Face recognition

Face blindness test on faces the page draws

Study six faces for four seconds each, then pick each one out of three at a viewpoint you never saw — 20 scored choices in about seven minutes, free, with the whole result on the screen at the end and nothing behind an email address. It reports percent correct with a 95% interval and how often guessing would match it, then repeats the task upside down so you can see the difference. The faces are generated from thirteen geometric parameters because the validated test’s faces and its cutoff belong to its authors.

100% free
No signup
20 scored choices
Faces drawn by the page
Upright and inverted

20 choices, three faces each, about seven minutes

You will be shown 6 faces one at a time for 4 seconds each, then asked 12 times which of three faces you studied — with the studied one turned about twenty degrees from the view you saw. Then 4 new faces and 8 more choices, all of it upside down.

Three sample faces at the two viewpoints the test uses. No hair, no ears, no color, no shading and no background — thirteen geometric parameters and nothing else, so there is nothing to remember except the face.

Start with a 2-face warm-up (4 choices with the right answer shown afterwards). It is excluded from every figure and exists so the first scored trial is not also the first time you have seen the format.

Answer with 1, 2 and 3 or by tapping a face. Nothing is timed — take as long as you like over a choice.

How the recognition run works

Study, then choose, then do it again upside down.

Watch six faces go by, four seconds each
There is no way back to a face once it has gone, and there is nothing written on it — no hair, no ears, no color, no shading, no background. Try to take in the arrangement of the whole face rather than fixing on one feature, because at test the face is turned about twenty degrees and a feature you singled out head-on will have moved. The study clock only advances while this tab is the one in front of you, so a phone call does not cost you a face.
Pick the studied face out of three, twelve times
Each choice shows the face you studied alongside two others generated from the same parameter space at a controlled distance from it — close enough to be confusable, far enough to be separable. Answer with 1, 2 or 3, or tap a face; nothing is timed, so take as long as you want. The turn alternates direction between repeats, which stops a visitor learning one silhouette and matching against that for the rest of the block.
Do the last eight upside down, then read both proportions
Four new faces are studied inverted and tested inverted. Expect this half to feel much harder, and that is the measurement: upright faces are processed as configurations rather than as collections of parts, so inversion costs far more on faces than on any other object class. The result shows both proportions with their intervals, the gap between them, and the exact probability that guessing would have produced each.

Technical specifications

Trials	20 scored three-alternative choices — 12 upright over 6 studied identities and 8 inverted over 4 more — plus an optional 4-choice warm-up with the answer shown afterwards, which is excluded from every figure on the page
How a face is made	13 geometric parameters on 0-1: head width, jaw taper, chin length, eye separation, eye height, eye size, brow gap, brow tilt, nose length, nostril width, mouth width, mouth height, lip thickness. Flat gray fill, no hair, no ears, no shading, no background
How far apart the faces are	Two studied identities are never closer than 0.28 root-mean-square in that 13-dimensional space, and each distractor is held between 0.17 and 0.30 from its target. Two random faces average about 0.41 apart, so every wrong answer on this page is a near miss by construction
The change of viewpoint	Study is frontal; test is turned by a sine of 0.34, about twenty degrees. Each feature moves sideways by its depth in front of the face plane times that sine — the nose tip travels roughly five times as far as the eyes — and the outline is compressed by the cosine. No pixel arrangement from the study phase reappears at test
Response and what voids a choice	Digits 1-3 on the number row or the keypad, or a tap on the face itself. A key pressed during the 0.4-second cross between choices voids that choice, since there is no stimulus on screen for it to be an answer to. A choice interrupted by this tab going to the background is discarded and is not re-asked, because its study list is already behind you
What is reported	Percent correct for each block with a 95% Wilson interval, the exact one-sided binomial probability against a chance floor of one in three, the discard count, and the upright-minus-inverted difference in percentage points with an explicit warning about how wide that difference's own uncertainty is
Reference figure	Developmental prosopagnosia prevalence: 2-2.5% — Kennerknecht et al. (2006), First report of prevalence of non-syndromic hereditary prosopagnosia. Printed as background and never applied to a visitor's score. No control mean, no cutoff and no percentile appears anywhere on this page
What leaves the page	Nothing. Faces are drawn as inline SVG paths computed in this tab rather than fetched as images, so the run needs no network after the page loads and there is no image request that could record which faces you were shown

Frequently asked questions

Why are the faces drawings instead of photographs?

