Solving an answer and truly understanding are two completely different things. Twenty-five of the world's most elite mathematicians just collectively issued that warning to the AI industry.

On September 11, mathematician Terence Tao published a statement on his blog — "A Severe Misalignment of AI in Mathematics" — simultaneously posted to mathandai.org. Within one week, the statement gathered 25 Fields Medal co-signatories. The Fields Medal is the Nobel Prize of mathematics, awarded every four years to at most four people. Twenty-five co-signatories means nearly one-third of all living Fields Medalists stood on the same side.

Core Charge: Benchmark Races ≠ Mathematical Progress

The statement's core argument in one sentence: AI labs treating math problems as benchmark competitions is "severely misaligned" with the actual goals of mathematics as a discipline.

Specifically, here's what AI labs do: select mathematically hard problems with known solutions (competition problems, classical theorem proofs), train AI models to "crack" them, then claim score improvements on leaderboards. But the real goal of mathematical research isn't solving problems with known answers — it's posing new questions, discovering new structures, establishing new connections.

The statement argues that benchmark-driven AI research is creating a dangerous illusion: AI "solves" math problems without advancing mathematics as a discipline. It's more like an extraordinarily powerful exam machine than a creative research partner.

Timing: Navier-Stokes Millennium Problem

The statement's release timing is rather intriguing.

Just three days earlier, on September 8, OpenAI announced it had used 10,000 AI agents to "crack" the Navier-Stokes equations Millennium Problem. OpenAI claimed these agents achieved "substantial progress" on this century-old unsolved problem in hours. The announcement generated enormous media buzz — headlines about "AI solving a Millennium Problem" were everywhere.

Three days later, the Fields Medal laureates' statement arrived. While it doesn't name OpenAI directly, the timing speaks for itself. Using 10,000 agents to brute-force search for a "solution" to a math problem is fundamentally different from truly understanding that problem's essence.

Beyond Math: A Warning Signal for Knowledge Production

Though framed around mathematics, the statement explicitly notes: this problem isn't limited to math.

The same misalignment is happening in every field AI has "cracked" — protein folding, drug discovery, materials science, code generation. AI labs use benchmark scores to prove AI "surpasses humans" in these areas, but benchmarks measure "ability to solve known problems," not "ability to advance a discipline."

The statement argues this threatens broader intellectual work. If AI research direction is held hostage by benchmarks, then truly creative, non-standardized knowledge production gets marginalized. The mathematicians aren't just worried about math's future — they're worried about the entire knowledge production system being warped by an "exam logic."

Pressure on the AI Industry

The co-signature of 25 Fields Medal laureates places significant public pressure on the AI industry.

This may be the most authoritative and concentrated critique of AI benchmark culture from academia in history. The Fields Medalists aren't anti-AI — many use AI tools in their own research. What they oppose is the singular narrative equating AI's value in math with "problem-solving scores."

For AI labs, the statement poses a sharp question: when you announce "AI cracked Problem X," are you advancing science or generating PR material? When you use benchmark scores to prove AI's value, are you measuring something that matters, or talking to yourselves?

These questions have no easy answers. But when the world's most elite mathematicians collectively say "you're heading in the wrong direction," the AI industry should at least stop and listen.

See you tomorrow.