OpenAI's disproof of the Erdős unit distance conjecture was generated in a single model run without human mathematical intervention
Assessment
Evidence favors the claim, but the chain is incomplete or the sources are secondary.
The available record favors the claim, but the record is almost entirely OpenAI's own account, relayed and endorsed rather than independently witnessed. OpenAI's proof document states that the problem "was solved in a completely automated fashion": the model received an AI-written statement of the problem, its output went to an automated grading pipeline, and only after the grader passed it did researchers begin reading. The nine mathematicians who verified the result restate this in the first footnote of their companion paper, describing a file "first mathematically generated in one shot" and afterwards refined only in exposition; Gowers calls it the first famous open problem solved by AI "with no human intervention once it had been trained and then given the problem to solve," and Litt describes being asked to check a solution the model had already produced. Bloom notes that humans at OpenAI and in the companion paper later improved the proof substantially, but that improvement followed generation and does not bear on the claim.
Two artifacts support the account beyond OpenAI's word. The problem statement OpenAI says the model received, reproduced in its document, asks the model to resolve the planar unit-distance problem and says nothing about counterexamples or number fields. And the published chain of thought, as read by Shankar, Korbonits, and others, spends most of its length rejecting unrelated approaches (hypercubes, rational points on the unit circle, Dirichlet units) before arriving at CM fields of growing degree, which is hard to square with a prompt that named the method. This is why the prompt contained no hints toward a counterexample or number-field methods stands as the claim's central premise and currently looks sound.
The dissent comes from commentary rather than from anyone with access to the process. One line holds that "one shot" hides millions of filtered runs; another infers, from Tsimerman's remark that Boris Alexeev had once suggested a number-field approach to him, that the same idea must have reached the model's prompt. Neither offers documentary evidence, and the second relies on a suggestion (bounded-degree fields) that differs from the move that worked (unbounded degree). What survives of the skeptical case is a genuine limit on verification: labs do not disclose how many failed attempts precede their announced results, the released chain of thought is a rewritten summary rather than the raw transcript, and no outside party observed the run. Gowers himself allows that more computation may lie behind each presented step. The claim is therefore well supported as a description of the successful run, and would move to verified only if OpenAI released the raw prompt, transcript, and run logs for independent inspection, or move against if such records showed steering or a primed context.
Full reasoning: the evidence and decisions behind this verdict
Sources read in full this pass: the companion paper "Remarks on the disproof of the unit distance conjecture" (Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, Wood; arxiv.org/html/2605.20695v1), Alex Korbonits's essay "The Verification Problem" (korbonits.com/blog/2026-05-23-the-verification-problem/), both of Anatol Wegner's AIchats posts (aichats.substack.com/p/openai-disproves-the-unit-distance and aichats.substack.com/p/openai-disproves-the-unit-distance-8cd), and the first half of an explainer page at cuberoot.me that paraphrases OpenAI's statement on AI use. OpenAI's proof document (cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf) and its announcement page could not be opened by the reading tool (encoding error and HTTP 403 respectively); the statement on AI use and the problem statement are known here through consistent quotation in Wegner's first post, Korbonits, cuberoot.me, and a search index excerpt of the PDF itself, which agree word for word on the key sentences.
What the claim asserts and how the evidence bears on each part. "Single run": OpenAI's statement describes one generation followed by machine grading; the companion paper's footnote says "generated in one shot." No source with knowledge of the process contradicts this. The skeptical reading that the surviving run was selected from a large pool is conjecture ("likely generated millions of ... traces") with no supporting information, and even if true would not make the surviving proof multi-run; it would bear on how to interpret "one shot," which is the point the linked subclaim about undisclosed failed attempts captures. "Without human mathematical intervention": three pieces of evidence. First, the reproduced problem statement ("Resolve Erdős's planar unit-distance problem completely: ν(n) ≤ n^{1+O(1/log log n)} as n→∞? Equivalently, determine whether there exist absolute constants...") is a neutral formulation of the conjecture, and OpenAI says it was itself AI-written. Second, the chain of thought, per Shankar's first-hand reading in section 8 ("trying out a vast array of ideas from a wide range of mathematics") and Korbonits's account of the rewritten transcript (Boolean hypercube, powers of (3+4i)/5, Dirichlet units, each rejected for a stated reason before the "Suppose optimistically" paragraph), shows a search that a primed prompt would have made unnecessary. Third, the verifiers' testimony: Litt was asked by Sellke and Sawhney to check a solution already produced; Gowers's "no human intervention once it had been trained and then given the problem." Bloom's remark that the proof was "significantly improved by the human researchers at OpenAI" concerns post-generation improvement (the companion paper itself describes simplifications such as Sawin's single-split-prime idea and the pro-2 tower), consistent with the footnote's "expositionally refined."
