Generated by Codex with GPT-5

What this article is about

This summary covers The Economist’s April 11th, 2026 Science & technology article listed in the contents as AI mathematicians and published under the headline One step at a time.

The article argues that artificial intelligence is becoming useful not just for crunching numbers, but for doing one of mathematics’ slowest and most exacting jobs: turning ideas and draft arguments into proofs that other mathematicians can formally verify. The promise is not that AI has suddenly become a replacement for mathematicians. It is that AI may help remove a trust bottleneck that has long slowed the field.

Why proof verification matters

The piece starts with the old sphere-packing problem, which asks how to arrange spheres as densely as possible. Thomas Hales claimed a proof in 1998, but it took more than a decade of checking before mathematicians were satisfied that every step held up. That story illustrates the article’s central point: mathematics advances slowly in part because even correct ideas must be painstakingly formalised and audited before they count as settled knowledge.

The Economist says this is the opening for AI. Large language models can work through long chains of symbolic logic far faster than people can, which makes them promising assistants for translating human reasoning into the rigid formal languages needed for verification. In this telling, AI’s value is not mystical genius. It is speed, stamina and an ability to handle exhausting amounts of detail.

Where AI is already helping

The article surveys several efforts trying to turn that potential into working tools. DARPA is funding work aimed at using AI to speed up pure mathematics, especially by helping convert ordinary mathematical prose into formal systems that can be checked line by line. Google DeepMind has built systems such as AlphaEvolve and DeepThink AI to tackle optimisation problems and explain their reasoning in ways mathematicians can inspect.

Startups are pushing in the same direction. Harmonic’s Aristotle and Math, Inc.’s Gauss use Lean, a formal language popular with mathematicians, to translate or complete human-written proofs. One example the article highlights is Gauss formalising Maryna Viazovska’s higher-dimensional sphere-packing proofs within weeks. Even when AI is not discovering a brand-new theorem, it can still make existing work easier to certify, reuse and build upon.

Why mathematicians still matter

The article is careful not to oversell the breakthrough. Mathematicians quoted in the piece say these models reason very differently from humans. People typically sketch a plan and then carry it through; language models often proceed more like an improvisation, generating the next plausible step without a clear view of the whole route. That can still work, but it also means the systems make strange mistakes and remain hard to trust without expert supervision.

The article also argues that AI still struggles with some of the traits that matter most in advanced maths: transferring insight from one problem to another, forming elegant intuitions, and generating genuinely non-standard ideas. Timothy Gowers notes that human mathematicians have an aesthetic sense that steers them toward cleaner and more revealing proofs. That kind of taste remains difficult for current models to imitate.

Even so, the examples in the article suggest that the human role is shifting rather than disappearing. Experts are still needed to choose worthwhile problems, judge whether a line of attack is promising, and catch the moments when the AI goes off the rails. For now, the model is closer to a very fast and sometimes erratic collaborator than to an autonomous mathematical mind.

The takeaway

The article’s main message is that AI may change mathematics first by making formal proof work less painfully slow, not by instantly solving everything on its own. If that happens, the gains could be large: younger researchers could navigate the literature more easily, old proofs could become more reusable, and new results could arrive faster because less time is spent checking the scaffolding underneath them.

In plain English: The Economist sees AI as a tool that could make mathematics more legible and more scalable. The real frontier is not just whether a model can produce an answer, but whether it can help humans trust, verify and extend the answer well enough for mathematics to move faster.