Ten of the things on this site are not games. Nobody is playing against you: the machine holds a ruler and writes down a number. This page collects what those numbers are usually worth — where the human ceiling sits, where the machine's sits, and who established either of them.
The ceilings below are not records. A record is what one trained person did once; these are the values an ordinary adult produces in a quiet room, and almost all of them come from published experiments rather than from this site. Where a figure is ours, it says so. Every number here carries the name of whoever produced it, because a number without one is an opinion dressed as a measurement.
| Ability | Human limit | Machine limit | Test on this site |
|---|---|---|---|
| Simple reaction to light | median ~250 msin a laboratory simple visual reaction runs 190–200 ms: Woodworth and Schlosberg, Experimental Psychology, 1954. The ~250 ms is this site's own figure; nearly all of the difference is equipment | under 1 msthe browser timer; resolution in fractions of a millisecond | Reaction |
| Visual memory for positions | 5–9 numbers heldthat is the span with rehearsal and chunking allowed (Miller, Psychological Review 63(2), 1956). Block both and about four is left: Cowan, Behavioral and Brain Sciences 24(1), 2001 | no limitan array in memory; reading one element costs the same however many there are | Chimp Test |
| Arithmetic in the head | ~15 sums a minutethis site's own figure on easy sums; mental-calculation champions are an order away, and that is training, not the ceiling of the species | billions of operations a seconda 3 GHz core: a clock is about 0.3 ns and it issues several additions per clock | Mental Arithmetic |
| Instant estimate of number | exact to 4, then ±15%Jevons, Nature 3, 1871; the shape of the curve from Mandler and Shebo, JEP: General 111(1), 1982; the adult Weber fraction of 0.11–0.15 from Halberda and Feigenson, Developmental Psychology 44(5), 2008 | exact at any countcounting objects is one linear pass over the image | Count the Dots |
| Updating working memory | 2–3 backat 4-back untrained adults are near guessing: Kane et al., JEP: LMC 33(3), 2007; participants were moved to n=4–5 only by nineteen days of daily practice — Jaeggi et al., PNAS 105(19), 2008 | any na fixed-length queue; n does not change what a comparison costs | N-Back |
| Suppressing automatic reading | conflict costs ~100 mssingle-trial interference against an XXXX control runs from 35 to 180 ms, usually 70–130: MacLeod, Psychological Bulletin 109(2), 1991. Stroop's own card format (Journal of Experimental Psychology 18(6), 1935) gave 110.3 seconds against 63.3 for a hundred items — +74%, or about 470 ms an item | no such effecta program has no reading that switches itself on and then has to be overruled | Stroop Test |
| Visual search | flat, or +18 ms per itemsingle-feature search has a slope near zero, conjunction search tens of ms per item: Treisman and Gelade, Cognitive Psychology 12(1), 1980; the distribution of slopes from Wolfe, Psychological Science 9(1), 1998 | linear, in microsecondsa pass over pixels: the slope is linear too, but the step is thousands of times smaller | Find the Target |
| Producing a random sequence | does not happen: the run is predictableHagelbarger's relay machine guessed 5218 of 9795 presses — 53% where a coin gives 50%: Hagelbarger, IRE Transactions on Electronic Computers EC-5(1), 1956 | a table over 3–7 movesShannon, A Mind-Reading (?) Machine, Bell Labs memorandum, 1953; the modern version is a few dozen lines of code | Mind Reader |
| Telling machine text apart | at chance, about 50%untrained readers on GPT-3 output: Clark et al., ACL, 2021; the same for self-descriptions in Jakesch et al., PNAS 120(11), 2023 | it writes the texts in questionautomatic detectors are unreliable and biased against non-native writers: Liang et al., Patterns 4(7), 2023 | Spot the Machine |
| Reading distorted images | ~75% right on reCAPTCHA samplesand three independent humans agree on the answer only 71% of the time: Bursztein et al., IEEE Symposium on Security and Privacy, 2010 | from a few per cent to 77%, by schemea generative attack on text CAPTCHAs: Ye et al., ACM CCS, 2018 | Reverse CAPTCHA |
The number everyone quotes is Miller's: seven plus or minus two, from a 1956 paper whose own title calls it magical. Miller was describing digit spans, and a digit span lets you rehearse the list under your breath, group it, and lean on long-familiar number words. What is measured is capacity together with every trick for stretching it. Cowan's 2001 review gathered the experiments that block those tricks — rehearsal suppressed, chunking impossible, items presented faster than they can be named — and the span collapses to about four, with a spread of one either way. Four is the raw figure; seven is four plus the tricks. Neither number is wrong, and the whole subject is the distance between them.
G. Miller, Psychological Review 63(2), 1956; N. Cowan, Behavioral and Brain Sciences 24(1), 2001.Throw a handful of beans on a table and glance: at one, two, three and four you simply know how many, in about the same time and with almost no errors. Jevons established this on himself in 1871 by tossing beans into a box, and Mandler and Shebo measured the shape of the curve in 1982: up to four, reaction time rises by 40–100 ms per item, then jumps to two or three hundred milliseconds per item, which is the price of actual counting. Past that range you stop counting and start estimating, and estimates obey Weber's law — the error grows in proportion to the number. For adults the coefficient sits near 0.11–0.15, which is where our own "exact to four, then about fifteen per cent" comes from.
