Welcome to the AI crisis in math

Axiom published a set of solutions to longstanding problems in mathematics. The article describes the release as a bombshell because these are not routine homework questions. They are problems that have resisted serious efforts for a long time and sit near the frontier of mathematical knowledge.

The significance depends on more than getting answers. A convincing solution must explain why the answer follows from accepted definitions and earlier results. If an AI system produced such arguments, it would show that AI can contribute to work traditionally associated with highly trained mathematicians.

The article presents this development through a conversation with Robert Hart, The Verge’s London-based AI reporter. It also reports an existential crisis among leading mathematicians. The event therefore raises both technical and personal questions: how should researchers evaluate AI’s work, and what becomes distinctive about human mathematical labor?

A longstanding problem is a question that mathematicians have been unable to settle for an extended period. It may ask whether a statement is always true, whether a particular object exists, or how a mathematical quantity behaves. The defining feature is that no accepted proof or disproof has been found.

These problems differ from ordinary exercises because their solution methods are unknown. A student may apply a theorem already taught in class. A frontier problem may require a new idea, a new construction, or a connection between distant areas of mathematics. Even a promising approach can take years to complete or fail under close inspection.

In the article, Axiom’s reported solutions mattered because they addressed problems with that history of resistance. The source does not name or describe each problem. So the central point is their status as persistent research challenges, rather than any one specific conjecture or field.

Ordinary mathematical exercises usually test known methods. The solver is expected to recognize a technique, carry out the steps, and reach an answer within a limited time. A longstanding research problem is different. It may have no known route to a solution and can remain open across many years of expert work.

That gap is the important comparison in the article. Axiom’s reported results concerned problems that had resisted mathematicians, rather than standard classroom questions. Solving one would therefore be closer to making a research breakthrough than completing an exercise. The hard part is often discovering the right idea and proving every step, not merely performing calculations.

The source does not provide exact durations, names, or difficulty rankings for the problems. It does support the broader contrast between ordinary exercises and long-running mathematical challenges. That contrast explains why the publication caused such a strong reaction and why mathematicians are reconsidering the scale of AI’s abilities.

The article presents AI as an active producer of the reported mathematical solutions. That means the system did more than calculate numbers or format an argument supplied entirely by a person. It apparently generated ideas or proof-like reasoning aimed at resolving difficult, longstanding questions.

The key mechanism is a chain of logical steps. An AI system can search patterns, suggest lemmas, combine known results, and write a candidate argument. In mathematics, however, fluent language is not enough. Every claim must follow from definitions, assumptions, and earlier justified steps. A plausible-looking answer can still contain a hidden error.

The source does not explain the exact Axiom model, prompting process, or amount of human assistance. It reports the solutions and the field’s reaction, not a complete technical audit. The forward question is whether AI can repeatedly produce independently checkable proofs, rather than occasional impressive results that require extensive human repair.

Reliable AI solutions would change what mathematicians spend time doing. Work that once required years of searching for a proof could become faster or partly automated. That could expand the number of problems researchers can investigate and make difficult mathematics more accessible to people without specialist training.

A mathematician might ask better questions, choose useful definitions, guide an AI system, and test its proposed argument. Humans would still need to judge whether a result matters and whether its concepts can be understood, reused, or connected to other areas. AI could provide the route, while mathematicians provide direction and interpretation.

The article says many leading mathematicians are experiencing an existential crisis. That reaction reflects uncertainty about professional identity, not proof that human work has ended. The outcome depends on reliability. If systems produce correct, readable, reusable proofs, mathematical labor may be transformed; if they mainly produce plausible errors, human expertise remains central.

The basic check is to inspect the argument line by line. A mathematician tests each definition, assumption, calculation, and inference, then checks whether the conclusion really follows. This is essential because an AI can produce confident prose containing a subtle gap, an invalid inference, or a citation that does not support the claim.

A stronger method is formal verification. The proposed proof can be translated into a precise language accepted by a proof assistant. Software then checks each permitted logical step against a trusted foundation. Independent mathematicians can also reproduce the argument, seek counterexamples, and compare it with established results. Agreement between separate checks increases confidence.

The source excerpt does not say how Axiom’s reported solutions were verified. It does establish why verification matters: the claims concern longstanding problems and have triggered major debate. Until solutions survive rigorous checking, they should be treated as reported results or promising arguments, not automatically as settled mathematics.

A mathematical proof starts with definitions, assumptions, and previously established results. It then connects them through valid logical steps until the desired conclusion follows. The proof is not just an explanation of why a result seems likely. It is a complete justification that another trained reader can inspect.

For example, checking many cases may suggest that a pattern always holds, but those examples do not cover every possible case. A proof must explain why no permitted case can fail. In formal mathematics, rules or proof-assistant software can check whether each inference follows from the stated foundations and earlier steps.

This distinction is central to the article’s AI discussion. An AI-generated solution may sound precise and contain useful ideas, yet still hide a gap. A verified proof can establish the result within its mathematical framework. The source does not give a formal definition, but its focus on AI solutions makes the difference between plausibility and proof especially important.