Computers now check AI-written math proofs, so few mathematicians read them — does a certified proof still mean anyone understands it?

This explores why AI-produced math proofs, such as recent Erdős problem results, can be accepted as correct while few mathematicians read or understand them. The short answer from the corpus: these proofs usually arrive already checked by software, so nobody has to read them to trust them. Erdős problems became a popular AI test bed because they cover many accessible areas of math at many levels of difficulty. Reports on this work treat two outcomes separately: a proof being formally certified, and a human understanding it. The second keeps falling behind the first Why did Erdős problems become a popular AI testing ground?. A typical case is Erdős Problem 728. An AI system produced a proof in Lean, a language that a computer can check line by line. Researchers then had to translate it into ordinary mathematical writing so people could follow it. The Lean proof is verified beyond doubt. Whether readers actually understand the human-readable version has not been tested Did an AI system truly solve Erdős Problem 728 autonomously?.

The less obvious point is that a proof was never only a certificate of correctness. When human mathematicians write a proof, the writing itself builds their understanding, and a published paper has traditionally signaled that someone gained real insight. AI-generated math breaks that link. A paper can be fully correct without anyone having gone through the process that produces understanding Does AI-generated mathematics break the link between proof and understanding?. The Leiden Declaration responds by giving human authors sole responsibility for correctness and credit. It argues that proof has two jobs, establishing certainty and conveying understanding, and that machine checking can only do the first Can AI-generated proofs ever replace human mathematical understanding?. Seen this way, unread proofs are not mathematicians being lazy. Readers simply aren't needed for the one job that machines now do.

There is also a capacity problem. Free automated checking doesn't remove human work. It moves that work somewhere else. Once a proof is checked, someone still has to confirm that the formal statement actually means the math problem people care about. Only experts can do that. One 2026 corpus had about 379 checked proofs for every statement that needed this kind of expert review Does free proof checking actually reduce verification burden?. If proofs pile up faster than experts can review even the statements, reading the proofs themselves drops far down the list. AlphaEvolve shows the same split. Automated scoring reliably confirmed its solutions across 67 problems, but humans could explain why those solutions worked only some of the time. The system even exploited weaknesses in its own checker Can automated scoring verify mathematical constructions without human understanding?.

Not everyone sees this as a crisis. Terence Tao argues that an opaque tool is fine if its output is backed by a reliable checker such as a proof assistant Can opaque machine learning models help prove new mathematics?. AI can also be pointed back at the reading problem. An agentic reviewer that checks proofs line by line found serious errors in STOC and ICML papers that human reviewers had missed Can inference scaling help reviewers catch errors humans miss?. What the corpus doesn't show is any way to make humans understand at the same pace. Machines now handle both verifying and reviewing proofs, and understanding has quietly become the scarcest resource in mathematics.