The primed-prompt hypothesis examined directly. Wegner's second post rests on Tsimerman's sentence: "On Boris Alexeev's suggestion, I thought about this problem with the idea of making a counterexample stemming from a varying family of bounded degree number fields." Three problems. Alexeev's OpenAI affiliation is asserted without source and could not be checked within this pass's search budget. The suggestion was bounded degree; Tsimerman says increasing degree "occurred to me" as his own extension, and Sawin's section explains why the bounded-degree approach gives no signal to vary the field, so the idea that reached Tsimerman is not the idea that worked. And the hypothesis predicts a chain of thought that begins with number fields, whereas the transcript as described spends most of its length elsewhere. The hypothesis is therefore unsupported, though not strictly refuted, since the raw prompt and any system context remain private.
Residual uncertainty, which keeps the verdict at supported rather than verified: the entire affirmative record is a first-party account; the released chain of thought is explicitly a rewritten summary, and Gowers notes that "a lot of additional 'actual thought'" may sit behind each presented step; the number of runs launched is undisclosed; and the supporting subclaim that researchers first examined the proof only after an automated grader rated it correct rests on the same statement. The credence of 0.8 applies to the literal claim (the successful proof came from one generation, with no human supplying mathematics between problem statement and output); a stronger reading on which no failed runs or prompt engineering preceded it would deserve a lower number, since that is precisely what cannot be checked.
Instances: three affirm (OpenAI's statement, the nine verifiers' footnote, Korbonits repeating the statement), two deny (Wegner's two chatbot-conversation posts, one author). The affirmations all trace to OpenAI's statement, so they are one voice on the process plus the verifiers' first-hand reading of the transcript; the denials offer no evidence about the process. The lopsidedness in standing, not the count, is what makes this supported rather than contested. What would change the verdict: release of the raw prompt, system context, transcript, and run logs (toward verified if they match the account; toward contradicted if they show steering or a primed context); or credible testimony from anyone involved that the prompt or context carried mathematical direction.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
OpenAI's statement on AI use describes the model receiving an AI-written statement of the problem and returning a proof that was machine-graded before any person read it, and the nine external verifiers report the same file as generated in one shot and only afterwards refined in exposition. Given that the prompt contained no hints toward a counterexample or number-field methods, and given that OpenAI researchers first examined the proof only after an automated grader had rated it correct, no human mathematics entered between the problem statement and the finished proof, and the claim follows.
Granting its premises, the inference goes through: a neutral problem statement, a single generation, machine grading, and human reading only afterwards leave no point at which human mathematics could have entered. The argument's weight rests on the prompt containing no hints toward a counterexample or number-field methods, which is supported by the reproduced problem statement and by the exploratory shape of the chain of thought but cannot be inspected beyond what OpenAI has published; the grader-before-humans sequence adds little independently, since it comes from the same statement. The caveat is that every premise is a first-party account relayed by verifiers who joined after the fact, so the argument establishes the claim to the extent that account is trusted, and no further.
Because AI labs do not disclose how many failed attempts precede their announced results, the "one shot" that reached the public may be the surviving trace of a large pool of runs filtered by machine graders, so that the process which produced the disproof was a selection procedure rather than a single run. Skeptics add that an OpenAI-affiliated mathematician had earlier suggested a number-field approach to the problem to one of the verifiers, and infer that the same idea reached the model through its prompt or hidden context, in which case the prompt was not a neutral statement of the problem and human mathematics did intervene.
