When AI Solves the Problems Mathematicians Were Saving for Later

Gold disappears as we mine it. What if the same is true of mathematics’ best unsolved problems? As AI gets better at finding proofs, it may consume intellectual territory that humans can explore only once.

An Unsolved Problem Is a One-Time Opportunity

Mathematician Terence Tao has compared public unsolved problems to a nonrenewable resource. Once a problem is solved, it never becomes genuinely unknown again.

That matters because mathematics is not merely a production line for correct answers. The struggle toward a proof creates new concepts, exposes weaknesses in existing theories, and trains the researchers who will tackle the next generation of questions.

Think of an unsolved problem as an unclimbed mountain. Once someone reaches the summit and publishes the route, others can repeat the ascent. They cannot recreate the original expedition, complete with its dead ends, surprises, and uncertainty.

If AI begins solving major problems in volume, mathematics will gain knowledge. But humanity may lose some of the terrain where mathematical intuition is formed.

The deeper question is not whether AI will replace mathematicians. It is how much unknown territory will remain for humans to explore firsthand.

AI Has Moved From Calculation to Proof

Earlier mathematical software mostly behaved like a powerful calculator or search engine. Newer systems can select strategies, combine ideas, and produce arguments that formal proof tools can verify.

Google DeepMind’s AlphaProof and AlphaGeometry 2 demonstrated that shift at the 2024 International Mathematical Olympiad. Together, they scored 28 out of 42 points, equivalent to a silver-medal performance. The systems solved proof-based problems in algebra, number theory, and geometry rather than merely performing arithmetic faster than humans.

An Olympiad problem is not the same as an open research problem. Contest questions are designed by people who already know that a solution exists. At the research frontier, mathematicians may not know whether a conjecture is true, false, or even framed correctly.

Still, the direction is hard to miss. AI is moving beyond retrieving known methods and into searching for new proofs. Paired with formal verification systems, it can also detect gaps that human reviewers might overlook.

That combination changes the economics of mathematical discovery. Proof generation becomes cheaper, faster, and easier to scale.

The Scarce Resource Is the Moment of Discovery

Mathematics itself is unlikely to run out. Every theorem creates new questions, and every solved problem reveals another layer of structure.

But valuable problems are not generated automatically. The best ones connect distant theories, expose important patterns, and push an entire field forward when solved. They resemble an intellectual commons that generations of researchers have collectively cultivated.

If AI races through that commons, the number of known theorems could soar. Yet fewer humans may develop the intuition that comes from spending months or years trapped inside a hard problem.

That creates an uncomfortable paradox. Civilization could possess more mathematical knowledge while having fewer people who understand why particular results matter.

The speed of production may outrun the speed of comprehension. A database full of machine-verified proofs is impressive, but it is not the same thing as a mathematical culture capable of interpreting them.

Should AI-Discovered Proofs Be Embargoed?

One proposal is to delay publishing proofs discovered by AI. Human researchers would get a protected window to attempt the problems independently, much as cybersecurity teams sometimes receive time to patch vulnerabilities before details become public.

The incentives point the other way. A major proof brings academic prestige, commercial advantage, and technological leverage. When universities, labs, and well-funded companies are competing, voluntary restraint is unlikely to survive first contact with a landmark result.

There is also a real cost to withholding knowledge. A mathematical breakthrough might unlock progress in cryptography, physics, computing, or biology. Preserving the pleasure of discovery is a difficult argument when delayed publication could slow practical advances.

A more workable approach would be to create separate tracks. Some competitions and research programs could remain human-only, while others explicitly encourage human-AI collaboration. Publications could also document which steps came from AI and which depended on human insight.

Mathematics education would need to change accordingly. Finding the answer will matter less when machines can do it cheaply. Formulating important questions, interpreting proofs, and connecting distant concepts will matter more.

Mathematics Will Need More Prospectors Than Miners

AI solving unsolved problems would not end mathematics. It would change the mathematician’s role from extracting answers to deciding where humanity should dig next.

The coming challenge is not a shortage of theorems. It is preserving enough room for humans to experience the uncertainty from which understanding grows.