Table of Contents
Teach Kids to Read Science News: The Erdős Proof Case Study
Solved, disproved, or claimed? Teach kids to read science news using the 2026 Erdős conjecture story, where the same result got three different headlines.
If you want to teach kids to read science news, start with this: in October 2025, an OpenAI executive posted that GPT-5 had solved ten unsolved Erdős problems. It hadn’t. The model had found solutions already published in the mathematical literature, and Thomas Bloom, who maintains the Erdős Problems database, called the post “a dramatic misrepresentation.” Seven months later, in May 2026, OpenAI announced something real: a model had disproved a conjecture Paul Erdős posed in 1946. Nine mathematicians published a companion paper confirming it. The model succeeded on roughly half its attempts. This one story contains every trap: the hype, the retraction, the real result, the honest caveat, and headlines that used four different verbs for the same event. Here is how to walk a 7-year-old, an 11-year-old, and a 15-year-old through it.
Key Takeaways
- Four verbs do different work in a science headline: claimed (someone said it), solved (a problem closed), disproved (a conjecture shown false), verified (independent experts checked). Most AI-math coverage in 2026 blurred all four.
- October 2025: OpenAI’s then-VP said GPT-5 solved 10 Erdős problems. It had rediscovered published solutions. That’s a retraction, not a breakthrough.
- May 20, 2026: an OpenAI model disproved the Erdős unit distance conjecture (1946) using infinite class field towers from algebraic number theory. Noga Alon, Melanie Matchett Wood, Thomas Bloom, and six other mathematicians published remarks confirming the result.
- The model succeeded on about 50% of repeated runs, and OpenAI did not publish failure rates or compute time. Both facts belong in any honest headline about it.
- On June 2, 2026, mathematicians published the Leiden Declaration calling for guardrails on AI in mathematical research; it gathered 1,590 signatures within three days.
What actually happened, in order
The Erdős unit distance problem asks a simple question: if you place n dots on a piece of paper, how many pairs can be exactly one centimeter apart? Erdős conjectured in 1946 that there is a ceiling on that count, and for eighty years nobody could prove or break it. Most human effort went into proving the ceiling existed. As Bloom later noted, comparatively little went into trying to break it.
On May 20, 2026, OpenAI announced that one of its reasoning models had found an entirely new family of point configurations producing more unit distances than the conjecture permitted. The conjecture is therefore false. The surprising part was the method: the model reached into algebraic number theory, using infinite class field towers, the Golod-Shafarevich theorem from the 1960s, and results by Ellenberg-Venkatesh and Hajir-Maire-Ramakrishna. Nobody had connected those tools to this geometry problem before.
Nine mathematicians, Noga Alon (Princeton), Thomas Bloom (Manchester), W. T. Gowers (Cambridge), Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood (Harvard), published a companion paper on arXiv explaining and confirming the proof. Gowers said that if a human had written the paper, he would have recommended acceptance without hesitation.
Then the caveats. Science News reported that the model produced a correct result on roughly 50% of runs, and that OpenAI had not disclosed failure rates or compute time. Wood raised a different problem: “It’s not clear that there’s a way for [AI] to reasonably attribute the source of the ideas.” Bloom raised a third: “It could be right. It could be nonsense. Who’s going to be able to check this?” On June 2, mathematicians published the Leiden Declaration calling for tight guardrails around AI in mathematical research. By June 5 it had 1,590 signatures.
By August, Quanta Magazine reported that Erdős problems were falling at a steady clip: OpenAI announced ten more advances on August 1, Google DeepMind resolved four problems and rediscovered nine forgotten solutions, and human-AI collaborations handled dozens more. Alon’s reaction was blunt: “Once AI started to solve them, there is no point anymore.” Tao collaborated with a model on one of them.
Teach kids to read science news with four verbs, not one
This is the lesson that transfers to every science story your kid will ever read.
