Mathematician Jacob Tsimerman is creating an independent institute that will try to replace informal AI safety assurances with statements that can be proved. The Mathematical A.I. Safety Institute is expected to begin work in the San Francisco Bay Area in January 2027 with between 10 and 30 mathematicians.
Tsimerman, a Fields Medal recipient who is also joining OpenAI’s safety team, argues that the field needs a much higher standard. Cryptographers can sometimes prove that a scheme remains secure under clearly stated assumptions. AI systems lack both an equivalent shortcut and a broadly accepted theoretical definition of “safe,” making the institute’s goal unusually difficult.
Initial questions include whether a system can be shown to act responsibly and produce correct results, whether interacting agents can avoid unwanted collective behavior, and whether defenses can hold against vulnerabilities that have not yet been discovered. One possible technique is a zero-knowledge proof, which demonstrates that a condition is true without revealing the private information used to establish it. That might let laboratories support safety claims without disclosing model details.
The institute is announcing a research agenda, not presenting a completed safety proof. Its value will depend on converting broad social requirements into precise assumptions that mathematics can test—and on whether those assumptions match how deployed AI systems behave.