AI Can Prove Theorems. Only Humans Can Decide Which Ones Matter

A theorem becomes powerful through human intuition, creation and collective judgment. An AI-generated proof does not replace this process at the heart of mathematics.

12 October 2026

“They have annihilated my field of research,” said Fields medallist Hugo Duminil-Copin, according to Le Monde. [1]

A conjecture can be the basis of a doctoral thesis or the focus of decades of research. In the past week, machine-generated answers have arrived for problems mathematicians were still trying to solve.

Many mathematicians have seen their careers disrupted. Their field of research is gone. And why would they move on to the next one if it, too, can be swept away by a single OpenAI announcement?

They can't even take part in this new way of doing maths: the powerful model that has proven all these theorems is not available to them.

Yet this shock brings an earlier human act into focus: OpenAI resolved hundreds of questions.

But this was possible only because human mathematicians had already decided which theorems were worth proving. That judgment had emerged from a uniquely human mix of intuition, creation and recognition within a human community.

That judgment has no universal scientific definition. A machine cannot settle it on our behalf by producing a proof or declaring a result important.

A machine cannot (yet?) know which of the infinitely many theorems one can formulate actually matter.

A research programme, suddenly overtaken

On 6 October, OpenAI published hundreds of mathematical manuscripts, at different stages of verification. [2, 3]

“So we've had whole research programs wiped out on Tuesday by their dump,” NYU mathematician Tristan Buckmaster told CNBC on 8 October, as reported by Business Insider. [4]

The results were public, but the model was not generally available. OpenAI said it was working to release it. Moving to the next problem offered little reassurance: the same machine might solve it too. [2, 3]

How does a theorem become powerful?

But choosing the next unsolved problem leaves a more basic question unanswered: why should that problem deserve our attention?

Mathematics offers infinitely many possible statements. Only some attract the sustained attention of a research community. A difficult result can remain isolated; a short theorem can open connections that transform an entire field.

But the effort required to obtain an answer does not, by itself, determine its importance.

Mathematicians describe results as remarkable, powerful, beautiful, singular or strong. In What is good mathematics?, Terence Tao explores several forms of mathematical quality, including insight, usefulness, beauty and vision. [5]

There is no universal scientific definition of what makes a mathematical result powerful. That judgment emerges from an intuition, a creative proposal and a social process that can unfold over decades.

How? A mathematician sees a possibility worth pursuing. Others investigate it, challenge it, use it and connect it to questions of their own. Through these exchanges, a result acquires the significance that leads a community to call it powerful.

Now consider the reverse. A machine produces a theorem and declares it powerful. That declaration does not make it so: until a human community recognises its significance, the machine has supplied an assessment, not conferred mathematical importance.

Only human judgment gives these questions and results their value as objects of research. Mathematics may or may not be universal; what deserves our time and attention is constituted through this human process.

The intuition that comes before the proof

The models did only half the job here: producing the proofs.

But it was human mathematicians who had formulated these questions and made them part of a research agenda before the models tackled them. [3]

We recognise the significance of the answers only because people have found reasons to care about the questions.

A mathematician's belief that a particular theorem might be worth proving is akin to Picasso's artistic wager that representing a face differently might reveal something worth seeing.

A mathematician makes a similar proposal when they say that discovering whether this conjecture is true or false would be worth the effort.

In mathematics, the selection of a question is therefore a uniquely human act. It proposes something for others to care about, and the ensuing conversation gives that proposal its place in mathematics.

There is no prompt for that

There is no prompt of the form “Find the theorems that could help achieve X and prove them,” where X is anything other than another theorem.

Here, X is the direction of research. That direction is only valid when a mathematician sets it through intuition, in a purely creative act, and human peers then validate it.

If humans aren't involved, it's worthless, by the very definition of how maths works.

Maths, by design, is all about finding results that mathematicians deem singular, powerful, beautiful and strong. These words have no scientific definition. They provide a common language for something felt and shared within the mathematical community.

A goal that no specification can supply

An engineering problem can begin with an external objective: build a bridge that carries a specified load within a given budget. Once that objective is set, it provides a way to assess a proposed solution.

Pure mathematical inquiry does not begin with such a specification. Today, among the countless conjectures we could formulate in pure, abstract mathematics, there is usually no concrete external deliverable to tell us which one deserves to become a research programme.

Proving a statement establishes an answer to the question we selected. It does not supply the reason for selecting it.

The value was never in proving an arbitrary statement. It was in deciding that this particular question deserved to be resolved, whether by a proof or a counterexample.

That choice creates the purpose of the work. A machine may produce the answer, but the answer matters because people have made the question matter.

Choosing where discovery should go

This changes what mathematicians must make central to their work, but not what that work has always been about. As AI expands the range of answers they can obtain, they will need to formulate questions, develop intuitions and discuss which directions deserve sustained effort.

A researcher might ask an AI to investigate a conjecture or search for counterexamples. An unexpected result might then justify changing the question entirely.

The decisions are concrete. Which possibility deserves a month of computation? Would a simpler proof teach us more? Does a new direction matter enough for others to devote years to it?

Directing research means continually judging what is worth understanding, then making that judgment available for others to challenge. And this is nothing new.

Access to the new tools

Access to the tools matters here. In its 6 October statement, the Advisory Group on Mathematics and Artificial Intelligence defends mathematicians' freedom to pursue their own questions and approaches, beyond those selected by AI laboratories. It connects that freedom to equitable access to research tools and computational resources. [6]

The profession will need to recognise the work involved in proposing consequential questions, redirecting an investigation and helping a community see why a discovery matters.

AI can produce answers. Only a human community can give those answers a place in what it considers worth knowing. And that is the essence of “doing maths”.

Sources

Sources checked on 12 October 2026. Statements about model availability refer to the 6 October announcement. The argument about mathematical value and its comparison with engineering expresses the author's position.

[1] Le Monde, report on Hugo Duminil-Copin's reaction to OpenAI's mathematics release, English edition, 12 October 2026. Opening quotation in the newspaper's English translation.

[2] OpenAI, Sharing AI progress in mathematics, 6 October 2026.

[3] OpenAI, Mathematics repository, README, accessed 12 October 2026. Verification status, named conjectures and description of how the results were produced.

[4] Business Insider, report on Tristan Buckmaster's CNBC interview, 10 October 2026; interview broadcast 8 October.

[5] Terence Tao, What is good mathematics?, 2007, section 1.

[6] Advisory Group on Mathematics and Artificial Intelligence, On OpenAI's Release of Mathematical Results, 6 October 2026.