Filtering under load

Selective attention test scored on your search slope

Say whether a tilted solid bar is somewhere in a scatter of gray bars, over displays holding 4, 8, 16 or 24 of them, and the run reports how many milliseconds each extra bar cost you. It runs the same question twice: once where the target is the only tilted thing on screen, and once where every other bar shares one of its two properties. The first usually comes back near flat and the second does not, and the gap between your own two slopes is the measurement — free, on screen immediately, with nothing stored and no account asked for. Nothing in the task uses color, so there is no second route needed around one.

100% free
No signup
128 displays
4 to 24 items
Slope in ms per item

A scatter of gray bars appears. One of them may be a tilted, solid bar; press J if it is there and F if it is not. The number of bars changes from trial to trial — 4, 8, 16, 24 — and how much that number costs you is the entire measurement.

The two searches, spelled out

Target: a bar both tilted and solid, identical in every trial and in both blocks
Others: upright solid bars in the one-feature block; a mixture of tilted hollow and upright solid in the two-feature block
Sizes: 4, 8, 16, 24 bars, drawn in a random order with no more than 3 of one size running together
Presence: half the displays hold the target, half do not, 8 of each per size per block
Blocks: two, run feature then conjunction for this run, drawn from the seed
Answer: J for there, F for not there, or the two buttons
Window: 5 seconds — long, because a twenty-four-bar array honestly takes a while
Field: 480 by 300 units, 8 by 5 cells, each bar jittered up to 13 units off its cell center

The bars are jittered off their grid on purpose. On a perfect lattice the eye can sweep row by row at a fixed rhythm, and the slope that comes out is a scanning rate rather than a search — the same reason the printed versions of this task scatter their items.

The sample puts 2 trials behind each point instead of 8, so its four medians are drawn from two numbers each and the line through them will move a lot between runs. It shows you the task; it does not give you a slope worth quoting.

How to take the selective attention test

One target, two kinds of clutter, and four display sizes shuffled through both.

Learn the target: tilted and solid, both at once
The bar you are looking for leans at 45 degrees and is filled in. A bar that leans but is only an outline is not it, and neither is a filled bar standing upright. That distinction is the whole design — in one block nothing else leans, so the target jumps out, and in the other every distractor matches it on exactly one of the two properties, so neither property alone finds it. Press J when you can see it and F when you are sure it is not there; the two buttons below the field do the same, and the run counts how many answers came that way.
Answer at the moment you know, and guess rather than stare
Every display holds a target on half the trials and nothing on the other half, so a run of empty screens is normal and is not a sign you are missing them. The window is five seconds, which is generous for twenty-four bars and impossible to hit if you are checking twice. A display you gave up staring at contributes a very long time to its set-size point and bends the line through it, so an honest guess is worth more to the measurement than a careful timeout.
Read the two slopes against each other, not against a table
The result is four medians per condition — one at each display size — and the least-squares line through them, printed as milliseconds per item with its intercept beside it. A slope near zero means the display size did not matter and the target was found however much surrounded it. A slope of twenty or thirty means each added bar cost about what checking one more bar would cost. The number that belongs to you rather than to your hardware is the difference between the two, because both were measured on the same screen with the same hand minutes apart.

Technical specifications

Trials	128 scored displays — two blocks of 64, each block being 4 display sizes × target present or absent × 8 repetitions. A 32-trial sample exists for checking legibility and puts only 2 trials behind each median
Display sizes	4, 8, 16 and 24 bars, shuffled with no more than 3 displays of one size in a row. Four points is the minimum a straight line can be fitted through and still be checked against its own scatter
Stimulus	Rounded bars 9 by 32 units, upright or leaning 45 degrees, either filled or outlined at 2.5 units of stroke, all in one pale gray on a dark field. No hue is used at any point in the task
Layout	A 480 by 300 field divided into 8 columns and 5 rows, one bar per occupied cell, each jittered up to 13 units off its cell center so the array cannot be swept as a lattice
Blocks	Two, run in an order drawn from the run seed rather than fixed, with a break and an explanation between them. Blocked rather than interleaved: mixing the two searches trial by trial adds a switch cost that lands unevenly across display sizes
Response window	5,000 ms from the array painting, with 400 ms of fixation before it and 400 ms of blank after the answer. Long deliberately — a 24-bar conjunction display genuinely takes over a second
How the slope is fitted	The median of the correct trials at each display size, with the outlier rule applied inside each cell rather than across the pooled run, then ordinary least squares of those four medians on display size
Reference figures	One, and it is a sort rather than a rank: slopes under about 10 ms/item are conventionally called efficient and over about 25 inefficient (Wolfe, 1998), with the same source noting the boundaries are conventions and the distribution has no gap in it

Frequently asked questions

Why does the second block feel so much slower than the first?

Because a target defined by one feature and a target defined by two are found by different machinery. When nothing else on screen leans, the tilt is available everywhere at once and the display size stops mattering — this is why the flat slope in the first block is the normal result rather than a good one. When half the distractors lean and the other half are filled, neither property on its own marks the target out, so the two have to be combined at each candidate in turn. That combining is what costs time per item, and it is the reason the same person produces two very different slopes ten minutes apart.

Where is the gorilla video? Every other page about this has one.

Deliberately absent, and the reason is in the study itself: the effect can only be measured on somebody who has not heard of it. Roughly half of observers counting basketball passes fail to notice a person in a gorilla suit walk through the scene, but a visitor who arrives already knowing that is no longer an observer of the phenomenon, they are an audience for it. Hosting the clip would produce a page where almost everybody spots the gorilla and concludes something flattering about themselves. The search task here can be repeated as often as you like and still measures the same thing, which is the property that made it worth building instead.

