Erdős Problems Solved by AI: What’s Really New?

Quick answer: Yes. AI has produced new solutions to some Erdős problems, while other reported successes were rediscoveries of published work or partial results. In January 2026, GPT-5.2 Pro and Harmonic’s Aristotle produced a Lean-checked proof of Problem 728. On May 20, 2026, OpenAI announced an AI-generated disproof of Problem 90, Erdős’s 1946 unit distance conjecture, which outside mathematicians checked and described in a companion paper. Each claimed result needs its own check of novelty, scope and verification.

Sources checked: November 22, 2026.

Paul Erdős left behind more than a thousand problems: short questions about numbers, graphs and geometry, some easy, some so hard he offered thousands of dollars for a solution. For decades they were a human playground. Then, in less than a year, AI tore through them. This is the story of how it happened, what was really solved, and why the mathematicians who loved these problems have mixed feelings. (New to the topic? Start with Erdős problems explained.)

Timeline: AI and the Erdős Problems

DateWhat happened
May 2023Thomas Bloom (University of Manchester) launches erdosproblems.com to collect the problems
Nov 2025AI searches find forgotten solutions in old papers; Mehtaab Sawhney and Mark Sellke report published solutions to 10 problems listed as open, plus partial results for 10 more
Dec 25, 2025An AI-produced proof of Problem 333 turns out to match a 1977 paper by Erdős and Newman
Jan 2026Problem 728 solved with GPT-5.2 Pro, with a Lean proof from Harmonic’s Aristotle (Jan 6); a 24-author Google DeepMind study classifies its results as two full resolutions, two partial results, four independent rediscoveries and five literature identifications
May 2026A DeepMind study reports 9 of 353 attempted problem statements resolved, including variants; on May 20, OpenAI announces an AI disproof of the unit distance conjecture (Problem 90)
June 2, 2026The Leiden Declaration calls for “tight guardrails” on AI in mathematical research
Aug 2026OpenAI says an unreleased model it calls Astra solved Problems 146, 180 and 183, with human help on the manuscripts and later Lean certificates; Quanta reports the site’s catalog at 565 solved and 652 open (counting all results, human and AI)
Oct 6, 2026Bloom removes solved/open labels and totals and pauses new comments and proof claims; the same day OpenAI publishes hundreds of research manuscripts at various stages of verification

Act 1: The Literature Search Surprise

The first wave was not new mathematics at all. AI models are good at reading enormous amounts of text, and many “open” Erdős problems had in fact been answered in obscure papers. In November 2025, Sawhney and Sellke reported that ChatGPT helped them find published solutions to 10 problems listed as open, and relevant partial results for 10 others. Useful work, but not discovery.

The lesson came on Christmas Day 2025, when an AI-produced proof of Problem 333 was posted and then found to match a 1977 paper by Erdős and Donald Newman. The proof was not wrong; it simply was not new. After that, members of the community stressed checking the literature before announcing anything.

Act 2: Real Proofs, Checked by Computer

In early January 2026, Kevin Barreto used GPT-5.2 Pro on Problem 728. After the community clarified which version of the statement Erdős intended, Harmonic’s Aristotle system produced a proof of that version in the proof language Lean on January 6. Later that month, a 24-author Google DeepMind study reported its own results honestly sorted: two full resolutions, two partial results, four independent rediscoveries and five cases of finding the answer in the literature. By May, another DeepMind study resolved 9 of 353 problem statements it attempted, including subproblems and variants, at a cost the team estimated at a few hundred dollars per problem.

These were mostly the easier problems on the list. Terence Tao and others pointed out that AI was harvesting “low-hanging fruit”: problems that were open mainly because nobody had looked hard at them.

Act 3: The Breakthrough

Then came a problem nobody called easy. In 1946 Erdős asked how many pairs of points among \(n\) points in the plane can be exactly distance 1 apart, and conjectured that square-grid-like arrangements are essentially best. On May 20, 2026, OpenAI announced that an unreleased general-purpose reasoning model had found a construction beating that conjectured growth rate. Mathematicians including Noga Alon, Thomas Bloom and Tim Gowers wrote a companion paper. Gowers wrote that no previous AI-generated proof had come close.

Experts also noted caveats: verification went relatively smoothly for mathematicians with the right background in algebraic number theory, and OpenAI did not share how many attempts failed. Gowers’s praise is an expert assessment, not a journal acceptance.

