Trust but Verify: Evaluating AI for the Public Interest
The Vals AI Public Sector Summit.
About: Government is making calls on AI without much evidence to go on: whether these chatbots are safe for teens, how far ahead of China the US actually is, and what role government itself should play in evaluating any of it.
Trust but Verify convenes the policymakers, agency leaders, and researchers building the evidence base that makes those calls defensible. Hosted by Vals AI at Station DC.
Overview
Youth Wellbeing in the Age of AI: Clinical, policy, and child safety perspectives on what it actually means for an AI system to be safe for young people: what current evaluations can and cannot tell us, and what evidence companies should be expected to produce before releasing a product to teens.
Negotiating AI with China: As US and China leadership meet this month to talk AI, a discussion on what the United States should actually be willing to agree to, where Chinese model capability sits, and which commitments could be verified.
Government's Role in Evaluating AI: What an independent verification body would look like in practice, who it answers to, and how the government can adequately resource it.
Plus research poster presentations from the Vals AI team spanning our research in AI Child Safety, Cybersecurity, and Environmental Impact. Reception with food and drink to close.
Agenda
1:30 to 2:00 Doors open: check-in, refreshments, research poster boards
2:00 to 2:15 Opening remarks
2:20 to 2:30 Topic overview: Measuring AI Safety for Youth
2:30 to 3:10 Panel: Youth Wellbeing in the Age of AI
3:15 to 3:25 Topic overview: Measuring AI in National Security
3:25 to 4:05 Panel: Negotiating AI with China
4:05 to 4:15 Break
4:15 to 5:00 Government's Role in Evaluating AI
5:00 to 6:00 Reception
Speakers
To be announced. Full lineup shared closer to the event.
About Vals AI
Vals AI is an independent, third party evaluator of frontier AI models, backed by Andreessen Horowitz. Our results are used by the major AI labs and by the US government, and we brief policymakers in the US and abroad on what current evaluations can and cannot show.
Audience is roughly 100 policymakers, agency leaders, researchers, and advocates. Bipartisan by design. Space is limited and registration is required.