Because a photograph gives you a dozen ways to pass a face test without remembering a face. Hair, a hairline, an earring, a collar, stubble, one person's lighting, one camera's angle — encode any of those and you can be right every time while your face recognition sits out the whole exercise. The published instruments control this by cropping hair away and photographing each identity from several angles, and they still carry a second problem these drawings do not: a photograph has a race, and people are reliably better at recognizing faces from groups they have spent time around, which means a low score on a photograph-based test is partly a fact about whose faces you have seen. Thirteen numbers and a flat gray fill have none of that in them.

Why is the last third of the test upside down?

Because it is the control that tells you what kind of remembering you were doing. Turning a face upside down damages recognition far more than turning any other kind of object upside down, and the standard explanation is that upright faces are processed as whole configurations — the arrangement of the features relative to each other — while inverted ones fall back to being processed feature by feature. So somebody whose upright and inverted scores are close was probably working feature by feature in both halves, which is a real strategy and produces a real score, but not a score about face perception. The gap between your two numbers is the diagnostic thing here, and it is the reason eight trials are spent on a condition nobody enjoys.

I got every upright trial right. Am I a super-recognizer?

No, and the ceiling is the reason. Twelve trials cannot separate the top few percent of the population from the top half of it: a 12-out-of-12 run still carries a 95% interval running from roughly three-quarters to certainty, because twelve is a small number. The upper tail is a real thing — Russell, Duchaine & Nakayama (2009) described people who recognize faces from a single brief encounter years later, and identifying them takes long tests with hard items and a comparison group. What a clean sweep here does tell you is that whatever this task asked for, you have it comfortably.

Why is there no score out of 72 like the Cambridge test gives?

Because this page is not that test and a number shaped like its number would invite you to read it as one. The published instrument learns six faces and asks 72 three-alternative questions across three stages, the last with visual noise laid over the images, and it is scored against a control sample that was collected with its exact face set. Reproduce neither the faces nor the sample and the raw total means nothing on that scale; print the total anyway and somebody will compare it with a friend's real one. So the figure here is a proportion with an interval, in units that carry their own uncertainty and cannot be mistaken for anybody else's.

I recognize faces fine but never remember names. Is that the same thing?

It is a different failure and this test does not touch it. Recognizing a face is deciding that you have met this person before; retrieving their name is pulling an arbitrary label out of memory with no help from what you are looking at, and names are famously the hardest thing to retrieve about a person — people routinely recall someone's job, their city and where they met them while the name stays out of reach. Nothing here asks you to attach a label to anything: every face is anonymous, and the only question is whether you have seen this arrangement of features before. If names are the problem, this page is measuring something you are probably good at.

My eyesight is poor, or I am using a screen reader. What does my score mean?

It means whatever your eyes did, which is not what the page is trying to measure. This is a purely visual task with no audio route and no text route, and there is no honest way to build one — a description of a face in words is a description somebody wrote, so a test built on it would measure verbal memory for that description. If the faces are hard to resolve at your screen distance, or you cannot see them at all, the right conclusion is that this instrument does not apply to you rather than that you scored low. Nothing about the faces depends on color, so a color vision deficiency changes nothing here.

Can I retake it, and will the faces be the same?

Retake it as often as you like; every run generates a fresh set from a new seed, and the seed is printed with your result. That matters more here than on most tests, because the second run of a fixed face set is a memory test for the first run. What you should not do is average several runs into a better estimate of yourself — 20 trials taken three times is not 60 trials, since practice at the format, at the viewpoint change and at what these faces vary in all accumulate across runs and push later scores up for reasons that have nothing to do with faces.

