The Answer Without Understanding

 

Can we possess knowledge before we comprehend its meaning?

In October 2026, Quanta Magazine reported on a remarkable mathematical event: OpenAI had released hundreds of manuscripts addressing hundreds of open questions. Some of the proofs had been translated into Lean, a language in which a computer can check whether each logical step follows from its premises. Mathematicians began the slow work of assessing what had actually been proved, what was important, and what might be useful.

Within days, three manuscripts were withdrawn after a sign error was identified. Even at the frontier of artificial intelligence, mathematics still demanded its ancient discipline: show your work.

The news raises an unsettling question. What does it mean to possess an answer we cannot yet explain?

The theorem in the margin

In 1637, Pierre de Fermat famously claimed to have found a proof of a theorem for which his margin was too narrow. Generations of mathematicians pursued it. Andrew Wiles finally presented a proof in 1993, but a gap was found; a corrected proof followed in 1994 and was published in 1995.

For centuries, Fermat’s claim was not established knowledge. Yet the unanswered question was productive. It inspired techniques, collaborations, failures, and new mathematics. Its value did not wait until the final proof.

A solution can be the end of one problem and the beginning of an entire intellectual landscape.

A proof is not an explanation

Proof and understanding are related, but they are not identical.

A verified proof establishes that a conclusion follows from stated assumptions. It need not reveal why the result is surprising, which ideas made it possible, or where those ideas might lead.

Lean is a formidable verifier, not an oracle of significance. It can check a formal argument, but humans must still ask whether the formal statement captures the intended question and whether the result changes our understanding.

Now suppose a machine produces ten million correct steps. The proof checks out, but nobody can see the larger pattern. We have gained certainty about a proposition. Have we gained understanding?

Perhaps some. But not necessarily all the understanding the proof could offer.

A tool is worth more than a trophy

Imagine two proofs of the same theorem. One is immense and laborious. The other reveals a mathematical connection that opens a new line of research.

Their conclusions are identical. Their gifts are not.

The greater achievement may lie in the new method: a tool that other mathematicians and physicists can learn, teach, and carry into questions the discoverer never imagined.

If an AI creates such a method, why deny that it has contributed to mathematical knowledge? The validity of the result does not depend on whether the author is flesh or silicon. Its usefulness does not depend on whether we can reconstruct the discoverer’s private moment of insight.

Yet attribution, responsibility, and publication norms remain unsettled. Producing manuscripts faster than specialists can assess them is not the same as advancing shared understanding at the same rate.

Mathematics is a collective undertaking. Its proofs must be inspectable; its ideas must travel.

Knowledge that outlives its discoverer

Perhaps we have been looking for understanding in the wrong place.

The machine discovers a relationship. One mathematician verifies it. Another finds a clearer proof. A teacher makes it comprehensible. A physicist puts it to work. A student glimpses an application none of them anticipated.

Where is the knowledge?

Not imprisoned in the machine, nor owned by the first person who understands it. It moves, changes form, and grows as minds engage with it.

A proof may establish a truth. But understanding begins to flourish when someone can explain it to another—and that other person can do something new with it.

The question, then, is not simply whether a machine can provide an answer. It is whether the answer can become part of a living body of understanding.

And that leaves us with a second, stranger question: Must an intelligence experience the world to understand it?

Inspired by Quanta Magazine’s October 2026 reporting on AI and mathematics. Sources: Quanta’s report and follow-up; OpenAI’s October 6 announcement.