OpenAI's publication of more than 300 mathematical proofs produced with artificial intelligence has sparked concern in the academic community. Amid new claims of solutions to fundamental problems in mathematics and computer science, a question has emerged: are we approaching the moment when machines definitively surpass humans in mathematics? However, Dan Margalit, chair of the mathematics department at Vanderbilt University, offers a less pessimistic view. In his opinion, AI's current achievements do not negate the human role but largely build on decades of human research.

In an Axios article, Margalit highlights five reasons why mathematicians should not see artificial intelligence's successes as the end of their profession. Chief among them is that mathematics is not simply about getting the right answers, but about trying to understand why they are right.

AI reaches solutions by following paths humans have already laid out

OpenAI's latest achievements are indeed impressive. The company has published hundreds of mathematical papers presenting solutions to difficult problems, and previously claimed to have solved a famous Millennium Prize Problem: the existence and smoothness of solutions to the Navier–Stokes equations.

OpenAI then announced that its models had disproved the exponential time hypothesis, one of the fundamental conjectures in theoretical computer science. On the same day, Anthropic announced that it had independently obtained a similar result. According to The Information, the company's models used about 14 million tokens on this work.

Such developments reinforce the impression that AI is beginning to advance entire scientific disciplines on its own. But Margalit suggests paying attention to exactly which problems the machines were able to tackle.

According to him, AI primarily solved problems that researchers had already come particularly close to solving. Human publications and accumulated knowledge had effectively laid the groundwork for a breakthrough. Meanwhile, the systems have not achieved comparable results on thousands of other problems for which the scientific literature has yet to point to a sufficiently clear direction.

In this picture, artificial intelligence acts less as a pioneer than as an implementer capable of bringing to completion ideas that humans have spent a long time developing and refining.

AI's mathematical proofs still need to be checked

Another reason not to jump to conclusions is that the proofs themselves contain errors. Even if a system produces a convincing mathematical text, that does not guarantee that its reasoning is flawless.

Margalit notes that mathematicians found logical errors when reviewing the published work. According to him, some papers may be withdrawn because of flaws in their proofs. In this respect, AI repeats typical human mistakes: skipping important steps, drawing unjustified conclusions, or presenting faulty reasoning as correct.

For this reason, the mathematical community cannot yet simply take machine-generated results on faith. They must be carefully examined and checked, just like proofs submitted by humans.

This does not diminish the value of the new tools, but it shows that the ability to generate proofs and the ability to guarantee their correctness are different tasks.

Machines explore possibilities rather than ask interesting questions

Margalit also suggests comparing AI's current mathematical achievements with computers' successes in chess and Go.

According to his account, OpenAI deployed tens of thousands of agents that repeatedly combined known theorems in millions of different configurations. This approach allows a machine to work without tiring and test an enormous number of possible paths to a solution.

This is impressive computing power, but exploring possibilities does not in itself imply the emergence of mathematical intuition. AI can efficiently search for a proof within the space of already known ideas, yet, in Margalit's assessment, it does not currently demonstrate a comparable ability to formulate genuinely interesting questions on its own.

The distinction here is fundamental: finding a solution to a given problem is not the same as understanding which problem is worth posing in the first place.

AI can free mathematicians from routine work

Margalit also sees a more practical opportunity in artificial intelligence: turning it into a general-purpose computational tool.

He points out that mathematicians once devoted entire books to calculating integrals. The advent of specialized software automated this work, but did not destroy mathematics. On the contrary, researchers gained the opportunity to focus on more difficult questions.

Something similar is happening today. If a complex research problem involves an integral that can be calculated using software, a mathematician no longer has to spend a substantial amount of time on it. AI can extend this effect by helping researchers complete the technical stages of their work more quickly.

As a result, researchers can devote more attention to concepts, connections between mathematical objects, and new directions for their work.

A correct answer is not the same as understanding

Margalit's most important argument concerns the very nature of mathematics. In his view, its purpose goes beyond obtaining correct results.

He draws on the ideas of mathematician Bill Thurston, set out in his well-known 1994 essay. If a computer produces a list of ten thousand prime numbers, that does not explain why they exist or what patterns connect them.

A machine can provide an answer, but an answer in itself does not necessarily bring understanding. It is the human desire to grasp causes, structure, and connections that turns computation into mathematical research.

Margalit emphasizes that even if a computer can solve a problem, this does not mean that people will stop wanting to understand the solution. For him, mathematics is above all a human activity aimed at understanding the world around us.

What role remains for humans in the age of AI?

The achievements of OpenAI and Anthropic show that artificial intelligence is becoming an increasingly powerful tool for scientific work. It can explore an enormous number of possibilities, combine known results, and find solutions to problems that previously required considerable effort from researchers.

But this does not mean that human contributions are losing their importance. Machine-generated proofs need to be checked, many successful solutions build on work humans have already done, and generating an answer is no substitute for understanding.

Margalit suggests viewing these developments not as a competition in which humans must beat machines, but as another stage in the evolution of scientific tools. People have already learned not to worry about computers playing chess better than they do. Perhaps something similar will happen in mathematics: AI will solve more and more problems, but the human need to understand how the world works and why mathematical truths hold will not disappear.