Skip to main content

GPT-5 literature searches prompted six Erdős database status changes after finding earlier human results

Between 12 and 14 October 2025, GPT-5-assisted literature searches helped the Erdős Problems project identify earlier human results for Problems 339, 494, 621, 822, 903, and 1043, after which maintainers changed the six database statuses.

The Reality Check

These were literature-recovery events, not six new GPT-5 proofs. Contributors examined the cited papers and decided that existing human results matched the database problems. The public record does not expose every query, rejected candidate, or unsuccessful search, so this is not a controlled benchmark of GPT-5’s literature-search reliability.

Context

The Erdős Problems project maintains more than one thousand problems spanning decades of mathematical literature, different terminology, and many subfields. A problem can remain listed as open even when a relevant human result exists elsewhere in the literature.

Researchers used GPT-5 through ChatGPT, the OpenAI API, and internal tools to search many problems in parallel and filter candidate references. Human contributors then checked whether the papers actually answered the database entries.

A 12 October repository commit changed Problem 339 from open to proved and Problem 1043 from falsifiable to disproved. A later commit changed Problems 494, 822, and 903 from open to solved and Problem 621 from falsifiable to solved. The project’s maintained contribution table continues to classify all six as GPT-5 literature-search recoveries.

The project warns that its collection is not a controlled benchmark and that success counts are affected by selection and reporting bias. Later GPT-5 work expanded beyond this six-entry milestone and should not be conflated with it.

THE TAKEAWAY

GPT-5 did not solve six previously unsolved Erdős problems from scratch. It helped researchers find earlier human results that had been missed by a maintained problem database, and those findings survived expert review strongly enough to change six public status records. The milestone is evidence for AI-assisted knowledge maintenance and retrieval, not autonomous theorem proving.

Continue the Thread

AI-Assisted Scientific Literature Discovery

Tracks evidence that AI systems can find, interpret, and connect scientific literature well enough to change expert-maintained knowledge records or research decisions.

Related threads

Sources

Early science acceleration experiments with GPT-5

arXiv / OpenAI researchers and collaborators

Primary EvidencePreprint · Developer / Vendor Claim

Used for: Search methodology; use of ChatGPT, the OpenAI API, and internal tools; human verification; the later ten-problem full-solution list; partial-progress findings; detailed examples; and reproducibility limits.

AI contributions to Erdős problems

Erdős Problems project / GitHub

Strong ArtifactTechnical Artifact

Used for: Maintained attribution and dates for GPT-5 literature-search contributions, the distinction between literature recovery and primary mathematical contributions, and project warnings about incompleteness, selection bias, and benchmark interpretation.

Mark various problems as solved, add 1081

Erdős Problems project / GitHub

Strong ArtifactTechnical Artifact

Used for: Direct repository evidence that Problems 494, 621, 822, and 903 changed to solved on 14 October 2025, establishing the event date for the six-entry cluster.

Update 339 to proved and 1043 to disproved

Erdős Problems project / GitHub

Strong ArtifactTechnical Artifact

Used for: Direct repository evidence that Problem 339 changed from open to proved and Problem 1043 changed from falsifiable to disproved on 12 October 2025.

Last checked Methodology 2.0.0