AI Cracks an 80-Year-Old Maths Puzzle: Why It Matters Beyond Mathematics

For about eighty years, mathematicians struggled with a famous unsolved problem first posed in 1946 by the Hungarian mathematician Paul Erdos. In May 2026, an artificial intelligence (AI) model managed what humans could not, by making a genuine breakthrough on it. The challenge is called the planar unit distance problem, and it belongs to a large collection of brain-teasers known as 'Erdos problems'. Mathematicians regard the new result as a landmark because, they say, had a human produced it, it would have deserved publication in a top mathematics journal, a bar that no earlier AI-generated proof had reached.

The problem sounds simple. Take a sheet of paper and place dots on it. As you add more and more dots, even up to millions or trillions, how do you arrange them so that you get the largest possible number of pairs of dots that are exactly the same distance apart? Erdos guessed that the best arrangement was something close to a square grid. For decades mathematicians believed this was true but could not prove it. The AI model, instead of proving the old guess, did something more surprising: it disproved it. It discovered an entirely new family of arrangements that does better than the square grid, finding more equal-distance pairs. The new pattern is hard to even picture, and it was built by combining ideas from different branches of mathematics.

What makes this result stand out from earlier, doubted AI claims is how the model reached it. In the past, some announcements of AI 'solving' maths turned out to be cases where the system merely found answers that already existed in published papers, which experts dismissed as misleading. This time, according to the mathematicians who checked it, the AI read academic papers, understood them well enough to apply their methods in fresh ways, and connected fields that experts had not linked before, such as using tools from algebraic number theory to solve a question in discrete geometry. Nine mathematicians verified the result. Around the same time, another AI research lab said its own system had solved several Erdos problems too, suggesting this is part of a broader wave.

Why does maths matter so much as a test for AI? Unlike creative writing, where people can argue endlessly about quality, a mathematical proof is either right or wrong, and anyone who understands it can agree. That makes maths a clean, honest measuring stick for whether an AI is truly reasoning rather than just guessing or copying. AI systems also have one advantage over human specialists: humans tend to know one field deeply, while an AI is less trapped by assumptions about which fields naturally connect, so it can spot useful links across distant areas of knowledge and search tirelessly for solutions.

Still, experts urge caution and balance. The same large language models that crack a deep problem in one attempt can fail at simple arithmetic in another, and they still produce confident but wrong answers, a flaw called hallucination. Human checking remains essential. One lab relied on human mathematicians to verify and simplify the output, while another connected its AI to formal proof-checking software (such as the system called Lean) that can confirm a proof line by line. The deeper significance, researchers say, is the future: if AI can read and extend maths papers, it may one day do the same in biology, physics, medicine and engineering. For aspirants, the takeaways are the idea of AI doing original research, the continuing need for human verification, the persistence of hallucination as a limitation, and the role of Indian institutions, with an IIT Delhi professor among the experts who assessed the breakthrough.