(Image Credit: tungnguyen0905/pixabay)
Jacob Tsimerman, a leading mathematician and 2026 Fields Medal winner, has recently launched the Mathematical AI Safety Institute (MAISI). This institute is a nonprofit organization focusing on building rigorous mathematical foundations for AI safety. Announced September 8th, 2026, the Bay-Area-based MAISI will use mathematics to help address risks that AI systems may pose.
Tsimerman is serving as MAISI’s scientific director, even though he will join OpenAI’s safety and research team. The institute operates independently of OpenAI. Mathematician and AI safety researcher Andrew Critch, who applied algebraic geometry to machine learning models, co-founded MAISI and serves as executive director. The scientific advisory panel consists of Fields Medalist Timothy Gowers, Stanford mathematician and American Mathematical Society president Ravi Vaku, former UK AI Security Institute chief scientist Geoffrey Irving, and Alignment Research Center director Paul Christiano.
Compared to other AI safety efforts, MAISI is more foundational and narrow. The goal is for it to develop definitions, measurements, and solution concepts. Researchers could then precisely define AI safety problems and reason about them with mathematical rigor. The founders say there isn’t an agreed-upon, theory-level definition of what makes a powerful AI system safe.
MAISI operates as a short-term visiting program. Mathematicians spend a semester or two working on AI-safety problems with one another. The organization expects to bring in researchers and mathematicians from the wider AI-safety community. MAISI is different from traditional academic mathematics as it expects to quickly share concepts as live documents instead of putting results into journal papers years later.
The institute covers research topics like AI cooperation, deceptive behavior detection, and verifying AI system properties. Zero-knowledge proofs, a cryptographic technique, are an example application in AI safety. This allows someone to verify if a statement is true without revealing the underlying original information. These types of methods could be applicable for verifying claims about AI systems while keeping proprietary information private.
Coefficient Giving announced that it has given MAISI a $10 million grant in September for support as part of its focus on AI and biosecurity. According to the institute, mathematics is only a part of a larger approach to AI safety. It’s not intended to replace other measures like regulations, infrastructure security, and technical safeguards. It wants to add another 10-30 mathematicians for a program starting in January 2027, which runs through May. This program is expected to grow, adding 30-100 participants for a bigger program later next year.
While AI systems become more advanced, MAISI says that the toughest safety questions may benefit from mathematical rigor. The institute aims to apply methods in fields like number theory and cryptography.
Have a story tip? Message me here at element14.