How AI’s Brute-Force Math Breakthrough Threatens to Erase the Soul of Pure Mathematics
From Suno’s uncanny, eerily lifelike generative elevator music to Tilly Norwood’s sharp, customer-service-themed one-liners, artificial intelligence has already sent shockwaves through every corner of the creative industry. But now, following OpenAI’s bombshell announcement that it had solved the decades-old Navier-Stokes existence and smoothness problem using thousands of coordinated AI agents, we can add the field of pure mathematics to the list of sectors upended by the tech.
“The artists and musicians already went through this growing pain a while back,” notes Juspreet Singh Sandhu, a mathematician at Colorado State University. Contrary to the framing of high school timed exams that reward fast, correct final answers, professional mathematicians do not prioritize speed at all costs. In fact, for a field stereotyped as rigid, practical and formula-bound, the process of developing new mathematical frameworks has far more in common with artistic exploration than most people realize—think of it less like solving a worksheet, and more like inventing a brand new game or puzzle from scratch. Mathematicians work slowly, deliberately, step by step, to unpack and refine their ideas. OpenAI’s brute-force shortcut to a proof skips almost all of that intentional process, in a move that risks eroding the core of human mathematical understanding.
Math as a craft of patterns, not just answers
The 20th century British mathematician G.H. Hardy summed up this artistic side of the field perfectly in his iconic 1940 essay A Mathematician’s Apology, writing that “A mathematician, like a painter or poet, is a maker of patterns.” Where a painter builds their work from shapes and pigment, and a poet weaves patterns out of words, a mathematician constructs elegant, consistent patterns out of abstract ideas. A committed pacifist writing in the middle of World War II, Hardy wrote the essay to argue that mathematics should be pursued for its own inherent beauty, entirely separate from real-world applications—especially use cases tied to wartime military work. He celebrated what he called “useless” mathematics, holding up Carl Friedrich Gauss’ famous praise of number theory (the study of integers) as the highest, most perfect example of beautiful, purposeless pure math.
Hardy’s core comparison of math to art has held up perfectly, but he was proven wrong on one critical point: the “useless” number theory he admired for centuries turned out to be the backbone of the modern encryption protocols that secure everything from your personal email to your bank account today. For now, though, OpenAI’s Navier-Stokes proof falls firmly in that original “useless” category. The problem it solves has been nothing more than a fascinating intellectual obsession for mathematicians for decades, named for the 19th-century Navier-Stokes equations that were originally built to model the flow of viscous fluids. Engineers rely on those exact equations to calculate airflow for commercial aircraft design—but for pure mathematicians, the appeal was never the practical use case.
“Mathematicians were never drawn to these equations for engineering applications, that much is certain,” explains Jared Speck, a mathematician at Vanderbilt University. They chased a solution to the Navier-Stokes puzzle for the inherent intellectual richness, for the joy of cracking a long-unbroken puzzle. Solving it will not help anyone design a more aerodynamic airplane wing. To use a silly but effective analogy: plenty of cakes are roughly cylindrical, but studying the mathematical equations that define the geometry of a cylinder will never teach you how to bake a better sponge.
Long unsolved problems like this accrue a sort of mythos within the math community. The allure of the Navier-Stokes problem is identical to the draw of Sudoku or competitive chess: they offer no tangible real-world benefit, but deliver endless intellectual stimulation and joy. The equations are only an approximate description of how fluids flow in the real physical world, but mathematicians got curious about how they would behave in extreme, almost science-fiction edge cases, just for the fun of the thought experiment. The core puzzle boiled down to one simple question: do the Navier-Stokes equations allow for a scenario where a fictional fluid could spontaneously explode for no physical reason at all?
Mathematicians are well aware that approximate, imperfect models like Navier-Stokes can produce these nonsensical, impossible results in extreme edge cases—and that’s exactly what makes them so exciting. These strange, counterintuitive outputs often spark entirely new, never-before-seen mathematical frameworks. For decades, the global math community had been building a deep, intricate, beautiful body of work around the equations, according to Speck, and were very close to cracking the puzzle on their own when OpenAI announced that its AI agents had already proven that yes, those fictional sci-fi fluid explosions are mathematically possible.
The gap between a correct answer and actual understanding
While applied mathematicians often work on results-driven projects, from designing new encryption protocols to building optimization algorithms for robotic systems, the broader pure math community does not care only about getting a yes or no answer to a problem. The work of unpacking, contextualizing and fully understanding that final result is just as, if not more, important. When human mathematicians write proofs, they build their arguments directly on decades of prior published research, citing relevant work every step of the way. They spend months, even years, refining their writing and walking colleagues through their logic to convince their peers of the validity of their new insights. It is the equivalent of workshopping a poem with other writers, but built on logic and rigor rather than emotional resonance.
None of that collaborative, iterative process happened with OpenAI’s Navier-Stokes proof. Large language models cannot reliably cite their sources, and they are notoriously bad at explaining their internal logic in terms human researchers can follow. The 166-page proof is still under peer review, and field experts have not had the time to fully confirm if it is even correct. Mathematicians across the field have publicly criticized OpenAI for its lack of transparency around how the AI arrived at the final solution: New York University mathematician Tristan Buckmaster has noted that the model appears to draw heavily on his own published work and that of other colleagues without proper attribution.
“We don’t even know how autonomous the process was, or how much hidden human expertise was required to scaffold the AI to get to this end result,” Sandhu points out. As a result, Speck says, “almost nobody in my field fully understands what this proof actually means. The whole process was backwards: the computer handed us the final result first, and now we’re all scrambling to go back and figure out what the underlying logic even is. We’re still at the very earliest stage of that work.”
