Why the Future of Mathematics Belongs to Human Judgment

Anonymous

Artificial intelligence is rapidly crossing boundaries once thought exclusive to human mathematicians. It can search enormous spaces, discover patterns, formalize arguments, and increasingly prove difficult theorems. This raises an unsettling question: if machines can out-search, out-deduce, and out-prove us, what remains uniquely valuable about the human mathematician?

The answer is evaluation.

Machines may become extraordinary at producing mathematical truths. But mathematics is not merely the accumulation of true statements. Its deeper task is deciding which truths matter.

From Inherited Problems to an Ocean of New Ones

Much of AI’s mathematical success still takes place on intellectual terrain chosen by humans. Famous conjectures were identified as important because generations of mathematicians judged them to illuminate something fundamental about number, space, symmetry, computation, or the physical world.

But this inherited catalog is finite.

What if all existing conjectures were solved? How could we move forward?

Terence Tao has proposed Math 2.0 for the future. In my view, Math 1.0 could be called the conjecture era, and Math 2.0 is the post conjecture era.

AI companies have exploited the significance of existing conjectures, but they have forgotten why those problems are important.

It’s because we believe they are.

As AI becomes increasingly capable, it will not merely solve existing problems. It will generate new conjectures, definitions, structures, and research directions on a scale no human community could possibly examine.

At that point, the central scarcity in mathematics will no longer be proofs or problems. It will be attention.

Truth Is Not the Same as Significance

Henri Poincaré understood this distinction more than a century ago:

“Invention is discernment, choice. …… The sterile combinations do not even present themselves to the mind of the inventor.” [1]

This insight becomes even more important in the age of AI.

A sufficiently powerful system could generate an enormous number of correct theorems and extraordinarily difficult problems. But difficulty alone does not imply depth. A problem may be hard because it connects profound ideas—or merely because it is an artificial maze created from arbitrary definitions.

The future mathematician therefore need not compete with AI in producing more theorems. The more important role will be to look at what machines produce and ask:

Does this matter?

Does it reveal a hidden structure?
Does it simplify our understanding?
Does it open a path toward physics, computation, or another fundamental question?

A theorem can be perfectly correct and still be mathematically uninteresting.

Mathematical Taste

This is what mathematicians traditionally call taste.

Beauty, depth, naturalness, explanatory power, and importance are not properties captured simply by the formal correctness of a proof. They arise from relationships among mathematics, human understanding, scientific goals, and intellectual culture.

Liu Cixin dramatizes almost exactly this problem in his science-fiction story The Poetry Cloud [2]. A godlike intelligence uses overwhelming computational power to generate the entire space of possible classical Chinese poems, yet still cannot determine which of them constitute genuine poetry. Exhaustive generation solves the problem of production, but not the problem of selection. Selection has been discussed more widely in our community, such as [3-5]. The parallel with AI-generated mathematics is striking: a machine might someday generate an ocean of valid theorems without thereby telling us which ones deserve our attention.

AI may learn remarkably sophisticated models of human mathematical taste. It may even predict which problems mathematicians are likely to find interesting. But prediction is not the same as ultimate valuation.

A machine cannot compel humanity to care about a theorem.

Mathematical importance exists partly because finite human beings choose to devote their attention to certain ideas rather than others.

The Economy of Attention

This is why the future abundance created by AI may make human judgment more valuable, not less.

Human attention is limited. A mathematician has only one career. A research community can pursue only a small fraction of possible questions. Universities can teach only an even smaller fraction of mathematical knowledge.

The decisive questions therefore become:

Which ideas deserve the attention of an entire field?

Which discoveries should become part of our shared intellectual inheritance?

When mathematical production becomes nearly unlimited, selection becomes the bottleneck.

From Provers to Cartographers

AI may therefore transform rather than eliminate the role of mathematicians.

For centuries, mathematicians have combined two roles: conceiving and evaluating ideas, and performing the immense technical labor required to prove them. AI may increasingly absorb the second task—searching proof spaces, verifying lemmas, exploring cases, and constructing formal arguments.

The human mathematician can then move upward in abstraction.

Instead of being primarily a producer of proofs, the mathematician becomes a cartographer of significance: identifying worthwhile territories, recognizing deep connections, choosing meaningful questions, and organizing an overwhelming landscape of machine-generated mathematics into something humans can understand.

In an age when answers are cheap and difficult problems can be manufactured automatically, discernment becomes scarce.

Poincaré’s principle therefore acquires new force:

“Invention is discernment, choice.” [1]

The future of mathematics may belong not to whoever can generate the most truths, but to whoever can recognize which truths are worth knowing.

If everyone practices this principle, stops obsessing over longstanding conjectures, and chooses problems they truly value rather than being swayed by AI companies, we will have the collective power to shape the future of mathematics.

