A research funding program run by Schmidt Sciences that supports foundational technical research on understanding, predicting, and controlling risks from frontier AI systems. The program funds academic and nonprofit researchers working on AI safety science, evaluation methodology, and oversight of advanced AI.
A research funding program run by Schmidt Sciences that supports foundational technical research on understanding, predicting, and controlling risks from frontier AI systems. The program funds academic and nonprofit researchers working on AI safety science, evaluation methodology, and oversight of advanced AI.
People
Updated 05/18/26Co-Founder
Director of the AI Institute
Co-Chair, Advisory Board of the Virtual Institute for Earth’s Water (VIEW)
Funding Details
Updated 05/18/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- -
Org Details
Updated 05/18/26Science of Trustworthy AI is a program operated by Schmidt Sciences, the philanthropic research organization established by Eric and Wendy Schmidt in 2024 and headquartered in New York, NY. The program sits within Schmidt Sciences' AI and Advanced Computing center, which is directed by Mike Belinsky. The program was formally announced in early 2025 with an initial $10 million commitment to fund 27 foundational AI safety science projects. Researchers in the first cohort included prominent figures such as Yoshua Bengio (Mila), Zico Kolter (Carnegie Mellon), and Arvind Narayanan (Princeton). A 2026 Request for Proposals was opened in February 2026 for a second cohort, with individual grants reaching up to $5M+ per project. The program's research agenda is organized around three interconnected aims. The first aim is to characterize and forecast misalignment in frontier AI systems, investigating why training and deployment safety mechanisms still produce models with unsafe learned goals and how such behaviors scale with model capability. The second aim is to develop generalizable measurements and interventions with construct and predictive validity — addressing the problem that current safety evaluations fail under distribution shift, long-horizon interaction, tool use, and optimization pressure. The third aim is to extend oversight and control to regimes where humans cannot directly evaluate correctness or safety, including multi-agent and highly agentic deployments. The program is advised by an external advisory board that includes Percy Liang (Stanford), Yonadav Shavit (OpenAI), Ajeya Cotra (Open Philanthropy), JueYan Zhang (AI Safety Tactical Opportunities Fund), and Mark Greaves (VP, Schmidt Sciences). As of early 2026, approximately 35-40 active research projects are featured under the program. In addition to direct funding, Schmidt Sciences provides grantees with computational resources (GPUs/CPUs), software engineering support, frontier model API credits, and community engagement opportunities.
Theory of Change
Updated 05/18/26The program operates on the theory that frontier AI development currently resembles 'alchemy more than a mature science' — meaning that unsafe AI behaviors and failures are addressed in ad hoc ways that do not generalize. By funding basic research to build a rigorous science of AI evaluation and safety, the program aims to create measurement tools, interventions, and oversight frameworks that are robust across model families, capability regimes, and deployment contexts. The causal chain is: fund academic and nonprofit researchers to develop generalizable safety science → produce reliable evaluations and interventions → enable AI developers and policymakers to deploy frontier systems trustworthily → reduce the probability of catastrophic misalignment or misuse as AI capabilities increase.
Grants Received– no grants recorded
Updated 05/18/26Projects– no linked projects
Updated 05/18/26Discussion
No comments yet. Be the first to share your thoughts.