Twenty-five Fields Medal winners question how AI labs present mathematical solutions and warn of risks to attribution, verification, and open research.
The race by artificial intelligence laboratories, such as OpenAI, to solve mathematical problems of maximum difficulty is triggering an unprecedented reaction within the scientific community.
Twenty-five mathematicians awarded the Fields Medal, considered the highest honor in mathematics, signed an open letter questioning the practices of AI labs when announcing solutions to research problems that remain open.
The conflict intensified this week after Tristan Buckmaster, a professor at New York University, accused OpenAI of pressuring him not to publicly acknowledge the work of a researcher linked to Anthropic who had solved a relevant mathematical problem.
Buckmaster also raised doubts about whether work done with Codex might have contributed to OpenAI later developing its own proof during an extensive inference session. OpenAI has not publicly confirmed that interpretation.
The problem does not end when a proof appears
For the mathematicians who signed the letter, solving a problem is only one part of the scientific process. A proof must be able to be examined, understood, communicated, related to previous work, and subsequently incorporated into the discipline’s accumulated knowledge.
That process becomes especially complex when an AI lab rapidly presents a solution generated by advanced models. Researchers warn that competitive incentives may favor premature announcements before there is sufficient time to fully verify the proof, identify novel methods, recognize previous contributions, and establish how the result fits within mathematical literature.
The issue has a fundamental technical dimension. A proof that a machine can generate but that specialists cannot yet verify does not automatically equate to consolidated mathematical knowledge. Independent validation remains necessary to determine if there are errors, hidden assumptions, or connections to preceding work that were not recognized.
Open research enters a pressure zone
The second risk identified by mathematicians affects the culture of open research. Frontier AI labs have computational resources that can be used to rapidly explore a line of research when they detect it could lead to an important result.
The difference in economic scale can modify traditional incentives. An academic researcher or group may work for years on a problem, while a tech lab can allocate vast amounts of computational capacity to try to find a solution before the academic work has been published or fully developed.
This can favor secrecy. If sharing a preliminary idea allows another actor with enormous computational resources to quickly produce a proof and claim the spotlight, researchers will have fewer incentives to disclose methods, conjectures, or intermediate progress.
OpenAI also changes what it means to be an author
The mathematical controversy also raises a problem that has already appeared in programming, engineering, and other intellectual professions: determining what part of the result belongs to the tool and what belongs to the researcher who formulated the problem, designed the strategy, interpreted the results, and verified the solution.
The Leiden Declaration, published in June by a group of mathematicians, had already addressed these transformations and proposed recommendations for researchers, institutions, and policymakers. The new letter takes the debate to a more concrete point: the scientific community needs mechanisms to preserve attribution and traceability when an AI actively participates in generating new results.
This does not mean that mathematicians reject the use of artificial intelligence. Advanced mathematical reasoning capability could become an extraordinary tool for discovering new structures, formulating conjectures, and solving problems that currently remain beyond human reach.
The precedent could extend to all of science
The dispute between OpenAI and part of the mathematical community serves as an early laboratory for a problem that will likely appear in other disciplines.
When an AI can generate hypotheses, designs, proofs, algorithms, or experimental results faster than human researchers, the validation, attribution, and transmission of knowledge will become as important as its generation.
For engineering, the implications are especially relevant. A model could propose a geometry, a control algorithm, a structural solution, or a design modification in seconds.
That does not mean the proposal can be automatically incorporated into a critical system. It will be necessary to know its origin, verify its assumptions, reproduce the results, and determine who assumed the technical responsibility for accepting it.
The mathematical experience thus anticipates a possible rule for AI-assisted science: the ability to produce a solution does not eliminate the obligation to demonstrate why that solution is correct.
SOURCE: https://techcrunch.com/