My conjunction slope came out around 25 ms per item. Is that bad?

It is the ordinary result and it is not a score. Combining two properties is slow in everybody, which is precisely why the paradigm is used to demonstrate the difference between the two kinds of search rather than to sort people. The number is also a property of this display as much as of you: bars packed more tightly, a smaller tilt difference, or more items on screen would all move it, so a slope measured here is not comparable to a slope from a paper or from another site. The comparison the page offers is your own two slopes against each other, and even that is one session.

Why does the page make me answer when nothing is there?

Because the trials with no target are where the shape of the search shows up most clearly. If you are working through the items one at a time, you find a present target after checking about half of them on average, but you cannot declare an absent one until you have checked all of them — so the absent slope comes out at roughly twice the present slope. The page prints that ratio, and a value near two is the signature of a search that was worked through rather than seen at once. Without absent trials there would also be nothing stopping a visitor from pressing present on everything.

What is the intercept for, and why is it so much bigger than the slope?

The intercept is what a display of zero items would cost: registering that a trial has begun, deciding, and moving a finger, plus every millisecond your monitor and keyboard add on the way. It is typically four to six hundred milliseconds against a slope of tens, and it is printed because a slope quoted alone hides how much of each response never varied with the display at all. It is also where the on-screen buttons load their cost, if you used them — aiming a pointer adds a constant per trial, which lifts the intercept and mostly leaves the slope alone.

Would color-blindness affect my result?

No, because there is no color in the task to affect. The textbook version of this experiment separates its features by hue — the classic demonstration searches for a green T among green Xs and brown Ts — and a visitor with a red-green deficiency taking that version has their color vision measured under the name of their attention. Using orientation and fill instead gives the same conjunction structure in a single gray. It is worth stating because it is the kind of choice that is invisible when it has been made correctly.

I missed several targets that were there. Does that ruin the run?

It changes what the slope means, and the results panel counts those errors separately for that reason. Calling a target missing is the error that flatters a search: it ends the trial early, contributes a short time to that display size, and pulls the line down. Calling an empty display full does the opposite and is rarer. A handful of either across 128 trials is ordinary; a long list of missed targets at the largest display sizes usually means the five-second window was being used as an excuse to stop looking, and the honest reading of that run is that it measured patience.

What a search slope tells you that an accuracy score cannot

The phrase “selective attention” covers two operations that behave nothing alike, and the whole value of measuring a slope rather than a score is that it tells them apart. Some properties of a scene are registered across the whole visual field at once — orientation is one, and so are size, motion and brightness — so a target that differs from everything else in one of them is available immediately, and adding distractors costs almost nothing. A target defined by a combination of two properties has no such shortcut: each property on its own is shared with some of the clutter, so candidates have to be considered in turn, and every added item adds time. The first pattern produces a line near flat; the second produces a line with a real gradient. An accuracy score would show almost nothing here — people are near ceiling in both conditions — and would miss the entire distinction.

Two things follow that most browser versions of this task get wrong. The first is that a slope is uninterpretable without the display it was measured on. Item spacing, how large a tilt separates target from distractor, and how many items fit on the screen all move it, so this page states its field size, its grid, its jitter and its bar dimensions in the specification table above rather than quoting a slope as though it were a property of the person. The second is that the absolute times are largely not yours: display latency, input latency and browser overhead sit inside every one of them, add up to a constant of tens of milliseconds, and land in the intercept. That constant divides out of a slope almost completely, which is the reason this measurement survives being taken in a browser at all when a raw reaction time barely does.

The neighboring phenomenon everybody arrives asking about is inattentional blindness — the gorilla walking through the basketball game. It is real, it is large, and it is a different claim from anything on this page: it is about an unexpected event during a demanding task, not about finding a target you were told to look for.

Inattentional blindness rate

≈ 50%

Simons & Chabris (1999), Gorillas in our midst: sustained inattentional blindness for dynamic events, Perception

About half of observers miss the unexpected event, and the rate rises with how hard the counting task is — so this is a property of the task, not a trait of the person, and nobody who misses it has a deficit. It also cannot be measured twice: a visitor who has heard of this study is no longer an observer of it, which is why a page using this paradigm gets one honest trial per visitor and should say so before asking.

That last sentence is why this page describes the study and does not stage it. A test you can only take once, and only if you have never heard of it, is a demonstration rather than a measurement — and this site is the second thing. If you want the same question asked with the distractor sitting immediately beside the target instead of scattered around it, the flanker task puts two of them on each side; with the distractor inside the target itself, the Stroop test makes the word fight the ink. For what happens to the same filtering after ten minutes of it, the attention span test and the continuous performance test hold the task constant and vary the time instead, and the concentration test puts the same needle-in-clutter problem on a clock. The listening equivalent, picking one interval out of a texture, is the relative pitch test.

A reaction time here is the interval between the frame that painted the stimulus and the timestamp the browser attached to your key, both read from the same monotonic clock. What neither can see is the display pipeline behind it, so on a 60 Hz screen roughly 16 ms of every figure below is the machine rather than you. That is the timing floor: two numbers closer together than that are the same number, and this page reports no precision it cannot support.

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 an attention deficit, a visual processing difficulty or anything else. Only a qualified professional, working with more than a browser, can make that judgment.

Where the 128 search times are kept

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 bar positions are generated in this tab from a seed printed with your result, so the arrays you saw can be regenerated from that number alone and never need to be stored. The times themselves sit in one array until you close the page, and the copy button is the only thing that moves any of it anywhere — into your own clipboard.