The argument does not reach its conclusion even granting what it can establish. That labs do not disclose how many failed attempts precede their announced results is well attested, but an undisclosed number of failed runs does not make the successful run more than one run, nor does selection by a machine grader amount to human mathematical intervention; it limits how far the claim can be verified rather than showing it false. The primed-prompt branch needs the prompt-neutrality premise to fail, and it offers only an inference from Tsimerman's remark about a colleague's earlier suggestion, which concerned a different approach (bounded rather than growing degree) and which the published problem statement and the chain of thought's early dead ends cut against. The argument would become live if prompt or run records surfaced showing steering.
Provenance
Where this claim has been said, linked to its canonical form.
Every affirmation of this claim traces to a single first-party document: the statement on AI use in OpenAI's own proof paper, which says the model was given an AI-written problem statement, its output was machine-graded, and only then did researchers look at it. The nine external verifiers restate that account in their companion paper's first footnote and in Gowers's and Litt's remarks, but they were brought in after the fact and did not witness the generation; what they add first-hand is a reading of the model's chain of thought and the observation that later human work improved the exposition rather than supplying the mathematics. The denials come from one author across two chatbot-conversation posts, which quote the same OpenAI statement accurately and then conjecture, without evidence about runs or prompts, that many filtered attempts or a primed prompt lie behind the "one shot". A reader should open the companion paper first, then OpenAI's proof document for the exact wording of the AI-use statement and the reproduced problem statement.
Throughout this document, phrases such as “AI proof” or “GPT proof” all refer to the same file, which was first mathematically generated in one shot by an internal model at OpenAI, and then expositionally refined through human interactions with Codex.
First footnote of the nine-mathematician companion paper presenting a human-verified version of the proof. Gowers adds in his section that this is the first famous open problem "solved by AI with no human intervention once it had been trained and then given the problem to solve," and Litt calls it "the first example of a result produced autonomously by an AI" that he finds exciting in itself. Bloom notes the original AI proof was valid but "significantly improved by the human researchers at OpenAI" afterwards.
Asserted without evidence of the source's own. The nine verifiers state the one-shot generation as fact but were not present for it; their knowledge of the process comes from OpenAI, and the paper offers no independent check of how the file was produced. What the paper does supply first-hand is a reading of the model's chain of thought, which several authors describe as exploring many approaches before settling on number fields, and the observation that later human work was expository and improving rather than generative. Worth reading closely: The footnote, Gowers's and Litt's remarks on autonomy, Bloom's note on later human improvement, and Tsimerman's remark about Boris Alexeev's suggestion are the passages both sides of this dispute build on.
OpenAI didn’t just ask the model once and get a perfect proof. They likely generated millions of 100+ page long reasoning traces and used another AI to filter them until one looked promising enough to hand to a human. The human didn’t guide the math step-by-step this time, but the computational search space required to find this “one shot” was likely astronomical.
A published conversation with Gemini critiquing the announcement. It quotes OpenAI's statement on AI use, then argues that "one shot" hides a large filtered pool of runs and months of scaffolding work, while conceding that humans did not guide the mathematics step by step in this case. The denial is framed as likelihood ("likely"), with no evidence about the actual number of runs.
Asserted without evidence of the source's own. The piece is a conversation with a chatbot, and the denial is the chatbot's conjecture ("likely generated millions of ... traces") offered without any information about the number of runs. It quotes OpenAI's statement on AI use accurately and concedes that humans did not guide the mathematics step by step. Its later point that the published chain of thought is a rewritten summary rather than the raw token sequence is accurate and is the piece's most substantive observation.
The proof was produced in one shot. According to the paper’s own statement on AI use, the model’s output was first sent to an AI grading pipeline, which reported high confidence the solution was correct. Only after that did human researchers at OpenAI start reading.
An essay arguing that verification, not proof generation, is the bottleneck for AI mathematics; it restates OpenAI's statement on AI use and reads the rewritten chain of thought as a sequence of stated dead ends before the decisive idea.
