‘Pure insanity’: Mathematicians will need years to make sense of Axiom’s latest drop

Axiom released a huge research repository containing AI-generated mathematical results, rather than a tightly focused set of finished papers. The manuscripts span many subjects and sit at different stages of verification. Axiom even published guidance to help readers navigate the collection.

The release included nearly 400 top-line results across 719 manuscripts. Some manuscripts contain Lean formalizations, which can let computers check proofs. However, only 300 top-line results had been formalized, and the quality and relationship between code and written claims were inconsistent.

That makes this collection different from research that readers can assess paper by paper with a clear verification record. Mathematicians must first sort useful results from possible errors and unclear writing. Researchers told The Verge that understanding the entire release could take years, while Axiom may continue producing more work.

The scale is unusually large: Axiom released nearly 400 AI-generated results across more than 700 manuscripts. The company identified 719 manuscripts in its formalization figures. That volume alone makes a quick, reliable assessment difficult.

The collection covers combinatorics, several branches of geometry, number theory, theoretical computer science, algebra, topology, probability, statistical mechanics, and mathematical physics. Researchers said even reading the roughly 40-page table of contents and abstracts took much of an hour.

This breadth matters because no single mathematician is likely to assess every area equally well. Specialists must examine work in their own fields, while also checking whether formalizations match the written claims. The release therefore creates both an enormous research opportunity and a major evaluation burden. Mathematicians said understanding everything could take years.

Axiom reported that 300 top-line results out of 719 manuscripts had been formalized. That equals around 42 percent. Fewer than half of the manuscripts therefore appeared to have formal descriptions when The Verge examined the collection.

The percentage does not show that the remaining results are false. It shows that they lack this particular verification route, or were still awaiting it. Axiom acknowledged that the manuscripts were at different verification stages and promised more formalizations as they became available.

Reliability is therefore uneven, not automatically high or low for the entire collection. Even formalized results require inspection, because researchers must check what the code actually proves. The article also reports that statements verified in Lean did not always map neatly onto claims in the accompanying manuscripts. Readers must still evaluate each result carefully.

Lean is a programming language and proof assistant used to express mathematical statements and proofs in a form a computer can check. This matters because a computer can verify whether the formal steps follow logically, reducing reliance on a reader catching every hidden mistake.

For example, a manuscript may claim a theorem and include Lean code formalizing it. If the code successfully checks, researchers gain confidence that the formal statement is logically correct. The article says these formalizations helped assess Axiom’s previous mathematical claims, even when researchers did not fully understand the arguments.

Lean does not make the surrounding manuscript automatically clear or reliable. Researchers must still determine whether the formal statement matches the paper’s claim and whether the formalization covers the intended result. In Axiom’s release, only 300 top-line results had been formalized, so Lean supported evaluation of only part of the collection.

Computer verification checks the formal code, not necessarily the meaning a paper assigns to that code. A formalization can be logically correct while expressing a narrower, different, or incomplete statement. That is why researchers still need to inspect it.

The article gives a direct example: some Lean statements verified by computer did not appear to map neatly onto claims in the accompanying manuscripts. Researchers must check what the code proves, how it relates to the written theorem, and whether important assumptions were included.

This review takes time, especially across hundreds of manuscripts. The article also reports inconsistent quality in the supplied formalizations. Lean is therefore a powerful filter, not an instant substitute for mathematical judgment. It can increase confidence in a precise formal statement, but people must connect that statement to the claimed result and assess the surrounding explanation.

The central risk is an evaluation bottleneck. AI systems could produce more mathematical claims than specialists have time to read, verify, explain, and credit. Researchers would then spend increasing amounts of effort sorting possible discoveries from confusing or erroneous material.

The Axiom release illustrates the mechanism. Hundreds of manuscripts already overwhelmed specialists, and some mathematicians said understanding the collection could take years. Kevin Buzzard said he faced a choice between reading possibly incorrect material, waiting for others, or waiting for formalization before judging its correctness.

Careers and academic status could shift as a result. Useful work might be buried in a flood of low-quality papers, while mathematicians’ roles move toward filtering and verification. The article says careers were upended overnight and academics feared Axiom would move on before they could assess the release. More output could deepen that mismatch.

A trustworthy mathematical result starts with clear definitions and a precise statement. Definitions establish exactly what the objects and conditions mean. A logical proof must then show that the conclusion follows from those definitions and accepted assumptions. Lean can help check this formal structure computationally, as the article describes.

Peer review adds independent scrutiny. Other mathematicians inspect the definitions, reasoning, examples, and limits of the claim. They may find errors or ask for clarification. Attribution supports trust in a different way: it identifies earlier ideas and gives credit, allowing readers to understand the result’s context and check whether claims are genuinely new.

These safeguards work together rather than replacing one another. A Lean proof may verify a formal statement, but researchers must confirm that it matches the manuscript. The article reports poor attribution and mismatches between formalized statements and written claims, showing why formal verification alone is insufficient.