References

[1] Henri Poincaré, “Mathematical Creation,” translated from the French by George Bruce Halsted, The Monist, Vol. 20, No. 3, July 1910, pp. 321–335; especially p. 325. https://henripoincarepapers.univ-nantes.fr/chp/hp-pdf/hp1910moa.pdf

[2] Liu Cixin, “The Poetry Cloud” (《诗云》), translated by Chi-yin Ip and Cheuk Wong. https://www.cuhk.edu.hk/rct/pdf/e_outputs/b7778/v7778p087.pdf

[3] https://proofsandprompts.com/2026/10/19/recommendations-on-the-use-of-ai-for-phd-students-and-supervisors

[4] https://proofsandprompts.com/2026/10/30/lets-stop-writing-papers

[5] https://proofsandprompts.com/2026/10/21/mathematics-of-the-future


Received 10 November 2026.

Filed under
More by →

6 responses to “Why the Future of Mathematics Belongs to Human Judgment”

  1. . Avatar
    .

    Agree, this could be a possible future, and not a bad one. After all, our capability of understanding and learning does not scale up with AI, even if we have this assistant who can answer all questions, we do not learn faster, mostly slower as learning happens through friction and not through a convenient chatbot. Thus, for clusters and expert networks of humans who are able to understand and teach to be maintained, a great deal of selection has to occur.

    One example that could apply at the current stage (but maybe not anymore like this tomorrow) is, that instead of using AI to generate a random paper, one should focus on questions that fits into one’s natural program, and use for one research output at least the same amount of time and energy as one did pre-ai. This could mean that AI suggests a strategy, but one does not have to stop there, one can try to connect it in a meaningful way, and likely based on own creativity, to something one self (and other colleagues) are more familiar with, one could work out a more complete picture, one could try with creative own inputs (that may require again long own pondering) to push through the wall AI seems to hit at that point. Those who do that likely are recognizable within the field, and they are likely those who are relied upon when expert judgement is needed.

  2. Scott Guthery Avatar
    Scott Guthery

    This is what was historically called Recreational Mathematics, wherein every mathematician had their own personal field of mathematics. There were two now-defunct journals that published papers from these fields: Recreational Mathematics Journal (1961-1964) and Journal of Recreational Mathematics (1968-2014). There were also regular columns in Scripta Mathematica and the MAA’s Mathematics Magazine. Finally, arXive is filled with papers from these fields. The line between recreational mathematics and many existing and abandoned scholarly fields of mathematical specialization is, by no means, bright. It’s all great fun but essentially a hobby that, as you point out, is probably not threatened by AI.

    Suppose, however, that AI made up its own field of mathematics and that the field was found useful by applied mathematicians. Now, that would be interesting.

  3. Una Avatar
    Una

    This is truly encouraging.

    I want to be a judge instead of an understander, since no one would pay me to understand.

    However, understanding is still important for being a good judge.

    I think we should combine Poincaré’s “choice” and Thurston’s “understanding”. That will be a meaningful future.

  4. Soda Avatar
    Soda

    Who is Liu Cixin by the way?

  5. Sirawit Pongnakintr Avatar
    Sirawit Pongnakintr

    I don’t really understand why one would think that AI cannot automate mathematical taste. Like… I think I can even code an algorithm to evaluate the beauty of a picture (perhaps just a small 1M-param CNN with MSELoss and Adam optimizer trained on data of paintings and scores coming from human votes), and I don’t think it’s a hard task to do so. To me, the only surviving argument isn’t that machines cannot judge “whether this theorem matters or not”–I bet it can, sooner or later. The surviving argument is that “whether such (platonically) valuable theory is valuable to [me] or not”, and this [me] must be a human. The argument that

    The future of mathematics may belong not to whoever can generate the most truths, but to whoever can recognize which truths are worth knowing.

    is self-defeating in that scenario, as long as “which truths are worth knowing” is a mean to some theorem-production end.

    The only way to make this argument survive is to accept that “perhaps the goal of mathematics is not to create a best future development of mathematics, but is to satisfy a human interest”. And if you’re willing to bite that bullet, then my opinion is in favor of this.

    In particular, I advocate against all kind of material judgements of mathematics (including money, fame, and glory). I want mathematics to remain a safe place for whoever who wants to know truth, understand why, and develop their own taste. Such human taste will never be replaceable. And such human attention is likely to become scarcer as time goes.

    1. Una Avatar
      Una

      The author didn’t say AI can’t automate taste.

      The key is humans can reject AI’s proposal.

Comments are moderated. Read our comment policy.

Add to the discussion

New posts by email.

Prefer a feed reader? Subscribe by RSS.

Also on Mathstodon.

Latest comments across the site.