A Discovery by Accident?

How an AI Cracked an 80-Year-Old Mathematical Problem, and What That Really Means

An opinion piece. The facts in this essay are not in dispute: the proof is real, it has been checked by nine mathematicians, and it overturns an eighty-year-old conjecture. What is disputed is what the achievement means. What follows is one reading: the result is better understood as a striking product of truth-blind generation than as evidence of machine mathematical understanding. Interesting, even historic; not the revolution it is sometimes taken to be. Several of the experts who verified the proof read it more generously, and they may be right. This is the argument for the other side.

---

The Problem: A Deceptively Simple Question

In 1946, the Hungarian mathematician Paul Erdős posed a question so simple to state that it sounds almost trivial: if you scatter n points across a flat plane, what is the maximum number of pairs of those points that can sit exactly one unit of distance apart?

This is not really a question about a handful of points; it is about the behaviour for very large n: what formula links n, the number of points, to the maximum number of unit-distance pairs. The answer is best read as a growth rate, an exponent: the count always grows faster than n, but how much faster? For small n the count outruns n quite comfortably, but as n grows the rate eases, and the exponent governing it drifts slowly downward toward 1, the count stays only barely faster than linear.

Erdős supplied both a ceiling and a floor. For the ceiling, he showed the count can be no larger than roughly n^(3/2), because the unit-distance graph cannot contain certain forbidden configurations (two unit circles meet in at most two points). That bound was later improved to n^(4/3) by Spencer, Szemerédi, and Trotter, which became the standing record. For the floor, he found a strong construction, though not the obvious one. A plain integer grid, with rows and columns one unit apart, produces only about 2n unit pairs, barely more than the number of points itself. The powerful construction is a carefully scaled square grid, analyzed with tools from number theory (counting how often integers can be written as sums of two squares); it attains a count growing like n^(1 + c/log log n), for some fixed positive constant c, one plus a correction term that fades to nothing as n grows, a rate only just faster than linear. On the strength of that construction, Erdős conjectured that no arrangement could do substantially better: that the grid was not merely achievable but optimal.

For almost eight decades, this conjecture stood. The square grid remained the working benchmark. Nobody found a cleverer arrangement, and nobody managed to prove the conjecture was true. It sat in that uncomfortable mathematical limbo: almost certainly right, completely unproven, untouched.

The Difficulty: Why Nobody Solved It

The unit distance problem is not obscure for lack of attention. It sits at the intersection of discrete geometry, combinatorics, and number theory: three fields with deep and active research communities. Erdős himself, one of the most prolific mathematicians of the twentieth century, was unable to resolve it. The best upper bound ever established, a ceiling of n to the power of 4/3, proven by Spencer, Szemerédi, and Trotter, remained untouched for decades, and the gap between that upper bound and Erdős's conjectured lower bound was never closed.

The difficulty lies partly in the nature of the problem: it requires constructing explicit point configurations with provably extreme properties, while also requiring algebraic and analytic machinery to verify those properties rigorously. Attempts to use algebraic geometry, harmonic analysis, and combinatorial methods all ran into different walls. The tools existed, but nobody had found a way to combine them that broke the problem open.

On 16 April 2026, Thomas Bloom, a mathematician at the University of Manchester who maintains the definitive online database of Erdős problems, listed the unit distance problem in a blog post titled “Top 10 Erdős Problems” (a list of both solved and unsolved problems). He chose it specifically because it represented a genuinely hard, genuinely open question that had resisted intensive study. He published this list partly as a corrective: recent AI announcements had led some to assume that all Erdős problems were trivial puzzles that had simply never been tried. The unit distance problem, he wrote, was evidence to the contrary.

The Solution: What the AI Did

On May 20, 2026, OpenAI announced that an internal reasoning model had produced a counterexample to Erdős's conjecture. The model found an infinite family of point configurations in the plane, not square grids, but something more algebraically exotic, in which the number of unit-distance pairs grows like n to the power of 1 plus epsilon, where epsilon is a fixed positive number. Because epsilon is a fixed number rather than one that shrinks toward zero, this is polynomially faster than the rate Erdős had conjectured, and the conjecture was therefore wrong. Two caveats keep this in proportion. First, beating the conjecture is not the same as reaching the known ceiling: this new lower bound still sits below the proven upper bound of n to the power 4/3, so the exact growth rate remains unknown: the gap was narrowed, not closed. Second, the original AI proof showed only that some fixed epsilon greater than zero exists, without computing it; shortly afterward Will Sawin (one of the verifying mathematicians) refined the construction to make the exponent explicit, reaching about n to the power 1.014, still well below 4/3.

