AI Disproved an 80-Year-Old Math Conjecture

Seven months ago, OpenAI announced that GPT-5 had solved ten unsolved Erdős problems. It hadn't. The solutions already existed in published literature. The post came down, the rival taunts arrived, and OpenAI absorbed a public credibility hit that was entirely self-inflicted.

This week, they tried again. And this time, the mathematicians are nodding.

What Actually Happened

OpenAI's new general-purpose reasoning model produced an original proof disproving the Erdős unit distance conjecture — a geometry problem first posed by Paul Erdős in 1946 that has sat unsolved for nearly 80 years. The conjecture concerned how many times a single distance can appear among a set of points in a plane. For decades, mathematicians assumed the best configurations looked roughly like square grids. The model found an entirely new family of constructions that performs better, disproving that assumption.

Critically, OpenAI didn't just publish the claim. They published companion remarks from working mathematicians including Noga Alon, Melanie Wood, and Thomas Bloom — the same Thomas Bloom who runs the Erdős Problems website and previously described OpenAI's last announcement as "a dramatic misrepresentation." His presence in the supporting materials is not incidental. It's a direct response to the credibility problem the company created for itself last October.

Why This Is Different From Benchmark Wins

AI systems beating humans on standardized tests, coding challenges, and reasoning benchmarks has become almost routine news. This is different in a specific way.

The Erdős conjecture wasn't a test with a known answer. It was an open problem — genuinely unsolved, actively worked on by human mathematicians, with no existing solution hiding in the literature for a model to surface and misattribute. Producing an original disproof requires not just retrieval or pattern-matching but the construction of something new: a chain of reasoning that didn't exist before the model produced it.

OpenAI's characterization is that this demonstrates AI systems are now capable of holding together long, complex reasoning chains and connecting ideas across fields in ways researchers hadn't previously explored. That's a meaningful distinction from "AI scored well on a math exam."

The Implications Are Larger Than Geometry

The unit distance conjecture is, on its own, a problem most people will never encounter. But the category of capability it demonstrates has implications well beyond discrete geometry.

If a general-purpose reasoning model — not a system purpose-built for mathematics — can produce original proofs in pure math, the same class of capability applies to any domain structured around long chains of dependent reasoning. Drug interaction modeling. Materials science. Climate system simulation. Protein folding variants. The problems that have resisted solution not because we lacked data but because the reasoning required to connect that data was too complex and too long for human attention spans to sustain.

That's the actual story underneath the geometry result. The conjecture is the proof of concept. The implication is a general-purpose reasoning system that can operate usefully at the frontier of hard problems across disciplines.

The Credibility Question Isn't Fully Resolved

OpenAI's track record on announcements like this earns some skepticism, and the mathematical community will subject this proof to scrutiny that takes longer than a news cycle. The presence of credible mathematicians in the companion remarks is encouraging, not conclusive.

There's also a broader pattern worth naming: AI labs have strong incentives to announce breakthroughs and weaker incentives to announce the corrections. Last October's mistake didn't cost OpenAI much beyond some embarrassing coverage. The asymmetry between the reach of the original claim and the reach of the retraction is a structural problem in how AI progress gets reported and understood.

What's different here is that the verification was built into the announcement rather than retrofitted after criticism. That's a process improvement, even if the underlying incentive problem hasn't changed.

What This Means for Anyone Building With AI

For marketers and business leaders, the practical takeaway isn't about geometry. It's about what class of problems AI reasoning systems are now plausibly useful for. Complex, multi-step analytical work — the kind that requires holding many variables in relationship across a long chain of logic — is moving into scope. That has real implications for how you think about research, strategy, forecasting, and any domain where the bottleneck has been reasoning capacity rather than data availability.

The cathedral of mathematics, as Thomas Bloom put it, has an unseen wing. So does your industry.

Build With AI That's Actually Ready

Knowing when AI capability is real versus announced is half the work. Winsome's growth team helps you cut through the noise and build programs on AI that actually performs. Let's talk.