Can Exams Measure Intelligence? Click Here To Know
Share
Results day has a strange way of turning a number into a verdict. A student opens an envelope, sees a mark, and in that instant a whole year of effort gets compressed into a single figure that somehow feels like it’s saying something about who they are, not just what they did on one morning in one exam hall. High scores get celebrated as if they confirm something. Low ones get treated the same way, in reverse. The question worth actually asking is whether that reaction has ever been justified — does an exam measure intelligence, or does it just measure performance on a given day?
What an Exam Is Actually Built to Test
Exams are, by design, fairly narrow instruments. They test how well someone can recall information, apply it under time pressure, and organise an answer within a fixed window — memory, preparation, and composure under a clock, essentially. Those are genuinely useful academic skills, and there’s a reason education systems have relied on them for over a century: they’re standardised, they’re comparable across large groups of students, and they’re relatively hard to game through favouritism.
But that same rigidity is where the criticism starts. Standardised tests tend to reward a fairly specific profile — strong verbal reasoning, comfort with maths, the ability to perform well under acute pressure — while having almost nothing to say about a student who thinks brilliantly but freezes in an exam hall, or one whose intelligence shows up in how they build things, read people, or solve problems that don’t come with four multiple-choice options. Psychologists have been raising this exact critique for decades. Howard Gardner’s theory of multiple intelligences — now a fixture in education debates — argues that intelligence was never one measurable quantity to begin with, but a spread of distinct capacities: linguistic, logical-mathematical, spatial, musical, interpersonal, and several others, most of which a written exam was never built to detect.
There’s also a more uncomfortable finding sitting underneath all this: exam performance can be surprisingly unstable over time. A widely cited Brookings Institution analysis found that a large share of year-to-year improvements in test scores — somewhere between 50 and 80% in the study — turned out to be temporary statistical noise rather than genuine gains in learning. If a score can swing that much without any real change in a student’s underlying ability, treating it as a fixed measure of intelligence starts to look shaky on its own terms.
Test anxiety adds another wrinkle entirely. A student who understands a subject deeply but performs poorly under acute exam pressure isn’t lacking intelligence — they’re having a very normal physiological response to stress, one that has nothing to do with whether they’ve actually learned the material. It’s part of why, since 2020, nearly a thousand US colleges and universities have moved to test-optional or test-blind admissions, no longer requiring SAT or ACT scores at all — an acknowledgment, at scale, that a single exam score was never telling admissions officers the whole story.
Worth noting, in fairness to the exam’s defenders: it isn’t a one-sided picture. Research from Opportunity Insights, tracking outcomes at highly selective US colleges, found that SAT and ACT scores do meaningfully predict college grades and other academic outcomes — better, in some cases, than high school GPA alone. So the honest position isn’t that exams measure nothing real. It’s that what they measure well — a fairly specific kind of academic aptitude — is narrower than what people mean when they say “intelligence.”
Now Machines Are Taking the Same Exams
The rise of AI has made this question harder to dodge, in a genuinely new way. Large language models are now being tested directly against real, high-stakes human exams — researchers evaluating one AI system on South Korea’s notoriously difficult national maths exam, the CSAT, found that with enough computational reasoning applied, the model’s score climbed close to a perfect mark. Machines that have never taught themselves anything from lived experience are increasingly capable of matching or beating students on exactly the kind of test we’ve long treated as a proxy for human intelligence.
That should make anyone pause before equating “correct answer under time pressure” with genuine understanding. An AI system doesn’t understand a concept the way a person does — it isn’t reasoning from lived experience or genuine comprehension. It’s identifying statistical patterns across enormous amounts of training data and producing the most probable correct response. If a system built entirely on pattern-matching can now clear the same bar we use to certify human intelligence, that bar was probably measuring something narrower than intelligence all along — call it pattern-application under exam conditions, which is real and useful, but isn’t the whole of what a mind can do.
What a Score Can’t Capture
Human intelligence does things no exam script and no AI model currently replicates convincingly: thinking creatively when there’s no template to follow, reading another person’s emotional state accurately, asking a question nobody else in the room thought to ask, making a genuinely difficult ethical call, imagining something that doesn’t exist yet. Two students can walk out of the same exam with an identical mark and go on to live completely different lives — one building a company, one running a lab, one making art that moves people — because what actually drove their outcomes wasn’t the score itself but everything a score doesn’t measure: curiosity, resilience, the ability to communicate, emotional intelligence, and a knack for solving problems that don’t come with a marking scheme.
So What’s the Exam Actually For?
None of this makes exams pointless. They’re still a genuinely useful way to track academic progress, enforce a baseline of discipline in how students study, and give large education systems a comparable measure across huge numbers of students, imperfect as that measure is. The mistake isn’t using exams — it’s mistaking them for something they were never built to be. The same goes for AI: extraordinarily useful for narrow, well-defined tasks, and not a measure of understanding in any deeper sense either.
Real intelligence was never going to fit inside a mark out of a hundred. It shows up in how someone thinks, adapts, creates, empathises, and eventually does something useful with what they’ve learned — none of which shows up on a scorecard, human or algorithmic. Knowledge can be tested. Data can be calculated. Intelligence, it turns out, is the part that keeps slipping past both.