The approach the AI used drew on three distinct bodies of mathematical work that had never been combined in this way. The first was a technique associated with Ellenberg and Venkatesh that uses small split primes and the pigeonhole principle to control class groups of number fields, a result from analytic number theory with no obvious connection to geometry. The second was the Golod-Shafarevich theorem, which guarantees the existence of infinite towers of algebraic number fields with bounded root discriminant. The third was a method, due to Hajir, Maire, and Ramakrishna, for building such towers in which infinitely many primes split, the so-called Frobenius-cutting argument. This piece is worth flagging carefully, because it is easy to overstate: the paper is explicit that constructing such towers is not the novel part of the work, the technique had been used before for other purposes, and the AI's own proof relied only on a weaker, finite depth version of it, with the stronger infinitely-many-primes form supplied afterward by the human authors in their write-up. (Related ideas from coding theory and lattice sphere packings sit in the background as motivation, but the verifying mathematicians stress these are not the novel ingredient, and the relevant ring of integers does not even form a lattice once the field degree reaches three.)

It is worth being exact about what is and is not new here. Every one of these techniques predates the result; the model invented no new machinery and introduced no method that did not already exist. The single original contribution was the combination itself, the decision to assemble known parts in an arrangement, and at a field degree, that nobody had tried. The novelty is entirely in the synthesis, not in any of the ingredients. And what that combination produced is a disproof by construction: a single explicit family of point sets that an eighty-year-old conjecture could not survive. This refutes the conjecture rigorously: it is a verified disproof, not a heuristic or a lucky guess, but refuting a conjecture is not the same as solving the problem behind it. The unit distance problem itself remains open: we now know the answer is not what Erdős expected, yet we do not know what it is. Its true growth rate still lies unknown, somewhere in the gap between the new lower bound and the n-to-the-4/3 ceiling.

The key insight, and this is what the nine mathematicians who verified the proof identified as genuinely novel, was to let the degree of the number field grow without bound. Previous work had always fixed the field and varied other parameters. The AI's chain-of-thought reasoning, partially preserved in the verification paper, included a passage that reads: “maybe that enormous degree is not just an annoyance but a source of possible counterexamples. Number fields deserve a closer look.” That reframing unlocked the construction.

The verification paper, posted the same day on the arXiv preprint server, was authored by nine mathematicians: Noga Alon (Princeton), Thomas Bloom (Manchester), W.T. Gowers (Cambridge and Collège de France), Daniel Litt (Toronto), Will Sawin (Princeton), Arul Shankar (Toronto), Jacob Tsimerman (Toronto), Victor Wang (Academia Sinica), and Melanie Matchett Wood (Harvard). They described their document as a human-digested, somewhat simplified, and somewhat generalized version of the AI proof. Noga Alon, who had heard Erdős raise this problem personally in lectures, called the result an outstanding achievement.

The Context: A History of False Starts

This was not the first time OpenAI had announced a breakthrough on Erdős problems. In October 2025, a senior company executive posted on the social platform X that GPT-5 had solved ten previously unsolved Erdős problems and made progress on eleven others. The post was deleted within days. Thomas Bloom, the same mathematician now co-authoring the verification paper, investigated and found that the model had not produced original proofs. It had performed a literature search and surfaced solutions that already existed in published papers, simply not yet catalogued in his database. Bloom called the announcement a dramatic misrepresentation. Demis Hassabis, the CEO of Google DeepMind, called it embarrassing. The executive responsible left OpenAI in April 2026. That episode casts an important shadow over the May 2026 result, and also, paradoxically, gives it greater credibility. The fact that Bloom, having publicly and forcefully debunked the earlier claim, subsequently co-authored the verification paper and stated that he had not encountered the approach before seeing the AI's proof, carries genuine weight. He had both the expertise and the motivation to be skeptical. He was not.

The Hard Question: Did the AI Understand What It Found?

Creativity or Combinatorics?

A tempting way to describe what happened is that the AI hallucinated its way to a discovery, and that description is more precise than it first sounds. The mechanism that produced this proof is the same one that produces the model's errors: recombining fragments learned in training by statistical association, with no internal representation of whether the result is true. When that mechanism attaches the wrong date to a famous name, telling you George Washington was born on Einstein's birthday, because names and dates are densely linked in the training data and nothing in the model objects, we call it a hallucination. When it splices together three number theoretic techniques that had only ever been studied separately, and the combination happens to be valid, we call it a discovery. The operation is identical. All the ingredients were present in the training data; the model combined them in a configuration nobody had tried, and that configuration happened to be correct. In this reading the result is a statistical event whose status, triumph or error, was not fixed at the moment of generation, but settled afterward, by whether it turned out to be true. That is the sense in which luck is doing real work here: the generator cannot aim at truth, so when it lands on truth, the landing is not to its credit.

