The Unlit Proof
Truth Arrives Before Meaning
Shanaka Anslem Perera
30 June 2026
The Capture
On the twentieth of May, 2026, a conjecture that had shaped a central question in geometry since 1946 was overturned, and the strange part was not that a machine had overturned it. The strange part came afterward. Nine of the most accomplished mathematicians alive, a group that included a Fields medalist, sat down together and wrote what they openly called a short, digested, human-verified version of the machine’s argument. They were not only checking the proof for errors, though they did that too. They were translating a truth that already existed into something a human mind could hold. The truth had arrived first. The understanding came second, as a separate and later act of labor.
This is a small thing and an enormous thing at once. The problem itself, Paul Erdős’s unit distance conjecture, asks a question a child could picture. Scatter some dots on a page. Count the pairs that sit exactly one centimeter apart. As you add more dots, how fast can that count grow? For eighty years the best mathematical intuition said that grid arrangements were close to the limit. An internal reasoning model at OpenAI found that intuition was wrong, and it found why, by reaching into a corner of number theory that few had expected to bear on the question. The model did not merely assist. It produced the decisive construction, the specific idea the proof turns on, and the construction held up, and humans then spent real effort making it legible to themselves.
We are used to the opposite order. In the classical ideal of science, a result became knowledge only when some person or community understood it well enough to vouch for it. Comprehension was not a bonus that arrived after the discovery. Comprehension was the discovery, or at least its final and necessary stage, the moment the thing was truly known rather than merely true. What happened in May was that this final stage detached and floated free. A mathematical truth sat in the world, checked by leading mathematicians and reliable enough to build on, and the light of full human understanding had not yet fallen on it.
Call it an unlit proof. The phrase will do more work than it seems to, because the condition it names is spreading, and it is not spreading evenly. In some regions of science, machines now generate and confirm results faster than any community can absorb them, and the backlog is a pile of truths waiting to be understood. In other regions, machines generate candidate results faster than anyone can check whether they are true at all, and the backlog is a pile of claims waiting to be trusted. These two situations look superficially alike and are in fact opposites, and telling them apart is the beginning of seeing what artificial intelligence is actually doing to the structure of knowledge.
It is not accelerating science evenly. It is reorganizing science around a hidden landscape, a topography of how cheaply and quickly a given kind of truth can be checked. Where checking is cheap, the machines race ahead and leave human understanding behind. Where checking is expensive, the machines flood the field with possibilities that no one can yet confirm. The shape of that landscape, and the question of whether the slow human work of understanding and validation can be made to keep pace with the fast machine work of generation, is the subject of this essay. The short version is the one the mathematicians lived in May. Truth is beginning to arrive before meaning. The question is what we do in the gap.
The Foundation
To see why this is new, it helps to be precise about what was old. For most of its history, the practice of science fused four distinct acts into one, performed by the same mind in a single continuous motion. There was generation, the act of producing a candidate, a conjecture, a hypothesis, a design. There was verification, the act of checking that the candidate was actually correct. There was explanation, the act of understanding why it was correct and weaving it into the rest of what we know. And there was abduction, the oldest and most mysterious act, the judgment of which question was worth asking in the first place, which anomaly deserved attention, where the frontier actually lay.
The philosopher Charles Sanders Peirce gave abduction its name and treated it as the engine of discovery, the leap that proposes an explanation worth testing. What matters here is that these four acts were never really separate. A mathematician who proved a theorem had chosen the problem, generated the approach, checked every step, and understood the result, and these were not four people or four moments but one person thinking. The unity of the four acts is, more or less, what we have always meant by the word understanding. To understand something was to have done all four with respect to it.
The first crack in that unity is older than the current wave of machines, and it is worth remembering because it tells us the shape of what is coming. In 1976, Kenneth Appel and Wolfgang Haken proved the four color theorem, the claim that any map can be colored with four colors so that no two bordering regions share a color. Their proof was unlike any before it. It reduced the problem to a large number of cases and then had a computer grind through them, a labor no human could perform or check by hand. The philosopher Thomas Tymoczko argued that this forced mathematics to redefine its own central concept. A proof had always been something a competent person could survey, follow, and hold in mind. Here was a proof that no one could survey. For the first time in a way the discipline could not ignore, verification had detached from comprehension, and some mathematicians never fully accepted it.
The classification of finite simple groups, completed across thousands of pages by many hands over decades, made a related point in a different way. It is a theorem too vast for any single mind to contain. The knowledge it represents lives not in any person but in a distributed community and its libraries. And in the summer of 2022, a third tremor arrived from biology, when the AlphaFold database expanded to more than two hundred million predicted protein structures, covering very nearly every protein known to science, an increase of roughly two hundredfold in a matter of months. No laboratory community on earth could validate or even examine that many predictions. The structures arrived as a flood, far faster than the science could metabolize them into understanding.
