The following document was written by ChatGPT. The text represents the opinions and arguments of the human author.
October 2026
Some intellectual disciplines mainly study a rich world that already exists. History studies past events, geology studies the Earth, and biology traditionally studied living organisms. Other activities are mainly generative: literature, music, engineering, and much of mathematics create new objects.
Artificial intelligence raises the possibility that mathematics could move from the second category toward the first. If machines can autonomously generate enormous amounts of interesting mathematics, then mathematicians may increasingly explore, classify, and interpret a mathematical world produced partly by machines. But this scenario depends on a crucial premise that has not yet been established. Current AI can generate many correct statements and can solve or extend many difficult problems selected by humans. There is much less evidence that it can autonomously choose fruitful directions and generate large numbers of theorems that mathematicians would regard as interesting. The future of mathematics may depend on whether this gap is crossed.
There is a useful, although imperfect, distinction between disciplines that study complex systems already present in the world and disciplines that create their objects.
History begins with events that have already happened. Geology begins with the Earth. Economics studies an economy already operating. Traditional biology begins with organisms supplied by evolution. Their characteristic activities include observation, classification, reconstruction, and explanation.
Other fields are more overtly generative. Writers create literature, composers create music, engineers create machines, and mathematicians invent definitions, formulate conjectures, construct examples, prove theorems, and build theories.
Mathematics is philosophically ambiguous: one may believe mathematical objects are discovered rather than invented. But the day-to-day activity of mathematicians has historically been strongly creative. New mathematics was also scarce because producing it required substantial human effort.
It is tempting to extrapolate from current AI progress and imagine that machines will soon generate millions of new theorems. But that statement is too weak to have much significance.
One can generate infinitely many correct but uninteresting theorems. The difficulty is not producing statements that are true. The difficulty is producing mathematics that matters.
There are therefore three increasingly demanding capabilities:
generate many true theorems;
generate many nontrivial or useful theorems;
autonomously generate many theorems that mathematicians judge interesting.
The first is easy. There is now meaningful evidence for the second. The third remains largely open.
Several recent systems show that AI can produce substantial new mathematics when humans provide the surrounding problem, formal system, or evaluation criterion.
AlphaEvolve was applied to more than fifty problems in analysis, geometry, combinatorics, and number theory. According to Google DeepMind, it rediscovered state-of-the-art solutions in roughly three quarters of the cases and improved the best known result in about one fifth, including a new lower bound for the kissing-number problem in dimension eleven. [1] This is impressive evidence that AI can generate nontrivial mathematical advances quickly. But humans selected the problems and supplied evaluators that told the system which candidates were better.
Likewise, DeepMind's Aletheia research agent has solved research-level problems and was evaluated on hundreds of open problems. [2] Again, the targets came from humans.
A different line of work asks whether useful theorems can emerge with less human mathematical input. A 2026 system starting only from axioms and inference rules generated tens of thousands of formally proved theorems. Some of these were useful as lemmas for solving human-written benchmark problems. [3] This goes beyond trivial theorem generation. Yet usefulness for proof search is still not the same as mathematical significance.
The central difficulty is that mathematical interest is not easily reduced to correctness or proof difficulty.
A recent paper, Learning to Discover Interesting Mathematics, makes this problem explicit. The authors propose a quantitative proxy for interestingness based on the relation between the complexity of a theorem's proof and the size of its statement. Their system can generate candidate theorems, rank them, prove selected ones, and iteratively expand a formal library. [4] This is probably one of the closest current examples to autonomous mathematical exploration.
But the proposed notion of interestingness is necessarily incomplete. A short statement with a difficult proof can be artificial. A theorem with an easy proof can be conceptually important if it introduces the right point of view.
Another recent study of automated conjecturing illustrates the problem. Rewarding a system for producing difficult conjectures caused it to drift toward long artificial statements, and no single selection rule worked best in every tested domain. [5]
This suggests that the main obstacle to abundant autonomous mathematics may not be theorem proving. It may be selection: deciding what is worth proving in the first place.
If AI eventually crosses this threshold, then the character of mathematics could change substantially.
Suppose autonomous systems generate not merely correct statements but enormous families of genuinely interesting results, examples, definitions, and theories. No individual mathematician could read more than a tiny fraction of them.
The mathematician's role would then begin to resemble that of a naturalist.
A naturalist encounters an abundant world and asks which specimens belong together, which distinctions are fundamental, which exceptional cases matter, and what organizing principles explain the observed diversity. A mathematician facing millions of machine-generated results might ask similar questions:
The theorem would become, in part, a specimen.
Human interaction with mathematics could acquire an empirical-looking layer:
observe patterns → classify phenomena → formulate organizing principles → prove them.
Proof would still determine mathematical truth. But the human task might shift from manufacturing local results toward exploring and organizing a mathematical environment too large to have been created or understood by any individual person.
The contrast with biology would then be striking.
Biology historically studied organisms supplied by nature. Modern synthetic biology increasingly aims to design genomes, proteins, cells, and biological systems with desired properties. Biology is moving from understanding what exists toward designing what could exist. [6]
Mathematics could move in the opposite direction. Instead of mathematicians primarily creating the objects they study, machines could increasingly populate the mathematical world and humans could increasingly analyze it.
The reversal can be stated succinctly:
Biology is becoming more synthetic, while mathematics may become more observational and curatorial.
Chemistry may be an even better intermediate model. Chemists both study existing compounds and generate large libraries of new ones. Once synthesis became cheap, screening and selection became central. Mathematics may eventually face the same problem: theorem generation may become inexpensive while deciding which theorems matter remains difficult.
Even in this scenario, mathematics would not cease to be creative. Creativity would move upward one level.
The important questions might become less often
How can I prove this theorem?
and more often
What is the right concept? Which region of mathematical space should be explored? Which machine-generated results reveal one phenomenon? What definition turns thousands of facts into one idea?
There would also be a feedback loop:
human concepts → machine exploration → human interpretation → new concepts.
For as long as humans remain better at choosing fruitful concepts and judging significance, mathematics would retain a strongly creative human component. If AI also becomes better at those tasks, the transformation would be much deeper.
The important question is therefore not whether AI can generate many theorems. It plainly can generate many statements, and increasingly it can generate difficult and useful ones.
The real question is:
Can AI autonomously generate large quantities of mathematics that humans, after the fact, recognize as important or interesting?
At present, there are early indications in restricted formal settings, but no convincing demonstration at the scale required for the natural-history picture.
If the answer eventually becomes yes, mathematics may undergo an unusual reversal. Science traditionally studied a world that nature created, while mathematics largely created the world mathematicians studied. Machines would begin creating a substantial part of that mathematical world, leaving humans increasingly to explore, interpret, and organize it.
If the answer remains no, then AI may instead become an extraordinarily powerful mathematical tool while humans continue to supply the questions, definitions, and directions that make the resulting mathematics worth doing.
That distinction is the hinge on which the analogy turns.