AI Erdős Conjecture Explained: How to Tell Your Kid About It

Here is the AI Erdős conjecture explained in one paragraph, because the coverage made it sound harder than it is. Paul Erdős asked a simple question in 1946: if you put some dots on a page, how many pairs of them can be exactly one inch apart? For almost 80 years mathematicians believed the best you could do was arrange them roughly like a square grid. On May 20, 2026, an internal OpenAI model proved that belief wrong by finding an infinite family of better arrangements.

You can do the first half of that with your kid at the kitchen table in ten minutes. And the second half teaches something more valuable than the result: the difference between proving something is always true and finding one example that breaks it.

Key Takeaways

• The unit distance problem asks for the maximum number of pairs of points exactly distance 1 apart among n points in a plane. Erdős posed it in 1946 and offered a cash prize.
• Erdős conjectured the answer grew barely faster than n itself. The AI found arrangements with at least n^(1+δ) unit-distance pairs for a fixed δ > 0.
• Princeton’s Will Sawin later refined the result to show δ can be taken as 0.014.
• The proof came from a general-purpose reasoning model, not a system trained for math, and OpenAI says the model produced correct solutions in about half of repeated runs on the same prompt.
• The technique came from a completely different field, algebraic number theory, using infinite class field towers and Golod–Shafarevich theory.

The AI Erdős conjecture explained: what the problem actually asks

The unit distance problem is a question about counting. Place n points anywhere on a flat page. Draw a line between every pair of points that are exactly distance 1 apart. How many such lines can you get, at most?

Call that maximum u(n). Some arrangements are easy to analyze. Put all n points in a straight line one unit apart and you get n−1 pairs. A square grid with spacing 1 gives about 2n pairs. OpenAI’s announcement notes that the best previously known construction, a rescaled square grid, does slightly better: about n^(1 + C/log log n) for some constant C.

That exponent matters. Since log log n grows to infinity, the extra bit in the exponent shrinks toward zero as n gets large. So the best known constructions grew only barely faster than n. Erdős conjectured that was essentially the ceiling: an upper bound of n^(1+o(1)), where the o(1) tends to 0.

The 2005 book Research Problems in Discrete Geometry by Brass, Moser, and Pach calls it “possibly the best known (and simplest to explain) problem in combinatorial geometry.” Noga Alon of Princeton described it as “one of Erdős’ favorite problems.”

The AI’s result: for infinitely many values of n, there exist configurations with at least n^(1+δ) unit-distance pairs for a fixed δ greater than zero. That kills the conjecture. The original proof did not give an explicit δ; a later refinement by Princeton’s Will Sawin showed you can take δ = 0.014.

Conjecture, proof, counterexample: the distinction that matters

This is the concept worth teaching, and most adults are fuzzy on it too.

Term	What it means	Unit distance example	How hard to establish
Conjecture	A statement believed true but not proven	”You can’t beat roughly n unit-distance pairs” (Erdős, 1946)	Free to state; needs evidence to be taken seriously
Proof	An argument showing a statement is true in every case	Spencer, Szemerédi, and Trotter’s 1984 upper bound of O(n^(4/3))	Hard; must cover infinitely many cases
Counterexample	A single case where the statement fails	The AI’s infinite family of better configurations	Sometimes easier than proof; one is enough
Disproof	Showing a statement is false, usually via counterexample	What happened on May 20, 2026	Needs only one valid break

The asymmetry is the lesson. To prove Erdős right, you would have to rule out every possible arrangement of every number of points, forever. To prove him wrong, you needed one family of arrangements that beats the bound. Thomas Bloom of the University of Manchester, who maintains the erdosproblems.com database, made exactly this point to Science News: proving the conjecture rather than disproving it would have been “truly incredible” and required a different kind of creative insight.

Mathematicians did have evidence for the conjecture. Matoušek, and separately Alon, Bucić, and Sauermann, studied the problem with non-Euclidean distances and showed that “most” such distances obey the conjecture in some sense. The evidence pointed one way. The answer was the other way.

How to Teach Your Kid About Conjectures and Counterexamples

All three activities need paper, a pencil, and something to measure with.

Ages 5–8: The One-Inch Game

Draw four dots anywhere on paper. Give your child a ruler and have them find every pair exactly one inch apart, drawing a line for each. Count the lines. Then rearrange the four dots and try to get more lines.

The best answer for four points is more than most kids expect, and the search is the point. When they find a better arrangement than their first try, say the sentence out loud: “You just found a better example. That is exactly what the computer did.”

Ages 9–12: Make a Conjecture and Break It

Have your child do the one-inch game with 3, 4, 5, and 6 dots, writing down the best count they find each time. Then ask them to predict the answer for 7 dots before drawing anything. That prediction is a conjecture.

Now try to beat their prediction. If they succeed, they have produced a counterexample to their own conjecture, which is how this actually works in mathematics. If they cannot beat it, that is not proof, and saying so is the lesson: “You didn’t find a better one. That doesn’t mean there isn’t one.”

Ages 13+: The Grid Assumption

Show your teen why the square grid feels optimal: in a unit grid, most points have four neighbors at distance exactly 1, so you get roughly 2n pairs from n points. Then ask the harder question: why would anyone assume that is the best possible?

The honest answer is that nobody proved it, they just could not beat it for 80 years. Then explain what the model did differently. Instead of geometry, it used algebraic number theory. Erdős’s original construction can be understood through the Gaussian integers, numbers of the form a + bi. The AI’s proof swapped those for more complicated number fields with richer symmetries, which produce many more unit-length differences. Arul Shankar, a number theorist commenting on the result, noted that in the model’s reasoning trace “a significant majority of the thoughts are trying to construct a counterexample to the widely believed upper bound, rather than trying to prove it.”