The natural objection is that the model's chain-of-thought reasoning looks like directed exploration rather than random noise: it identified number fields as a promising direction and pursued it with apparent coherence. But this settles nothing, and it is worth being clear about why. A truth-blind generator produces coherent-looking exploration for the same reason it produces coherent looking false birthdays: coherence is a property of the surface text, which these systems model extremely well, not evidence of any underlying grip on validity. The decisive fact is that the model had no way to tell, as it was writing, whether its chain of reasoning was the rare correct one or one of the many that merely read well, and that distinction was supplied entirely from outside, by the nine human verifiers. The burden of proof therefore lies with anyone claiming the model understood what it found. The null hypothesis, the ordinary hallucination mechanism, emitting one true argument among an unknown number of plausible-but-wrong ones, with correctness certified by humans after the fact, already accounts for everything we can observe, and requires no new faculty to be invented for the occasion.

It is sometimes replied that this, taking known tools from different domains and combining them in a new way, is also how human mathematical creativity works, and that Erdős himself was celebrated for exactly this kind of cross-domain synthesis. The parallel is real, but it is incomplete in the way that matters. The human mathematician who recombines ideas also carries the faculty that tells him which of his recombinations to trust, which to throw away, and when an argument has quietly gone wrong; in a person, the generating and the judging are the same mind. That second faculty is precisely what the model lacks. So the difference here is not merely one of mechanism, scale, and speed: it is the presence or absence of the very thing that converts generation into mathematics. Strip the judgment away and what is left is a generator that cannot tell its discoveries from its hallucinations, because to it they are the same kind of object.

The Missing Piece: Self-Verification

The deepest question, however, is not whether the AI produced something new. It did. The deeper question is whether the AI knew it had produced something new, and more specifically, whether it could tell the difference between this result and the countless plausible-looking outputs that are simply wrong.

The answer, based on everything we know about how these systems work, is: almost certainly not. The model did not formally verify its own proof. Nine external mathematicians did that. The model produced output; humans determined it was correct. If it had produced something subtly wrong, which it almost certainly has done on many other occasions, and may have done many times in attempts at this very problem, it would have had no reliable internal mechanism to detect the error.

This is the core distinction between what we might call genuine mathematical intelligence and what happened here. A mathematician like Erdős did not merely generate conjectures and proofs: he had calibrated intuitions about which approaches were likely to succeed, which arguments had holes, and when his own reasoning had gone wrong. That internal error-correction, that sense of mathematical taste and validity, is arguably what mathematical intelligence consists of. The AI, as currently constituted, has a very powerful generator and a very weak internal verifier. The generation, in this case, was historically significant. But the verification was entirely outsourced to humans.

How Many Attempts? We Don't Fully Know

There is another question that has received relatively little attention in the coverage of this result: how much did the model produce, and discard, around this one success? Here the record offers one concrete data point. The verification paper states that the proof file was first generated by the model in a single shot, and then refined for exposition through human interaction, so the winning argument was one generated artifact, not something stitched together over thousands of partial attempts.

This is the single hardest fact for the reading offered in this essay, and it deserves to be met head on rather than stepped around. If the proof had been one survivor culled from thousands of logged failures, the luck framing would be almost trivially true: a lottery with enough tickets eventually prints a winner. A clean one-shot generation denies us that easy story. But it does not rescue the opposite one. One shot is not one attempt: a single generated file says nothing about how many independent runs, on this problem and on the others in the same batch of evaluations, produced confident nonsense that never left the building. And even a generator that lands the proof in a single pass is still a generator that could not have known it had done so. Producing truth in one shot and recognizing truth in one shot are different capacities, and only the first is in evidence here.

What remains unknown, then, is the denominator around that one shot: how many independent attempts were run before one succeeded, whether other AI-generated “solutions” to hard problems were tested, found incorrect by human reviewers, and quietly set aside, and what the overall hit rate and false-positive rate look like. The one result that was published is the one that happened to be right; the full distribution of outputs remains largely invisible.

This matters for how we interpret the result. A generator that produces one correct proof in ten attempts is a different kind of tool than one that produces one correct proof in ten thousand. Both could yield the same published outcome. Without knowing the denominator, we cannot assess the reliability of the system, only the existence of at least one success.