Asserted without evidence of the source's own. The essay explicitly rests its one-shot statement on OpenAI's own statement on AI use and adds no independent knowledge of the process; its own contribution is a reading of the rewritten chain of thought as a chain of stated dead ends before the decisive idea.
The AI didn’t disprove the Erdős conjecture. Boris Alexeev and the OpenAI math team disproved it, and they used an LLM as a multi-million-dollar calculator to execute their intuition.
A published conversation with Gemini in which the author notes Tsimerman's remark that Boris Alexeev had suggested a number-field approach, asserts that Alexeev works at OpenAI, and the chatbot concludes the prompt was "almost certainly" primed with the method and that the autonomous one-shot narrative is "officially busted". No documentary evidence about the prompt is offered.
Asserted without evidence of the source's own. The entire case rests on one sentence in the companion paper, Tsimerman's remark that Boris Alexeev had suggested to him a counterexample from a varying family of bounded degree number fields, plus the unsourced assertion that Alexeev works for OpenAI. From this the chatbot infers what the prompt "almost certainly" contained. No prompt text, system context, or testimony is produced, and the piece does not engage with the problem statement OpenAI published, which contains no such hint, nor with the chain of thought's early exploration of unrelated approaches. The suggestion it cites (bounded degree fields) is also not the move that worked (fields of unbounded degree), which Tsimerman describes as his own further thought.
This problem was solved in a completely automated fashion. Our internal model was given an AI-written statement of the problem, and its output was sent to an AI grading pipeline, which indicated high confidence that the solution was correct.
The "Statement on AI Use" in OpenAI's own proof document, the originating account of the process; it goes on to say that internal researchers were involved only after this point. The document also reproduces the problem statement the model was given, a neutral request to resolve the planar unit-distance problem. The PDF could not be opened directly in this pass; the passage is as quoted consistently by secondary sources and a search index of the document.
This is the root of every affirmation of the claim; the verifiers, commentators, and press all take the process description from here. The document could not be opened in this pass, so the passage is known through consistent secondary quotation. By its nature it is a first-party account with no external witness, and the document reportedly also reproduces the neutral problem statement given to the model, which is the one checkable artifact bearing on the primed-prompt dispute. Worth reading closely: Reading the full statement on AI use and the reproduced problem statement would settle exactly what OpenAI claims about the prompt, the grading pipeline, and when humans first intervened.
How these sources relate
- https://arxiv.org/html/2605.20695v1 restates https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf, faithfully. The verifiers' account of the generation process is what OpenAI told them; they were brought in after the fact to check correctness. The footnote restates OpenAI's statement on AI use ("solved in a completely automated fashion") in slightly different words, adding the detail that later refinement was expository, which OpenAI's statement also implies. No strengthening detected. Judged from the citing document alone.
- https://korbonits.com/blog/2026-05-23-the-verification-problem/ restates https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf, faithfully. The essay names OpenAI's statement on AI use as its source for the sequence and restates it accurately. Judged from the citing document alone.
- OpenAI disproves the unit distance conjecture responds to https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf, faithfully. The piece quotes OpenAI's statement on AI use accurately and then disputes the inference from it; the quotation itself is faithful, the rebuttal is speculative. Judged from the citing document alone.
- https://aichats.substack.com/p/openai-disproves-the-unit-distance-8cd cites in support https://arxiv.org/html/2605.20695v1, stating it more strongly than that document supports. The quotation of Tsimerman is accurate, but the piece uses it to assert that the model's prompt was primed, a conclusion the companion paper nowhere supports and whose footnote and Gowers's remark directly contradict. Tsimerman's remark concerns his own earlier attempt, and the approach he names (bounded degree) differs from the model's (unbounded degree).
- OpenAI disproves the unit distance conjecture and https://aichats.substack.com/p/openai-disproves-the-unit-distance-8cd share an author. Both posts are by Anatol Wegner on the AIchats Substack; the second is labelled "Act II" and opens as "a continuation of a previous conversation" about the same announcement. Both are also mirrored on Medium under the same author.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Sep 16, 2026. Every judgment on this page is accompanied by a reasoning trace.