Prosopagnosia, super-recognizers and the spread in between

Prosopagnosia is the inability to recognize faces, and the version people find their way to this page for is the developmental kind: present from childhood, with no injury behind it, in someone whose vision and memory are otherwise unremarkable. It is commoner than its obscurity suggests — 2-2.5% of the population by Kennerknecht et al. (2006), which puts one in every large office — and the usual account from people who have it is not that faces look blank but that they cannot be held: the colleague is recognizable in the corridor where colleagues are, and unrecognizable in a supermarket. What that estimate cannot do is settle anything about one person. Estimates move with the criteria used, and looser thresholds put the figure higher. Self-report on its own substantially over-identifies: many people who find faces hard are within the ordinary range, and a page must report what the visitor scored on the matching task rather than converting a questionnaire into this prevalence. Face recognition also varies enormously among people with no condition at all, which is the more useful thing to tell someone who scored low.

The other end of the same distribution is what makes a single score so easy to misread. The upper tail has been described as a group in its own right — Russell, Duchaine & Nakayama (2009), Super-recognizers: people with extraordinary face recognition ability, Psychonomic Bulletin & Review — on the strength of four people who could place a face years after one brief encounter, which is a demonstration that the tail exists and not an estimate of how many are in it. Between the two ends sits a range of ordinary performance so wide that it is one of the most striking individual differences in cognitive psychology, and any test that hands you a number without showing you that range has told you something true and useless. This is also where the copyright rule and the honesty rule turn out to be the same rule. The test is distributed under the authors' control with a fixed face set that is not ours to reproduce, and this file could not verify the control mean and standard deviation to the digit. Both reasons independently rule it out, and the registry blurb for this page promising a score 'against published norms' overstates what is available — report that upward rather than inventing the norms to match it. The instrument that does carry a defensible cutoff is Duchaine & Nakayama (2006), The Cambridge Face Memory Test: results for neurologically intact individuals and an investigation of its validity using inverted face stimuli and prosopagnosic participants, Neuropsychologia — and this page is not it. Neither is it the self-report questionnaire that often accompanies it: The instrument is public and a suggested cutoff exists, but this file could not confirm the threshold value with enough confidence to print it, and a self-report cutoff printed one point wrong sorts people into the wrong side of a label they will repeat.

Which leaves the question of what a browser can honestly measure, and the answer is a proportion. Twenty three-alternative choices, a chance floor of one in three, a Wilson interval around whatever you scored, and the exact probability that guessing would have done as well — those four numbers are checkable, they carry their own error, and none of them pretends to be a rank. The inverted block is there so the proportion means something: a score that survives inversion intact was a shape-matching score, and a score that collapses was a face score. If the interesting part for you was the study phase rather than the choices — what happened in your head while a face was on screen — the aphantasia test asks whether you can call an image back at all, and a surprising number of people arriving here find that is the question they actually had. Two neighbors ask the same “is it only me?” question about other channels: the misophonia test for specific sounds, and the synesthesia test for cross-channel experiences that other people do not report.

Own-race bias in face memory

≈ 1.4 times the odds of a correct identification for a face from one’s own racial group, pooled across 39 research articles pooled.

Meissner & Brigham (2001), Thirty years of investigating the own-race bias in memory for faces: a meta-analytic review, Psychology, Public Policy, and Law

Quoted here as an odds ratio for correctly identifying a previously seen face, and it is background rather than anything measured on this page: the faces here have no racial features to be biased about. It is the reason a photograph-based face test scores you partly on whose faces you have spent your life around, and the reason a low score on one of those tests is ambiguous in a way a low score here is not.

One thing this page does not measure, and cannot: how quickly you answered. Every choice here is untimed on purpose, because face recognition is not a speed test and putting a clock on it would reward a strategy rather than an ability. Nothing on this page is reported in milliseconds, so the display and input latency that sets a floor under the speeded tasks elsewhere on this site never enters these figures.

This is a measurement exercise, not a clinical assessment. It reports what you did on this page against a stated reference and nothing more — it cannot establish prosopagnosia. Only a qualified professional, working with more than a browser, can make that judgment.

Where the faces come from and where your choices stay

Every number on this page is worked out by JavaScript running in the tab you are reading it in. Your answers, your reaction times and your score are never uploaded, logged or kept — which is also why the test carries on working after you disconnect from the network, and why nothing here can be held back behind an email address.

The faces are not images and were never uploaded anywhere: each one is a handful of SVG path strings computed from thirteen numbers while you watch, so there is no image request that could log which identities you saw or how long you looked. Your twenty choices live in an array in this tab, are summed once into two proportions, and disappear with the tab.