Act 4: The Backlash

Success brought problems. Proof claims poured in faster than anyone could read them. On October 6, 2026, Bloom removed the site’s solved/open labels and totals and paused new comments and proof claims, while continuing to record mathematical results and focusing on explaining the mathematics. The same day, OpenAI released hundreds of research manuscripts at different stages of verification, some touching Erdős problems; one independent tally two days later counted 40 affected entries, mixing claimed new solutions, alternative proofs and partial results. In August, Quanta had already quoted Noga Alon explaining why he stopped working on these problems himself: “Once AI started to solve them, there is no point anymore.”

The Leiden Declaration, posted on June 2, 2026 and signed by more than 1,500 people within days, named three norms at stake: checking proofs, crediting ideas fairly and keeping research open to everyone.

How AI Actually Solved Them

MethodWhat the AI didExample
Literature searchFound a forgotten paper that already answered the questionSawhney and Sellke’s 10 problems
Proof writing plus formal checkingWrote a proof that Lean confirmed step by stepProblem 728
Construction searchFound a clever example that breaks a conjectureUnit distance problem
Mass attemptsTried hundreds of problems cheaply and kept the successesDeepMind’s 353-problem study

What Counts as “Solved by AI”?

When you see the headline, ask four questions. Was the result new, or was it in the literature? Was the original problem solved, or an easier variant? Was the proof checked, by Lean or by experts? And how much did humans guide the AI? A result can be a rediscovery and still be useful; it just is not a breakthrough. For the bigger picture, see Can AI Do Math?

Common Misconceptions

“AI solved all the Erdős problems.” No. Hundreds remain open, and most AI successes were on easier problems.

“Every AI Erdős result is new.” No. A large share were rediscoveries of solutions already in print.

“Humans played no role.” No. People chose the problems, checked the literature, verified proofs and wrote companion papers.

Watch: AI and the Erdős Problems in Video

If you learn better by watching, these videos are worth your time.

AI just disproved the biggest math conjecture so far [Unit Distance Conjecture] (Dr. Trefor Bazett). A clear walk-through of the unit distance problem and what the AI construction showed.
The Erdős Breakthrough (OpenAI). OpenAI’s own account of the unit distance result; watch it alongside the independent coverage.
Terence Tao On AI Solving Erdős Problems (Dwarkesh Clips). Terence Tao on what AI successes on Erdős problems do and do not mean.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What happened with the AI proof of Problem 333 in December 2025?
Found a bug? Let us know!
2. Which tool checked the GPT-5.2 Pro proof of Problem 728?
Found a bug? Let us know!
3. What did an OpenAI model do on May 20, 2026?
Found a bug? Let us know!
4. What did erdosproblems.com change on October 6, 2026?
Found a bug? Let us know!
5. Which of these is NOT a new solution?
Found a bug? Let us know!

Frequently Asked Questions

Did AI solve Erdős problems?

Yes, some. AI produced new solutions to problems such as Problem 728 (January 2026, Lean-checked) and disproved the 1946 unit distance conjecture, Problem 90 (May 2026). Many other claimed solutions were rediscoveries of published results or partial results, and hundreds of problems remain open.

What was the biggest Erdős problem solved by AI?

The unit distance conjecture from 1946. On May 20, 2026, an OpenAI model found a construction that disproves Erdős’s conjectured growth rate, and mathematicians including Tim Gowers wrote a companion paper.

How did AI find old solutions to Erdős problems?

By searching huge amounts of mathematical literature. In November 2025, Mehtaab Sawhney and Mark Sellke reported that ChatGPT helped them find published solutions to 10 problems listed as open, and partial results for 10 more.

Why did erdosproblems.com stop showing solved problem counts?

On October 6, 2026, amid a flood of AI-generated proof claims, Thomas Bloom removed the solved/open labels and totals, paused new comments and proof claims, and refocused the site on explaining the mathematics, while still recording results.

What is the Leiden Declaration?

A June 2026 statement, signed by more than 1,500 people within days, calling for tight guardrails on AI in mathematical research around checking proofs, crediting ideas and keeping research open.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Erdős Problems Solved by AI: What’s Really New? - Effortless Math"?

No one replied yet.

Leave a Reply