Sandhu is one of hundreds of signatories of a public open letter titled A Severe Misalignment of AI in Mathematics, which warns that mass-producing AI-generated proofs risks “destroying fertile new ground rather than nurturing the growth of new mathematical ideas.” The letter’s first 25 signatories were all winners of the Fields Medal, widely considered the equivalent of a Nobel Prize for mathematics. Crucially, the declaration is not anti-AI at all: Sandhu notes that he, like the vast majority of working mathematicians, already uses AI as a regular tool in his day-to-day work. The letter simply states that mathematical culture is built first and foremost on shared understanding, and that the breakneck speed of LLM-generated proofs risks a future where we solve thousands of famous math problems without actually understanding any of the reasoning behind them.
In response to the widespread pushback from the math community, OpenAI announced this week that it would form a dedicated advisory board made up of working mathematicians to oversee the company’s use of AI for mathematical research. A company spokesperson wrote in an emailed statement that “the criticisms laid out in the open letter make clear that thoughtful, collaborative engagement between AI developers and the global math community is non-negotiable. We are committed to ensuring working mathematicians have a meaningful seat at the table as we shape how these tools are used.”
The hidden cost of outsourcing mathematical struggle to AI
Even if AI developers collaborate in good faith with the math community, the raw speed of large language models poses a quiet but serious threat to human mathematical creativity. While LLMs are still prone to frequent errors, they can crank through pages of long, tedious, unglamorous calculations that most human mathematicians find deeply unpleasant. But Lorenzo Gavassino, a theoretical physicist at the University of Cambridge, warns that outsourcing all of that hard, frustrating work to AI would mean missing out on the most valuable mathematical insights.
“Often, a calculation that turns out far harder than anyone expected, the kind that leaves you feeling stuck and discouraged for months, is exactly the moment where the biggest new breakthroughs are born,” he explains. That struggle is what pushes mathematicians to invent entirely new concepts that would never have existed otherwise. Gavassino cites the invention of the imaginary number i, the square root of -1, as a perfect example. Mathematicians spent generations banging their heads against the wall trying to solve cubic equations, before they eventually invented the number i to make the math work. That small new concept spawned an entire new subfield of mathematics called complex analysis. Centuries later, physicists used those long-developed theorems about imaginary numbers to map out the rules of quantum mechanics, forming the foundational math behind our modern understanding of electrons, atoms, and quantum computers. Just like the imaginary number, almost all transformative mathematical concepts start out as “useless” purely abstract ideas, and their real-world practical applications only unfold over generations. Mathematicians first spend years unpacking every layer of logic behind a new concept, building a full shared understanding of how it works, long before anyone finds real-world use cases the original inventors never could have imagined.
Gavassino shares a popular saying among mathematicians: “Good mathematicians prove theorems. Great mathematicians propose conjectures. The very greatest mathematicians invent entirely new definitions.” He frames the idea with a simple analogy: the good mathematician is good at winning a pre-existing game, the great mathematician comes up with clever strategies to win that game, but the absolute best invent the entire game from scratch. When it comes to the Navier-Stokes proof and other AI-generated mathematical results, LLMs have only demonstrated that they are good at winning existing games—they are nowhere near close to being able to invent entirely new ones.
Gavassino says he is not worried that LLMs will outperform human mathematicians at their own work. He sees the models as just another regular tool, one that currently has roughly the same level of reliability as “a drunk guy wandering home after a night out,” and that responsible, creative use of the tools is entirely possible. What he does fear is that bad-faith actors and policymakers will overinflate claims about AI’s mathematical capabilities to slash public funding for basic research. “I’m afraid people will start saying, ‘Why waste money paying mathematicians to do research? We can just let AI make all the new discoveries instead,’” he says. It is worth noting that OpenAI spent millions of dollars on computational power alone to generate its Navier-Stokes proof—enough money to pay the full salaries of dozens of academic mathematicians, or fund more than a dozen graduate students for years. That funding would have gone directly to supporting new, creative breakthroughs far beyond a single proof.
For centuries, the global math community has relied on a traditional teacher-apprentice model to sustainably pass down knowledge and nurture new generations of researchers. “When I was a young early-career researcher, my mentors gave me small, manageable problems that they could have easily solved themselves in a fraction of the time,” Speck recalls. “But they didn’t solve them for me. They invested in me, they gave me the space to struggle and grow and learn how to think like a mathematician.”
Right now, AI is positioned to pick all of the lowest-hanging fruit in pure mathematics, solving all the famous, long-standing puzzles that are traditionally used to train new students. That risks cutting off early-career mathematicians from the opportunity to work through those struggles, build their skills, and develop the creativity needed to invent entirely new fields of math. Over time, that could shut off the pipeline of human creativity that built the entire field of mathematics—and that eventually gave us the large language models themselves. Every part of modern AI infrastructure is built on centuries of that so-called “useless” pure math: differential equations and complex analysis made the lasers that etch the semiconductors powering AI chips possible, and neural network architecture runs entirely on the rules of linear algebra, a mathematical framework first formalized in the 17th century. Generations of human mathematical insight built the foundation for the trillion-dollar AI industry. Now, the very companies that profited from that centuries-long public work are at risk of trampling the math community that created their entire product, in a rush to chase the next quick breakthrough and hit their lofty valuation targets.
