Stanford HAI awards $3.5 million to seven teams building AI around human expertise
The grants support medical research experts can check, worker training and classroom assessment. Whether those systems deliver their intended benefits remains to be tested.
Listen to this story
The audio brief
Story brief
3 key pointsThe seven projects span tools for source-checkable medical research, infectious-disease diagnosis, scientific collaboration, legal decisions, worker training and classroom assessment. Their shared test is whether AI can extend human expertise without hiding weak evidence or learning gaps: one platform will let experts inspect and correct its reasoning, while another project will test whether AI-assisted learning holds up without AI. Stanford HAI says the grants bring its total allocation to $31 million across 25 teams; Google.org supported the 2026 program.
- 01
Each team receives $500,000, for a total commitment of $3.5 million.
- 02
The Scientific Researcher platform will start with breast cancer, reproducing landmark studies before examining outcomes for groups underrepresented in existing evidence.
- 03
A proposed self-driving microscope targets infectious-disease diagnosis in low- and middle-income countries, with clinicians in the Global South helping shape it.
Seven Stanford research teams are getting funding to build AI that helps people investigate treatments, interpret scientific data and learn new skills. Stanford’s Institute for Human-Centered AI named its 2026 Hoffman-Yee Research Grant recipients on November 16, awarding $500,000 per team—a $3.5 million commitment across medicine, science, law, education and work.
The program requires proposals to address AI’s human and social effects, augment human capabilities, and develop technology inspired by human intelligence. Google.org supported the program in 2026. Stanford HAI says it has allocated $31 million to 25 teams to date.
Medical evidence that experts can inspect
One team is developing Scientific Researcher, an open-source medical research platform designed to connect findings to verified sources. Experts would inspect and correct its reasoning, while theorem provers—tools for checking logical reasoning—would help check its work.
Starting with breast cancer, the platform will investigate unanswered questions using medical literature, patient records and public datasets. The team plans first to reproduce landmark studies, then examine treatment outcomes for patient groups poorly represented in existing evidence.
Another team plans an AI-based self-driving microscope for infectious-disease diagnosis in low- and middle-income countries. It also intends to create an open, multi-disease microscopy dataset, with clinicians in the Global South helping design the technology.
Scientific teams and public decisions
- Self-improving scientific models: One team aims to build focused systems that generate hypotheses, break problems into smaller questions, test solutions and learn from experience alongside human experts.
- An astronomy co-scientist: Another team proposes cooperating AI agents that draw on diverse data, expose evidence for challenge and coordinate follow-up observations.
Most AI systems for science are designed to help individual researchers. We need a system that works inside scientific teams to preserve context as collaborators come and go
Risa Wechsler, director of the Kavli Institute of Particle Astrophysics and Cosmology
The legal team plans systems supporting lawyers and policymakers in housing policy, benefits adjudication and other high-stakes settings. After initial development, it intends to pilot them with government agencies and public institutions.
Does AI help people learn—or hide the gaps?
The workforce project will use an AI interviewer to collect workers’ perspectives for TaskNet, an evolving map of professional tasks. Those findings would inform personalized learning pathways, tested through students’ practice with AI agents in a month-long study called Upskilling Together. The team wants to test whether augmentation is economically preferable to automation.
The classroom team will build and test assessment tools to distinguish durable learning from gaps masked by AI assistance. Its central question is whether a score earned with AI’s help holds up when the AI is absent.
Sources
- hai.stanford.eduStanford HAI Backs Ambitious AI Projects with 2026 Hoffman-Yee Grants | Stanford HAI
Reader comments
Newest comments first. Replies stay oldest first.