The question to ask: “If you can’t find a better example, does that prove there isn’t one?”

What is genuinely new here, and what is being oversold

The honest version has both parts.

Genuinely new: OpenAI states this “marks the first time that a prominent open problem, central to a subfield of mathematics, has been solved autonomously by AI.” The proof was checked by external mathematicians who wrote a companion paper. Fields medalist Tim Gowers wrote in that paper: “if a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation. No previous AI-generated proof has come close to that.” Jacob Tsimerman, who had himself tried to construct a counterexample and failed, called it “a really impressive piece of work.”

Legitimately caveated: Melanie Matchett Wood of Harvard called it “a beautiful piece of mathematics that has been discovered” while questioning whether it represents an AI breakthrough specifically, noting that publicly available models might replicate it, and raising a fair transparency concern: OpenAI has not disclosed its failed attempts or incorrect solutions. Sébastien Bubeck of OpenAI stated the model produced correct solutions in about 50% of trial runs on the same prompt, and Science News noted that the data supporting these claims “have not been released or peer-reviewed” and the compute time was not revealed. In June 2026, more than 1,500 experts signed the Leiden Declaration calling for guardrails on AI mathematical claims, citing attribution, verifiability, and access. Terence Tao made a related argument about machine-assisted proof in the Notices of the AMS in January 2025: interactive theorem provers are changing practice while raising real questions about what counts as a proof.

There is also a cautionary precedent, documented by TechCrunch. Seven months before this result, an OpenAI executive claimed GPT-5 had solved ten previously unsolved Erdős problems. It had not; the solutions already existed in published literature, and the post was removed after criticism from Yann LeCun and Demis Hassabis. Our piece on why the Erdős problems are falling to AI covers how the community now handles these claims.

What to actually do at home

Use the word “conjecture” out loud

Kids make conjectures constantly (“the tall glass always holds more”). Naming the category teaches the concept better than any explanation. Then test it.

Let them be wrong without being corrected

The most useful moment is when a child’s confident prediction gets broken by a single example. Do not soften it. That experience is the whole scientific method in miniature.

Separate “I couldn’t find one” from “there isn’t one”

This is the single most transferable idea in this article, and it applies well beyond math. Practice the distinction whenever it comes up: no counterexample found is not the same as no counterexample existing.

Show that ideas travel between fields

The reason this proof surprised people is that the tools came from number theory, not geometry. Bloom wrote that the result “shows that there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected.” If your kid likes two unrelated subjects, that is an asset, not a lack of focus. Our article on how an AI borrowed number theory to solve geometry goes deeper on the mechanism.

What not to do

Do not tell your child that AI has solved math. One problem out of hundreds of open Erdős problems fell, with a roughly 50% run success rate, and it needed human mathematicians to verify and contextualize it. Overstating this is how kids conclude that learning math is pointless, which is the opposite of what the result shows.

What to Watch For Over the Next 3 Months

• Week 4: Your child uses “that’s a conjecture” or “here’s a counterexample” unprompted about something unrelated to math.
• Month 2 red flags: “Why learn math if AI does it?” If you hear this, the answer is that a human refined the result (Sawin’s δ = 0.014) and humans decided the problem mattered in the first place.
• Month 3 self-check: Ask whether more Erdős problems have fallen. Bloom’s erdosproblems.com database tracks the status of roughly a thousand problems, and the pace is the thing to watch.

Frequently Asked Questions

What is the unit distance problem in simple terms?

Place n points on a flat page and count how many pairs are exactly distance 1 apart. The question is what the maximum count can be as n grows. Erdős posed it in 1946, offered a prize, and it is often called the simplest-to-explain problem in combinatorial geometry.

Did the AI prove or disprove the conjecture?

Disprove. Erdős conjectured an upper bound of n^(1+o(1)). The model constructed an infinite family of configurations achieving at least n^(1+δ) unit-distance pairs for a fixed δ > 0, which breaks that bound. A counterexample only needs to work once.

Did human mathematicians check the proof?

Yes. OpenAI says the proof “has been checked by a group of external mathematicians,” who also wrote a companion paper explaining the argument and its significance. Noga Alon, Tim Gowers, Arul Shankar, and Jacob Tsimerman all commented publicly on the result.

Why does the 50% success rate matter?

Because it tells you the model is not reliably correct. OpenAI’s Sébastien Bubeck said correct solutions came in about half of repeated runs on the same prompt. That means someone competent had to check the output, which is exactly why mathematical training still matters.

Can my kid actually try this problem?

The counting version, yes, with a ruler and a few dots. The full problem is unsolved in general: the best upper bound, O(n^(4/3)), dates to Spencer, Szemerédi, and Trotter in 1984 and has not been substantially improved. Nobody knows the true answer.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. OpenAI. (2026, May 20). “An OpenAI model has disproved a central conjecture in discrete geometry.” https://openai.com/index/model-disproves-discrete-geometry-conjecture/
2. Hulick, K. (2026, June 8). “AI guardrails and the Erdős math problem.” Science News. https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem
3. Kakaes, K. (2026, August 3). “Why the legendary Erdős problems are falling to AI.” Quanta Magazine. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
4. Bellan, R. (2026, May 20). “OpenAI claims it solved an 80-year-old math problem, for real this time.” TechCrunch. https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/
5. Brass, P., Moser, W., & Pach, J. (2005). Research Problems in Discrete Geometry. Springer. https://link.springer.com/book/10.1007/0-387-29929-7
6. Bloom, T. “Erdős Problems.” University of Manchester. https://www.erdosproblems.com/
7. Tao, T. (2025, January). “Machine-Assisted Proof.” Notices of the American Mathematical Society, 72(1), 6–13. https://www.ams.org/notices/202501/rnoti-p6.pdf