AI For Math Proofs: Organized Monkeys Or Superhuman Genius?
Ambuj Kumar, Simbian.ai co-founder and CEO, pioneered confidential computing, and builds AI agents to counter AI threats.
gettyOn October 6, 2026, OpenAI posted over 700 AI-written math manuscripts. About 300 come with proofs in Lean, a language a computer can check line by line. One claims a zero-free region for the Riemann zeta function, where the real part exceeds 7/8, a partial result related to the long-standing Riemann hypothesis. Months earlier, an unreleased Claude model raised the proven share of zeta zeros on the critical line from 41.6% to 67.2%, with a Lean proof.
This doesnβt mean that mathematicians are celebrating. In September, Terence Tao and 24 other Fields Medalists signed a declaration saying that βthe goals of the AI companies and the goals of the mathematical community are severely misaligned.β In a lecture at Caltech on October 9, Tao went further: βFurther blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics.β
So, are these machines geniuses? To judge that, start with how an LLM learns math.β
βMany labs train reasoning models using reinforcement learning, rewarding models when they produce correct answers. The approach works particularly well when those answers can be objectively verified. Whether an essay is good or a security alert is real is a judgment call, and a noisy reward teaches a noisy model.
Math offers a particularly useful advantage: When a proof is formalized in Lean, its checker can determine whether the argument follows from the stated assumptions, without requiring a human to review every step. βIt should therefore be no surprise that models trained this way can now produce solutions to problems that have stumped generations of mathematicians.β
However, it also creates a problem. When proofs are checked automatically, the model only has to be right once, and you can keep it trying for as long as you can pay for those GPUs. Give enough monkeys enough typewriters, and one will eventually type Hamlet. In practice it never happens, because probability is so low we would have to wait for trillions of years.
LLMs are stochastic systems that can generate mathematical arguments and Lean proof attempts. Many attempts will fail, but a model that can produce sufficiently promising candidates can be paired with an automated checker to identify formally valid results.β How many attempts would it take to solve a famous problem? What if a lab can spend tens of millions of dollars on compute, while successful announcements strengthen perceptions of its AI capabilities and potentially improve its competitive position? Those economics justify a very large search.
Only the successful results receive individual attention. The thousand failed tries, the tokens burned and the size of the model running the search largely stay out of view.ββ
βNone of this is an accusation. Math has long been the clearest test of intelligence, and labs should test their models on it.
The trouble is what we use math to measure. We admire mathematicians like Terence Tao for their ability to develop insights and solve problems that might take others months or years to understand. When Ramanujan filled notebooks with identities that took others decades to prove, part of the wonder was how little he seemed to need.β
The labs only disclose some of the cost, but not enough. Anthropic says its zeta result took 31 million output tokens, after a first session tried 650 ideas without success. OpenAI says its model averaged three hours of ChatGPT Pro thinking per result. Neither lab has said how big its model is. OpenAI has not published its prompts or per-result compute, although an advisory group of mathematicians at the Institute for Advanced Study asked for both on September 29.β
A mathematician who proves a theorem after a month of drafts has run a 20-watt brain for that month. A hypothetical 10-trillion-parameter model that needs a hundred attempts could consume vastly more energy, depending on the hardware and compute involved. From outside, the two announcements look the same, and nothing tells us whether we are looking at a mind like Ramanujanβs or a well-organized room of monkeys.β
βI propose that every AI-generated proof ship with a disclosure card, like a security researcher publishes an advisory for every vulnerability they disclose.β
The security world has developed a useful model for this. CVE provides standardized identifiers for publicly disclosed vulnerabilities, while CVSS offers a common framework for assessing their severity. Coordinated disclosure processes also establish expectations for documenting findings, notifying affected parties and appropriately attributing researchers.β
A similar discipline can be followed here. arXiv could add a metadata field for AI-assisted proofs, journals like the Annals of Mathematics could require a card at submission and the Lean mathlib community could require disclosure in every pull request that credits an AI contributor.
Defining what counts as an attempt will be hard, though, and labs will not shy away from calling model size and per-result compute trade secrets. Also, the lab that discloses first will look worse than the one that stays silent, which is why the first standard has to come from a venue that can refuse to publish without the card.
Once cost sits next to every result, a small modelβs first-try proof outranks a giant modelβs hundredth try, and labs have a reason to build small models that find short proofs mathematicians can read and build on. Until a result comes with its card, and until these models work alongside mathematicians, I believe we should hold back the respect we reserve for people who solve problems the traditional way.βββ
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
