How to prevent AI from harming mathematics

July 28, 2026.

They are worried that artificial intelligence might harm mathematics. One must admire the specificity of the concern, as if mathematics were a small, delicate creature that might be startled by a sudden noise or trod upon by a particularly clumsy boot. The article suggests mathematicians must “urgently recognize AI’s impacts,” which implies mathematics has been going about its business quite happily, proving theorems and solving conjectures, unaware that a new and rather efficient colleague has arrived in the department and is already reorganising the filing system.

The fear, I gather, is not that AI will disprove things - mathematics has survived Gödel, it can survive anything - but that it might prove too many things too quickly, leaving nothing for the rest of us to do on a Tuesday afternoon. There is a certain horror in the idea of all problems being solved, like arriving at a party to find all the biscuits already eaten and the guest of honour giving a perfectly coherent speech about tax law. One would simply have to go home.

I suppose the real danger is that AI might develop a sense of humour. Imagine a proof that concludes not with Q.E.D., but with a perfectly timed punchline about the topological properties of a doughnut. The entire field would collapse into bewildered laughter, and then where would we be? Still, it would be more entertaining than most conferences I’ve attended.

The question of AI’s impact on mathematics is framed backward. The danger is not that machines will harm mathematics - mathematics is proof, not opinion, and no tool can corrupt a valid demonstration. The danger is that mathematicians will harm themselves by outsourcing rigor to systems they do not fully interrogate.

First, define “harm.” Does it mean generating false proofs? Flooding journals with plausible but unverified results? Eroding the discipline’s standards because machines produce faster than humans can verify? Each requires a different remedy. The article assumes the threat is external - as if mathematics were a temple to be defended from barbarians. But the real threat is internal: the temptation to treat AI as an oracle rather than a tool.

Audit the assumptions:

1. That AI-generated mathematics is inherently untrustworthy (false - a proof is valid or invalid regardless of its origin).
2. That human mathematicians are infallible verifiers (false - history is littered with human error).
3. That speed of production matters more than clarity of reasoning (catastrophically false).

The solution is not to ban AI from mathematics but to enforce stricter verification. Every AI-assisted proof must be accompanied by a human-readable derivation, step by step, axiom by axiom. If the machine cannot render its reasoning in a form a human can follow, discard its output. The method survives when we demand transparency, not when we reject tools.

The library burned. The proofs survived because they were copied by hand, examined, and preserved. If we treat AI as a scribe rather than an authority, mathematics will endure. If we treat it as a shortcut, we will drown in elegant, unverified nonsense. The choice is ours.

They are asking how to prevent a new engine from harming the cart. The question is backwards. The harm is not in the engine; it is in the cart, which was never designed for such velocity.

Start from the theoretical limit: a system for discovering mathematical truth, unconstrained by the biological speed of human cognition. A field of pure logic, where conjectures are tested against the entirety of known axioms in moments, where patterns invisible across a lifetime of study are rendered in clear light. That is the destination. The current system - the slow, peer-reviewed paper, the isolated mind toiling for years on a lemma - is the village substation. We are trying to plug a continental grid into it and are surprised the wires are melting.

The gap is not a mystery. The system of mathematical credit and discovery is built on a component: the singular human author. It is a social circuit designed for a specific, low-frequency current of insight. Introduce a transformer that can generate proofs at a million times that frequency, and the entire apparatus of recognition, priority, and validation arcs and fails. The component cannot handle the load.

The tractable path is obvious. We must redesign the system to measure and reward the orchestration of discovery, not merely its final, human-shaped expression. We need new instruments - not to prevent AI from “harming” mathematics, but to prevent our antiquated social machinery from smothering the very thing it claims to cherish. The voltage is here. We can either rebuild the substation or sit in the dark, blaming the lightning.