Claimed means a person or company said something. It carries no truth value. A press release is a claim. So is a tweet from a vice president.
Solved means a stated problem now has an answer. But watch the direction: a problem can be solved by proving the thing true, or by proving it false. Those are opposite outcomes and the word “solved” hides which one happened.
Disproved means a conjecture was shown false. This is genuinely exciting in mathematics because it eliminates a possibility permanently, but it often reads as failure to non-specialists. “AI disproves 80-year-old conjecture” sounds worse than “AI solves 80-year-old problem” and means something more precise.
Verified means someone independent checked. In mathematics, that’s a mathematician reading the proof line by line, or a proof assistant like Lean mechanically confirming each step. Verification is the scarcest thing in AI-math coverage, and it’s the only one that should change your confidence.
How to Teach Your Kid About Reading Science Headlines
Ages 5–8: The “who says?” game
Read any headline aloud and ask one question: “Who says that?” Then find the answer in the first paragraph. A company, a scientist, a teacher, a friend. That’s it. You are building the habit of separating the claim from the claimer, which is the whole foundation. Do it with three headlines a week, including sports and weather, so it doesn’t feel like school.
Ages 9–12: Verb hunt and rewrite
Give them a printed headline and ask them to circle the verb. Then ask: does this verb mean someone said it, did it, or someone else checked it? Then have them rewrite the headline with the most accurate verb they can. “AI Solves 80-Year-Old Math Problem” becomes “AI Program Finds Counterexample; Nine Mathematicians Confirm It.” Less punchy. More true.
Ages 13+: Trace it to the source
Hand them the Erdős story. Their job: find the original announcement, find the companion paper by the mathematicians, and find one article that got it wrong. All three are online and free. Then write three sentences: what was claimed, what was verified, and what is still unknown. A teenager who can do this reliably has a skill most adults lack.
The question to ask: “What would have to be true for this headline to be wrong?” A kid who can answer that is reading critically. A kid who says “it’s from the news, so it’s true” is not reading at all.
Headline versus reality: the 2026 AI-math coverage
| The headline said | What actually happened | The verb that fits |
|---|---|---|
| ”GPT-5 solved 10 unsolved Erdős problems” (Oct 2025) | The model found solutions already published in the literature | Rediscovered |
| ”AI solves 80-year-old math problem, for real this time” (May 2026) | A model disproved the unit distance conjecture; nine mathematicians confirmed it | Disproved, then verified |
| ”AI achieves perfect score at the Math Olympiad” (Jul 2026) | Multiple AI systems scored 42/42 on IMO 2026 problems, graded by IMO organizers, after human contestants finished | Scored, officially graded |
| ”Erdős problems are falling to AI” (Aug 2026) | Real progress across many problems, plus concern about unvetted 100–200-page AI papers | Progressed, unevenly verified |
| ”AI is now a research mathematician” (various) | Models contribute; humans still verify, attribute, and decide what matters | Overstated |
Note the third row. At IMO 2026 in Shanghai (July 10–21, 117 countries, 666 contestants, 55 gold medals, seven perfect human scores), multiple AI systems were officially graded at 42/42 by the organizers, receiving the problems only after human contestants had finished, under strict time limits and no human intervention. That is a real, verified result, and it is still not the same as “AI does mathematics.” Our companion piece on what an AI perfect score at IMO 2026 actually proves unpacks the difference.
What to do at home this month
Keep a headline-and-reality notebook
One page per story. Left column: the headline as written. Right column: what the source document actually says. Three entries a week for a month, and your kid will start noticing the pattern without being told: the headline exaggerates the verb and drops the caveat.
Read one primary source together
Pick anything. The OpenAI post is short. The mathematicians’ remarks paper is mostly incomprehensible to a non-specialist, which is itself the lesson: real mathematics looks like that, and the headline version is a translation, and translations lose things.
