Project Grant 2541285

Award Summary

Brown University received a $442,685 CAREER award from the National Science Foundation's Division of Computing and Communication Foundations (CFDA 47.070: Computer and Information Science and Engineering) beginning September 1, 2026, through August 31, 2031. The project develops computational foundations for learning artificial intelligence (AI) specifications from imperfect human input by integrating reinforcement learning, Bayesian inference, and computational cognitive modeling with empirical studies of human decision-making. The research addresses a critical safety challenge: as AI systems increasingly influence high-stakes and everyday decisions across economic sectors, the systems depend on human-provided instructions that are often too rigid, too vague, or incorrect due to limited information and cognitive biases. The deliverables include probabilistic models of human reasoning that capture systematic specification errors, enabling AI systems to better interpret and infer human intent from flawed instructions.

The three-year research initiative is organized around modeling and inferring AI specifications, characterizing how people communicate goals to machines, and identifying where specification errors arise. Outcomes include safer decision-making technologies and new tools to help organizations deploy AI more effectively. The work conducted at Brown University's Providence, Rhode Island campus will advance foundational understanding of how intelligent systems can interpret imperfect human guidance, directly supporting NSF's mission to advance computing research that enhances both the reliability and safety of AI-driven applications across the economy.