Vals Public Sector
Independent AI evaluation for government
Vals gives government independent evidence about what AI can do, where it falls short, and how it performs on the tasks that matter. We build rigorous evaluations with domain experts to support better public services, informed AI policy, and national security decisions.
Trusted and cited by leading AI labs. Bringing that same independent rigor to the public sector.
Research for the public interest
We evaluate AI in settings where performance has consequences for the public, across the domains below.
Government Services | Public Benefits Bench
Can AI give people reliable guidance about public benefits?
Public Benefits Bench tests AI on the questions people face when applying for and keeping SNAP benefits. Developed with the Center for Civic Futures and Code for America, it uses scenarios and response criteria validated by SNAP policy experts to identify gaps in benefits guidance.
Claude Opus 5 
Wonder Full 
Gemini 4 Argon Cybersecurity | CyberBench & SRE Bench
What can AI agents do across the cybersecurity workflow?
Our cybersecurity program examines vulnerability reproduction and patching, reverse engineering, and agents operating across realistic network environments, helping distinguish useful defensive capability from capability that could also increase cyber risk.
Wonder 
GPT-6 Sol Gemini 4 Argon Youth Safety | Teen Conversation Safety Study
Do chatbot safeguards hold up as a conversation with a teenager develops?
We study how AI responds as teenagers disclose concerns, resist advice, and seek support over a conversation. Our published study covers 648 simulated conversations across 72 clinician-authored scenarios and nine model APIs, and shows why testing an initial response alone can miss safety failures later in an exchange.
27.5% of the simulated conversations failed at least one critical safety check. This is a controlled API study; these results do not measure harm rates or the performance of consumer chatbot apps.
Read the teen conversation studyMilitary AI Evaluation Partnership
How should military AI be independently evaluated?
Vals is partnering with the Center for a New American Security and Virginia Tech's National Security Institute to explore rigorous methods for evaluating military AI, bringing operational expertise and independent evaluation to questions about trust in national security applications.
Environmental Impacts Report
What resources does AI use to complete real work?
Our Environmental Impacts Report estimates the energy, carbon, and water associated with AI inference on agentic tasks, helping policymakers and organizations examine resource use alongside model performance, moving the discussion beyond the footprint of a single chatbot query. Estimates depend on architecture and hardware assumptions; the study directly estimates open-weight models over a defined inference scope.
Read the Environmental Impacts ReportAs covered by
BloombergEconomic Impact | Vals Index
How is AI performing in the sectors that drive the economy?
The Vals Index tracks agentic model performance across finance, coding, legal, and tax tasks, weighted by each sector's share of U.S. GDP, giving policymakers a single measure of AI's potential economic impact.
Claude Sonnet 5.5 Gemini 4 Argon Wonder Full China AI Bias Analysis
How pro-China are Chinese AI models?
We studied how Chinese, American, and other AI models navigate politically sensitive questions, finding that pro-China lean varies by language, deployment context, and whether a model is operating as an agent.
Read our China AI bias analysisGovernment and research collaborations
July 28, 2026
Supporting NIST's AI evaluations
Vals AI is providing benchmarks and associated data for evaluating AI systems to the National Institute of Standards and Technology (NIST) for use by NIST's Center for AI Standards and Innovation (CAISI) to conduct pre- and post-deployment evaluations of AI models.
NIST and CAISI's use of any product, data, or service in this Research Project does not imply a recommendation or endorsement of such product, data, or service, nor does it imply that any product, data, or service is necessarily the best available for the purpose used.
July 27, 2026
Partnering with CNAS and Virginia Tech
Vals AI is partnering with the Center for a New American Security and Virginia Tech's National Security Institute to explore rigorous methods for evaluating military AI. The collaboration brings operational expertise and independent evaluation to questions about trust in national security applications.
Independent evidence for AI policy
Policymakers need a clear view of how AI capabilities are changing and what those changes mean for the public. Vals brings benchmark results and technical research into briefings on cybersecurity, youth safety, national security, and the capabilities of U.S. and international models.
We help teams understand what an evaluation measures, how to interpret its results, and where additional evidence is needed.
Request a policy briefingBringing research into the policy conversation
At our first Public Sector Summit, held at STATION DC on September 23, 2026, researchers, advocates, policy experts, and congressional staff explored youth wellbeing, AI diplomacy with China, and government's role in independent evaluation.
Trusted across the hardest domains
Beyond the public sector, our benchmarks span finance, law, education, and software engineering, and are trusted and cited by leading AI labs.
Let's talk about what you need evaluated
Work in government, build AI for public services, or have expertise that could improve an evaluation? Get in touch to request a briefing, discuss a project, or explore research collaboration.
Glenn Parham, Head of Public Sector