Policy Research Grants

Read more about our 2026 grant recipients on the HAI blog.

This project is co-funded with the Hoover Institution Technology Policy Accelerator

This project examines how human-centered, interpretable AI systems can support the analysis of large-scale, multimodal open-source data to improve monitoring of nuclear proliferation activities. The multidisciplinary team of computer scientists, policy experts, imagery analysts, and open-source nonproliferation researchers will design and evaluate AI-assisted workflows for two policy-relevant use cases: detection of illicit procurement for nuclear proliferation, and detection of undeclared nuclear infrastructure. These workflows will leverage large language model-based AI agents and vision language models to integrate multimodal data streams, including multilingual text, online multimedia, and satellite imagery to support structured, human-interpretable analysis.

This project builds on prior work supported by Stanford HAI, which developed and tested AI-assisted workflows for satellite imagery analysis in nuclear monitoring contexts. That work demonstrated the feasibility of using foundation models to identify relevant signals across large-scale imagery datasets and highlighted key challenges related to uncertainty. This project extends that foundation by incorporating multimodal data sources, benchmarking foundation-model approaches against traditional CNN baselines. This project will formalize evaluation benchmarks and systematically assess the limitations of these approaches across different policy-relevant use cases to test transferability and reproducibility of the approach.

This project is co-funded with the Hoover Institution Technology Policy Accelerator

The future of AI will depend not only on technological advances, but also on public opinion. Public attitudes will shape how governments regulate AI, how firms invest, and how quickly new tools spread through workplaces. It is, therefore, essential to understand what the public thinks about AI, and why.

This project will build CrossInterviewer, a bilingual AI-powered interviewing tool, and use it to compare public opinion in the United States and China. CrossInterviewer will conduct interviews in both Chinese and English. It will also adapt survey questions based on each individual’s responses, allowing scholars to conduct nuanced, in-depth interviews — previously only possible with human interviewers — at the scale and efficiency of traditional online surveys. Using CrossInterview, we will study three themes:

1. General attitudes toward AI. Surveys consistently find that enthusiasm for AI is far higher in China than in the United States, but we know little about why this difference exists. By probing each respondent’s reasoning in their own words, CrossInterviewer will surface the beliefs, hopes, and worries that produce this gap.

2. Willingness to adopt AI at work. Our prior work mapped where U.S. workers across 104 occupations welcome or resist AI systems. This project will extend the analysis to China, producing the first systematic comparison of how workers in the two countries view AI in their own jobs.

3. Open versus closed models. Some AI developers release their models’ weights publicly, letting anyone download, inspect, and modify them; others keep weights private. Experts continue to debate the tradeoffs, but almost nothing is known about public preferences. We will investigate which approach people favor, and what considerations drive their views.

By analyzing public opinion on these important issues, the project will provide a firmer basis for government policy, corporate investments, and forecasts about the future of AI.

This project is co-funded with the Hoover Institution Technology Policy Accelerator

LLM-based agents are rapidly being deployed in politically consequential areas, including generating advocacy content, conducting policy research, and participating in political discourse. Existing research has documented that foundation models from the US and China handle politically sensitive content differently at the single-model level, with these effects persisting in private deployment. However, agents are not single-model. They operate in multi-step pipelines, read each other's outputs, and increasingly populate political systems where their actions feed downstream policy decisions. Whether foundation-model provenance—the bundled set of training data, alignment choices, and governance/regulatory regime under which a model was developed—matters in deployed agent systems is unknown. This question matters geopolitically because the deployment landscape is asymmetric: open-weight Chinese models (DeepSeek, Qwen) are attractive for operators who want to deploy at scale and avoid platform monitoring. As agents proliferate in the political information ecosystem, model provenance plausibly shapes what gets produced and what reaches decision-makers. 

The central question of this project is: does foundation-model provenance shape what agent-mediated political systems produce, and what does the existing agent-developer ecosystem suggest about how this is unfolding in deployment? We use agent-mediated regulatory commenting as the test case as it is a domain where agent deployment is real, outputs are observable, and political consequences are direct. These findings bear on whether provenance-aware procurement standards are warranted in government and political applications.