25 Fields Medalists Warn AI Math Races Can Erode Human Understanding
- Twenty-five Fields Medalists, including Terence Tao, Maryna Viazovska, Peter Scholze and June Huh, signed a declaration warning that AI companies' mathematical-problem race conflicts with mathematics' aim of conceptual understanding.
- The signatories say problem solving is a proxy, not mathematics' primary goal: a rush to mass-produce true-or-false results can leave too little time to isolate methods, write proofs clearly, cite prior work and discuss them with other mathematicians.
- They warn that rapidly announced AI-conceived ideas may raise attribution and plagiarism questions, and may never enter the mathematical canon without mathematicians who develop, explain and transmit them.
- The declaration says AI can accelerate mathematical study, but argues that its effect depends on decisions by the people controlling the technology; it calls for urgent action by mathematicians, AI companies and society.
- It extends the concern beyond mathematics: training in scientific and creative work develops understanding and the ability to ask new questions, while AI may directly produce outputs built on prior human work.
Hacker News opinions
I think the statement puts the burden on AI companies to propose a replacement mechanism. Tao says they released it quickly because they saw urgency, though he has made more concrete recommendations in his ICM slides.
Why should AI companies be responsible for repairing an institution that their innovation undercuts? That does not follow from pointing out the damage.
Tao has tried to be constructive before. His ICM talk, especially slides 46 to 51, has concrete recommendations, even if the work is unfinished.
If a machine unlocks understanding that no human has, what is the benefit? What incentive remains for people to learn the material from it?
I care about human understanding. Building things we do not understand leaves us exposed to consequences we cannot predict, and the statement's plagiarism concern is being waved away too easily.
I do not buy OpenAI's line that it cannot prove it did not plagiarize. These systems supposedly solve Millennium Prize problems, yet cannot trace basic data flows.
As far as I know, current proofs and disproofs use methods humans already developed, with AI being more exhaustive. I do not see current systems inventing the new methods needed for problems that resist existing approaches.
This is a worry across intellectual work. AI can make us dumber while producing more answers.
I see the same failure mode in nonfiction books. LLMs reportedly make up 80 to 90% of new arrivals in many Amazon nonfiction categories, crowding out human writing while extracting money from the system.
Why must humans understand if a cheap, inspectable, reproducible, widely available GPT-7 can do the work? Most people do not understand combustion engines or semiconductors either.
I expect a highly productive dark age in mathematics, where publish-or-perish incentives make knowledge grow faster than understanding and erode the skills needed to understand it. Smaller subfields may still become better places for curious mathematicians.
Existing solutions do not stop anyone from solving a problem for personal development. You can work on Navier-Stokes today, but you may lose grants and prestige, which are separate from conceptual understanding.
I see a conflict between mathematicians and the people funding them. Governments seek practical results, and AI may produce similar results without sustaining mathematics as a social activity.
That assumes governments have a single stable practical purpose and treats mathematics far too narrowly. History does not support either assumption.
There is a real contradiction: AI can make 20 years of education harder to justify, but without educated people, who can guide or inspect supposedly superhuman systems? Even Ronacher says he does not know what Astra is doing.