Teach the 50% number
The most useful single fact in this whole story is that the model got it right about half the time. Ask your kid: if a friend gave you the right answer half the time, would you copy their homework? That question does more for AI literacy than an hour of lecturing, and it connects directly to how they should treat a chatbot’s math. Our piece on what a 50% success rate teaches kids about AI reliability goes deeper.
Look for who disagreed
Every real science story has a dissenter, and the dissenter is usually the most informative person in it. In this story it’s Bloom asking who can check a 200-page AI proof, and Wood asking how you attribute ideas a model didn’t cite. Teaching a kid to find the dissenter is teaching them that science is an argument, not an announcement.
What not to do
Don’t use this to teach cynicism. “The media always lies” is as lazy as believing every headline, and it leaves a kid with no way to tell good reporting from bad. The Science News and Quanta pieces on this story were careful and specific. Point at them as examples of the job being done well.
What to Watch For Over the Next 3 Months
- Week 4: Your kid asks “who says?” unprompted about something they read or watched. That’s the 5–8 skill generalizing.
- Month 2 red flags: They can name the headline but not the source. They treat a company blog post and a peer-reviewed paper as the same kind of evidence. They describe AI results in absolutes (“AI is better than mathematicians now”).
- Month 3 self-check: Hand them a fresh science headline cold and ask for three sentences: claimed, verified, unknown. If they can produce all three without help, the skill is theirs.
Frequently Asked Questions
Does my kid need to understand the math to learn from this story?
No. The transferable skill is reading the verb and finding the source, and that works on a story about dinosaurs or vaccines just as well. The mathematics is deliberately over everyone’s head, which makes it a good test case: you cannot fall back on evaluating the content, so you have to evaluate the evidence chain.
Is it actually true that AI disproved the conjecture?
Yes, with the caveats. The result stands and nine named mathematicians published remarks explaining it. The model produced a correct proof on roughly half of repeated runs, OpenAI has not published its failure rates, and at least one researcher reportedly reproduced the proof with publicly available models, which complicates the “frontier capability” framing.
How do I explain “disproved” to a 9-year-old without math?
Try this: “I think every swan is white. You find one black swan. You didn’t solve swan colors, you proved my rule was wrong, and now everyone has to stop using it.” Disproving is deleting a wrong idea, and deleting wrong ideas is most of how science makes progress.
Should this change how much I trust my kid’s AI homework helper?
It should calibrate it. A model that disproves an 80-year-old conjecture half the time is also a model that produces confident wrong answers the other half. The lesson for homework is not “don’t use it,” it’s “never accept it without checking,” which is exactly the habit the mathematicians are asking for.
Where can I find science news that does this well?
Quanta Magazine and Science News both handled the Erdős story with the verbs and caveats intact. Read the same story in two outlets with your kid and compare the headlines. The comparison teaches faster than either article alone.
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
- OpenAI. (2026). “Model disproves discrete geometry conjecture.” May 20, 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/
- Alon, N., Bloom, T. F., Gowers, W. T., Litt, D., Sawin, W., Shankar, A., Tsimerman, J., Wang, V., & Wood, M. M. (2026). “Remarks on the disproof of the unit distance conjecture.” arXiv 2605.20695. https://arxiv.org/pdf/2605.20695
- Science News. (2026). “Mathematicians want guardrails on AI after it cracked an Erdős problem.” https://www.sciencenews.org/article/ai-guardrails-erdos-math-problem
- Quanta Magazine. (2026). “Why the legendary Erdős problems are falling to AI.” August 3, 2026. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
- TechCrunch. (2026). “OpenAI claims it solved an 80-year-old math problem, for real this time.” May 20, 2026. https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/
- International Mathematical Olympiad. (2026). “IMO 2026, Shanghai, China.” https://www.imo-official.org/editions/2026/
- TechXplore. (2026). “AI models match humans’ top score in international math contest.” July 2026. https://techxplore.com/news/2026-07-ai-humans-score-math-contest.html
