On May 20, 2026, a quiet revolution happened inside a data center.
OpenAI announced that one of its internal AI models had solved a cluster of mathematical problems — not trivial ones, but the kind that have been labeled "impossible" since the 1960s. The problems came from the legendary Hungarian mathematician Paul Erdős, who spent his life offering cash bounties for solutions to the hardest questions in combinatorics, number theory, and graph theory.
Most mathematicians didn't celebrate. They were confused. And then they became afraid.
The Problems No One Could Solve
Paul Erdős was unusual even among geniuses. He didn't have an office. He didn't publish papers in journals. He traveled from university to university, showing up unannounced, handing colleagues problems he'd written on scraps of paper, and offering money for solutions. Some bounties reached $10,000.
Most of his problems went unanswered for decades.
"OpenAI's model didn't just solve an Erdős problem," said a mathematician at MIT who requested anonymity. "It solved three of them. And the proofs it generated — well, they're not proofs in the way we understand them."
That's the word that keeps mathematicians awake at night: understand.
Traditional mathematics operates on a simple principle. A proof is valid if and only if other mathematicians can read it, follow it step by step, and agree it's correct. This human verification process is what gives mathematics its authority. It's what makes a theorem true rather than just plausible.
But OpenAI's model didn't generate proofs that human mathematicians could read.
What the Model Actually Did
The AI system didn't approach the Erdős problems the way a human would. It didn't use intuition or insight or the creative leaps that have defined mathematical discovery for millennia.
Instead, it found patterns in data that no human had ever noticed.
In one case, the model identified a previously unknown relationship between combinatorial structures and algebraic properties — a connection that would have taken a human mathematicians years, perhaps decades, to stumble upon. The result was correct. Other mathematicians have verified it independently. But no one can explain how the model found it.
"We're not just talking about a faster way to prove theorems," said one researcher. "We're talking about a fundamentally different way of doing mathematics."
The Implications Are Terrifying
This isn't just about Erdős problems. If AI can solve problems that humans can't, it means there are truths in mathematics that exist beyond human comprehension.
That's a profound philosophical crisis.
For thousands of years, mathematics has been the most reliable form of knowledge we have. We trust the Pythagorean theorem not because a god told us so, but because we can understand the proof. We trust Euler's identity because we can follow the steps from beginning to end.
What happens when the proofs exist, but only machines can produce them?
Some mathematicians argue that formal verification — using computer systems to mechanically check every step of a proof — could solve the problem. But this approach has trade-offs. Formal verification is rigorous, but it's also incredibly slow. And it doesn't give us the insight that makes mathematics valuable in the first place.
What Comes Next?
The mathematics world is at a crossroads.
Some mathematicians are embracing AI as a tool for exploration. They're using these systems to generate conjectures, to suggest new areas of research, to find patterns that human intuition can't see. Others are skeptical. They argue that a proof without understanding isn't a proof at all.
The most productive path forward seems to be a hybrid approach. AI generates insights and conjectures that humans then explore, refine, and understand. This preserves human agency while leveraging the unique pattern-recognition abilities of machine learning systems.
It won't be easy. There will be disagreements about which results are trustworthy, which AI methods are legitimate, and how much human understanding is required for mathematical knowledge to be valuable.
But one thing is clear: AI has changed mathematics forever.
The question now is how we'll adapt to that change. Not whether it will happen.
The Erdős Legacy Lives On
Paul Erdős died in 1996, never knowing that his life's work would one day be solved by machines. But his legacy is more important than ever.
His problems weren't just hard. They were designed to push the boundaries of what mathematicians thought possible. That's what made them worthy of bounties. That's what made them legendary.
And now, those boundaries have been pushed again — not by human genius, but by artificial intelligence.
The crisis in mathematics isn't the end of mathematics. It's the beginning of something new. And we're not ready for it.