AI Says It Cracked a 50-Year Math Problem. Mathematicians Aren't Celebrating Yet

You’ve probably noticed the headlines piling up: AI solves another math problem. This one lands harder. GPT-5.6 Sol Ultra has reportedly cracked the Cycle Double Cover Conjecture, a question that has resisted every attempt for nearly half a century. Yet the mood among mathematicians is strange. The first reaction isn’t applause. It’s “hold on a second.” Let’s unpack why.

Full disclosure up front: this story hasn’t been through the wringer yet. There’s no deep community vetting, no pile of peer commentary to lean on. So instead of pretending to confirm “AI proved it,” this piece focuses on something more useful — how to think about a claim like this when it lands in your feed.

What the Cycle Double Cover Conjecture Actually Asks

The name is a mouthful, so let’s strip it down.

Picture a set of dots connected by lines. Mathematicians call that a graph. Think of a subway map: stations are the dots, the lines between them are the edges. The Cycle Double Cover Conjecture asks a deceptively simple question. For any graph meeting some reasonable conditions, can you always find a collection of loops that traces every single edge exactly twice?

Sounds almost trivial. It isn’t. Since it was first posed in the 1970s, nobody has proven it in full. Plenty of special cases have fallen. But “for every graph” stayed stubbornly open — which is exactly why it sits on the short list of famous unsolved problems in graph theory.

Why an AI Proof Sparks a Fight

It comes down to one word: verifiability.

In math, the process matters more than the answer. It’s not enough to say “the result is correct.” Someone else has to be able to follow the reasoning, line by line, and see why it’s correct. AI-generated proofs tend to stumble here in two ways.

First, they’re often enormous. When a chain of logic runs hundreds of pages, checking it by hand becomes wildly impractical. Second, there’s the classic AI failure mode: the plausible-looking error. The prose is smooth, the argument seems to flow, and then somewhere in the middle a logical leap sneaks in unannounced. Catching that requires an actual human mathematician to sit down and grind through it.

We’ve been here before. There have been cases where an AI proof looked airtight at first, only for someone to later discover a step that quietly assumed the very thing it was supposed to prove. That history is exactly why mathematicians won’t pop the champagne this time either.

The Key to Verification: Formal Proofs

So is there a way out? Yes. It’s called a formal proof.

You may have heard of tools like Lean or Coq. They let you rewrite every step of a proof in a language a computer can check mechanically. No room for “this part’s obvious, let’s move on.” If a single step has a hole in its logic, the machine simply rejects it — no negotiation.

If GPT-5.6 Sol Ultra’s proof passes in a system like Lean, the whole conversation changes. It stops being “an AI claimed it” and becomes “a machine verified it.” But if the proof is just natural-language prose that hasn’t been formalized, then for now it’s a candidate, not a settled result. That distinction is the real fork in the road here.

So, Can We Trust It?

The honest answer right now is: wait and see.

An AI pointing toward a crack in a problem humans couldn’t solve is genuinely remarkable. AI is a powerful tool for scanning vast spaces of possibilities and surfacing approaches people overlooked. But “surfacing a lead” and “completing a proof” live on different floors of the same building. For the community to formally accept this, it has to clear formal verification or survive independent scrutiny from multiple experts. There’s no shortcut.

Here’s where it nets out. An AI knocking on the door of a 50-year-old problem is a real event. Whether that door actually opened is a call for cold, unforgiving verification to make — not the AI, and not the hype. Which raises the question worth sitting with: will we ever trust an AI’s proof the way we trust a human’s? Or are we drifting toward a new standard, where the only proofs we believe are the ones a machine has checked?