Why the ability to perceive that A is to B as C is to D may be the closest thing we can measure to general intelligence itself.
Consider the humblest-looking item on any aptitude test: A is to B as C is to —? It can look like a parlor trick, a puzzle for crossword hobbyists. It is, in fact, something close to a direct readout of the human intellect. What follows is a claim both ancient and, lately, very well measured: that analogy — the perception that one relation mirrors another — is not one mental trick among many, but near the center of thought itself.
When Charles Spearman gave us g — the general factor that emerges, stubbornly, from the positive correlations among all mental tests — he did not describe it as a mysterious quantity of brain-stuff. He described it as an operation: the eduction of relations and correlates. Given two ideas, perceive the relation between them; given an idea and a relation, generate the correlate. Read that twice, because it is the whole game. Given an idea and a relation, generate the correlate is not a description of intelligence in general — it is a description of an analogy. Reaching for the deepest layer of cognition he could name, Spearman named the very thing a verbal analogy asks you to do.
If Spearman was right, the best measures of g should be, in effect, analogy tests. They are. The instrument psychometricians reach for to estimate g with a single test is Raven's Progressive Matrices — and Raven's is a test of visual analogies: this figure transforms into that one as this third figure transforms into — which? In a now-classic 1990 paper, Carpenter, Just, and Shell asked what the Raven test actually measures, and answered, after modeling how people solve it: the processes that separate high scorers from low scorers are "the ability to induce abstract relations." We have been measuring intelligence with analogies all along; we simply did not always say so.
The modern picture tightens the knot. When Jan-Eric Gustafsson modeled a broad battery of tests in 1984, he found that fluid intelligence is statistically almost indistinguishable from g itself — the constraint forcing them to be identical did not even meaningfully worsen the model's fit — and that the inductive-reasoning factor (relational reasoning) loaded on fluid intelligence at a level "approaching unity." So follow the chain, each link load-bearing: analogy is the model case of relational reasoning; relational reasoning is the heart of fluid intelligence; fluid intelligence is nearly identical to g. Validated against other cognitive measures, a pure verbal-analogy test (the Miller Analogies Test) has correlated about .75 with general reasoning and .88 with verbal ability. Analogy is not adjacent to intelligence. It is intelligence, caught in the act.
Some go further. The cognitive scientist Douglas Hofstadter — with Emmanuel Sander, in Surfaces and Essences: Analogy as the Fuel and Fire of Thinking — argues that analogy is not a faculty the mind possesses but the very activity of which mind consists. Every concept, on this view, is a worn groove of remembered analogies; every act of recognition — that is a chair, that is a betrayal, that is the same mistake as last year — is the silent mapping of a new situation onto an old structure. A hub, with memory, language, and judgment arrayed around it as its expressions.
Intellectual honesty requires the rival account, and it is a strong one. Its champions — in a 1990 paper bluntly titled "Reasoning ability is (little more than) working-memory capacity?!" — argue that the deeper substrate is working memory: the sheer capacity to hold several elements in mind and operate on them at once. Their correlations between working memory and reasoning ran as high as .80 to .90 at the latent level. On this view, analogy is not the engine but the exhaust — the visible output of a more fundamental ability. The magnitude is contested (a large meta-analysis put the working-memory–g correlation nearer .48, "related but distinct"), but the challenge stands. So the deepest question remains open: is analogy the architecture of mind, or its most visible façade?
The data underdetermine the metaphysics, so let us separate what can be claimed from what cannot. Established, robustly: analogical reasoning indexes general intelligence about as well as any single cognitive operation we can name. The premier measure of g is an analogy test; relational reasoning is nearly identical to fluid intelligence, which is nearly identical to g. To be good at analogies is, measurably, to be generally intelligent. Not established: that analogy generates the other faculties. That remains a wager — but a good one. When the founder of intelligence testing, looking for the floor of cognition, found "the eduction of relations"; when the best single test we have is a wall of analogies; when the most g-saturated talent a person can display is the talent for seeing that this is to that as the other is to a fourth thing — the convergence is too clean to be coincidence.
There is a deflationary reading worth answering, because it is the one a skeptic reaches for: surely a verbal analogy test is merely a test of vocabulary — a measure of how many words one has happened to accumulate. A fund of knowledge. Erudition, not intelligence. It is comfortable, and it is wrong. Vocabulary is not a sideshow to general intelligence; it is its single best stage. In the standardization sample of the Wechsler Adult Intelligence Scale, Vocabulary carries the highest g-loading of any subtest — .87 — outscoring the matrix puzzles that look far more like pure reasoning. Decompose a vocabulary score and 55 to 61 percent of its variance is g, while only a sixth to a third is knowledge-specific. A vocabulary test is mostly an intelligence test in the costume of a knowledge test; crystallized ability as a whole loads on g at roughly .89. Cattell's investment theory explains why: crystallized knowledge is fluid intelligence invested, over years, into learning — the sediment a powerful current leaves behind. The objection survives in one form only: narrow, specialized knowledge is weakly tied to g (correlations near .4), while general verbal knowledge is the high-g signature (near .6 and above). Encyclopedic command of one trivial domain proves little; a wide, deep, general fund of knowledge proves a great deal. A high score on a verbal analogy test, then, is not a receipt for a well-stocked memory — it is, like the matrices, another reading of the same dial.
To build a contest scored on analogy — to rank minds by their command of A is to B as C is to D — is, knowingly, to build an instrument that measures something real and stable: not trivia, not luck, but the mind performing its most characteristic act. To see likeness across difference is not a trick at the edge of intelligence. It may be the center around which the rest revolves. Which leaves only one question, and it is not a theoretical one: how good is yours?
A publication of The Digital Areopagus.
← Return to The Analogy Arena