchers. This proposal adds what neither provides: routine national measurement of AI use, long-term cohorts of children and adults, real-world trials with published results, and a permanent home for the work.