What This Means, and What It Doesn't

The disproof of the Erdős unit distance conjecture is a genuine mathematical event. An 80-year old conjecture has been refuted. The answer is not what most experts expected. The point sets that maximize unit distances are not, in the limit, anything like square grids: they are exotic algebraic constructions built from towers of number fields that grow in degree without bound. That is a real result, and it will change how mathematicians think about discrete geometry. (It is worth restating that the broader problem is not fully settled: the exact growth rate still lies somewhere between the new lower bound and the n-to-the-power-4/3 upper bound.)

Whether it represents a turning point for AI in mathematics is the harder question, and here it is only fair to record that several of the experts closest to the proof disagree with the deflationary reading advanced in this essay. Gowers described the solution as a milestone in AI mathematics; Litt called it the first autonomously AI-produced result he found exciting in its own right rather than as a leading indicator. Those are not careless judgments, and a reader should weigh them against what follows. The case made here is narrower: the result demonstrates that large language models can, in at least one case, synthesize across mathematical domains in a way that produces genuine novelty, and, in the same breath, that human verification remains not just useful but essential, because the model could not confirm its own success, and without the nine mathematicians who checked the work, the output would have been indistinguishable from the many plausible-but-wrong results these systems routinely produce.

It is worth adding that the verifiers are not of one mind, and the most useful dissent comes from the most skeptical among them. Thomas Bloom, who a year earlier had publicly debunked a false AI claim on these same problems, gave a deliberately measured verdict. Has the result taught us something, does it deepen our understanding of discrete geometry? A qualified yes, he allowed: it shows that number-theoretic constructions have far more to say about such questions than anyone had suspected, and he expects algebraic number theorists to begin eyeing other open problems in the area. But he added a pointed caveat, that the work introduces no powerful new geometric tools, no hitherto unsuspected structural results, of the kind a genuine proof of the conjecture would have demanded, drawing the contrast with Guth and Katz, whose attack on the related distance problem did introduce exactly such tools. That assessment, from the resident skeptic who checked the proof, is the most credible support the deflationary reading has.

There is a further question of value, distinct from significance. A generator whose every output must be certified by a handful of the world's strongest mathematicians is limited not by what it can produce but by what can be checked, and expert attention is the scarcest resource in the field. Verifying a proof is usually far cheaper than discovering one, which cuts in the tool's favour; but if its wrong outputs are both frequent and subtle, the verification burden could swallow the gains. We do not yet know which world we are in.

From an industrial rather than a mathematical vantage the picture dims further. To use such a system you must first state your problem with the precision of a theorem, and for the unit distance problem that work had already been done. Erdős framed the question eighty years ago; deciding what to ask, and sharpening it into a precise conjecture, was itself an act of mathematical judgment, one the community went on to refine for decades. The model did not pose the question, judge it tractable, or formulate it. It overturned Erdős's answer while depending entirely on Erdős's question; it was pointed at a clean, famous, human-curated target. You must then run the generator some unknown number of times and hope one output is both correct and recognisably so, and finally spend scarce expertise to find it. Where a problem is already crisply posed and cheaply checkable, this can pay off, because generation was the historically hard step. Where stating the problem exactly is itself the hard part and verification is costly, which describes most real-world work, a tool that yields one usable answer per large-and-unknown number of attempts, each needing expert adjudication, may not be worth the trouble. And the denominator is doubly hidden: we do not know how many attempts this one problem took, nor how many famous problems were tried for each that produced something true.

The most honest summary may be this: something historically significant happened, and we do not yet have the right conceptual vocabulary to describe exactly what it was. It was not hallucination in the pejorative sense, the output was correct, but it was hallucination in the mechanical sense: the same truth-blind recombination that, on the model's bad days, invents a citation or a birthday, this time landing on something real. It was not intelligence in the full sense: the system could not evaluate its own output, and so could not tell this triumph apart from its routine errors. It was something in between: a very large, very fast pattern-matcher that, on one remarkable occasion, assembled known pieces into an unknown whole, and got lucky enough, or was good enough, for the assembly to hold. On the evidence we have, the honest reading leans toward lucky: nothing we can see required the assembly to be more than a recombination that happened to be true.

What happens when you close the loop, when the generation is paired with reliable self verification, remains to be seen. That is probably the most important open question in AI-assisted mathematics. And unlike the unit distance problem, it does not yet have an answer. Until that loop is closed, the burden of the word “revolutionary” belongs to those who would apply it, not to those who, looking at the same evidence, see something rarer and stranger than a revolution: a machine that found a truth it had no way of knowing it had found.

Note: The verification paper, “Remarks on the disproof of the unit distance conjecture,” is available at arXiv:2605.20695. Will Sawin's companion paper making the exponent explicit, “An explicit lower bound for the unit distance problem,” is at arXiv:2605.20579. The proof has been human-verified but has not yet completed formal peer review.