AI solves open math problems, raising questions about human understanding

Google DeepMind’s AlphaProof Nexus solved nine open math problems in a study published in Science.

The problems came from a collection linked to mathematician Paul Erdős, and the system used Gemini-powered agents, math software called Lean and, for difficult steps, a model trained for mathematics.

Researchers described how the system reached its results, and human mathematicians checked many of its proofs.

Two days later, OpenAI posted 722 mathematical manuscripts produced by an unreleased AI model, but the account gives no comparable details about how those results were made.

Mathematicians are concerned that AI-generated proofs may be hard to understand, even when they appear to solve a problem.

Thomas Bloom, a University of Manchester mathematician who runs a website tracking Erdős problems, stopped marking problems as open or solved and closed comments after people increasingly posted AI proofs they did not understand.

The challenge is to make AI advance mathematics in ways people can check and build on; researchers say AI is better at finding answers than judging their research significance.