Fields Medalist Breaks Down: AI 'Carpet Bombs' Mathematics Research

On October 6, OpenAI published 722 mathematical manuscripts on GitHub, organized into 372 result families. The collection contains 377 new proofs, with roughly one-third addressing major conjectures, another third offering significant value in their subfields, and the remainder comprising smaller results or variants of existing proofs. OpenAI stated the vast majority were produced by a single unreleased internal model trained from August 28, with each result consuming compute equivalent to about three hours of ChatGPT Pro reasoning time.

Immediate Fallout and Verification Concerns

Within days, mathematicians examining the manuscripts found that some did not actually prove their claimed conclusions. Even when a proof passes Lean formal verification, that only confirms internal logical consistency of the formalized statement; it does not guarantee the formalization faithfully captures the original natural-language argument. OpenAI subsequently retracted three erroneous papers, leaving 719 manuscripts in the repository.

The Value of Mathematical Research Beyond the Proof

The article emphasizes that mathematical research's value extends far beyond the final proof. In the process of seeking proofs, researchers introduce new concepts, forge new tools, and uncover connections between fields. Failed attempts often open doors to other problems. Hugo Duminil-Copin recounted that his multiple failures while attacking a probability conjecture generated a wealth of ideas later used in other research, yielding unexpected discoveries. AI delivering proofs directly may bypass this exploratory process.

Twentieth-century mathematician William Thurston reflected that after he rapidly proved a batch of theorems, the research direction gradually lost attention. As researchers left, the accumulated intuition, experience, and communication networks of that field began to dissipate. In his view, mathematical knowledge resides not only in papers but also in researchers' minds and the academic community.

Young Mathematicians Bear the Brunt

Traditionally, advisors assign graduate students problems expected to be solvable in a few months. Through repeated attempts, students train their intuition and learn to handle increasingly abstract and complex research. Today, exactly these training-ground problems may be swiftly solved by models. Established mathematicians can pivot to new directions, but PhD students and postdocs lack that luxury; their publications, positions, and career prospects often hinge on a handful of problems. If AI solves them first, years of accumulated work may lose value.

Open exchange may also suffer. Mathematicians routinely share nascent ideas at conferences, via email, and in discussions. If AI companies can harvest these clues and direct models to attack them, researchers may reduce public discourse. Duminil-Copin worries that the free-exchange culture sustaining mathematics is fracturing.

Emergence of "Alien Mathematics"

Reviewers of one AI-generated proof discovered the model combined two previously rarely connected ideas. The proof was completely correct and withstood scrutiny, yet its reasoning felt alien — dubbed "alien mathematics" by observers. Commentators suggest similar phenomena may soon appear in physics, chemistry, materials science, medicine, and engineering. The mode of scientific production is shifting, and new research paths may no longer follow human cognitive habits. The article concludes that today may be the last moment when science feels entirely "non-alien."

References: https://www.theatlantic.com/technology/2026/10/openai-math-proofs-future/688940/ https://x.com/miniapeur/status/2109237567832949109 https://x.com/haider1/status/2109353337489645622 https://x.com/imjustnewatai/status/2109102057885991250