W. S. Jevons, Nature 3, 1871; G. Mandler, B. Shebo, JEP: General 111(1), 1982; J. Halberda, L. Feigenson, Developmental Psychology 44(5), 2008.Simple visual reaction in a laboratory runs about 190–200 ms; on this site the median is nearer 250. Neither number measures thinking. Turning light into a signal in the retina costs tens of milliseconds before anything leaves the eye; the signal reaches primary visual cortex in 60–70 ms; and after the command arrives the muscle needs another 30–50 ms to develop force. What is left for deciding is a small remainder. The extra fifty milliseconds here is equipment: a timer inside a browser, a display refreshing sixty times a second and so adding up to 16 ms on its own, and an operating system busy with something else. Online experiment platforms have been measured adding 30–100 ms of lag. Improving a reaction therefore means improving the equipment or anticipating the signal, not speeding up conduction along a nerve: that speed is fixed and attention does not adjust it.
R. Woodworth, H. Schlosberg, Experimental Psychology, 1954; M. Schmolesky et al., Journal of Neurophysiology 79(6), 1998; A. Anwyl-Irvine et al., Behavior Research Methods 53, 2021.In 1935 Stroop printed colour words in conflicting ink and asked people to name the ink. Reading the words themselves was no slower; naming the ink of conflicting words took 74% longer than naming the ink of coloured patches — 110.3 seconds against 63.3 for a hundred items. Per item that is about 470 ms, but the card format is what costs that much: a whole list read aloud in one go. Modern experiments present one item at a time, and there the interference against a row-of-Xs control runs from 35 to 180 ms, usually 70–130; the hundred milliseconds in the table above comes from that work, gathered by MacLeod, and not from Stroop himself. The interesting part is not the size but that the effect cannot be switched off. You are not asked to read; you are asked not to read; a literate adult reads anyway, and the extra tenth of a second is the price of overruling something that was never under voluntary control. MacLeod's 1991 review of half a century calls it one of the most reproducible effects in the field.
J. R. Stroop, Journal of Experimental Psychology 18(6), 1935; C. MacLeod, Psychological Bulletin 109(2), 1991.In n-back the items arrive one at a time and you report whether the current one matches the one n steps back. The load is not storage but continuous updating: every new item has to be admitted and the oldest discarded, in order. At 2-back an untrained adult is usually accurate; at 3-back discrimination falls sharply; at 4-back most people are near guessing. In Jaeggi's 2008 training study participants started around n=3 and reached 4–5 only after nineteen days of daily practice, and the transfer to other tasks proved far weaker than the headline suggested — later replications made that plain. A machine has no n at all: a list is a list, and fetching the item n places back costs the same at n=3 and at n=3000.
S. Jaeggi et al., PNAS 105(19), 2008; M. Kane et al., JEP: Learning, Memory, and Cognition 33(3), 2007.Treisman and Gelade showed in 1980 that search comes in two shapes. Looking for a red circle among green ones, the time barely depends on how many green ones there are: the slope is near zero and the target pops out. Looking for a red circle among red squares and green circles — a conjunction of two features — the time rises with every distractor, in the region of tens of milliseconds per item. Wolfe's million trials later showed these are not two mechanisms but the ends of one continuous distribution. The other side of the same coin is change blindness: with a brief blank between two images, Rensink's observers needed many alternations to spot a large change, and in Simons and Levin's field study about half the passers-by giving directions failed to notice that the person asking had been swapped for someone else.
A. Treisman, G. Gelade, Cognitive Psychology 12(1), 1980; J. Wolfe, Psychological Science 9(1), 1998; R. Rensink et al., Psychological Science 8(5), 1997; D. Simons, D. Levin, Psychonomic Bulletin & Review 5(4), 1998.Asked for a random string of lefts and rights, people avoid long repeats, alternate more often than a coin would, and answer their own last move. That is enough to be predicted. In 1953 David Hagelbarger built a relay machine at Bell Labs that kept a small table of what the player had done after each recent run of moves and bet on the same again: over 9795 presses it guessed right 5218 times — 53% where a coin gives 50%. Shortly after, Shannon built a smaller machine in the same lab, and in the play-off between the two, his won. The modern version is a few dozen lines with a five-move memory. The defence is not to feel random but to take the digits from outside your own head.
D. Hagelbarger, IRE Transactions on Electronic Computers EC-5(1), 1956; C. Shannon, A Mind-Reading (?) Machine, Bell Laboratories memorandum, 1953.Two of the tests here point in opposite directions. Distorted text is the classic case where people managed and machines did not; but people are not flawless at it either — Bursztein's group measured about 75% correct on reCAPTCHA samples, with three independent humans agreeing only 71% of the time, and generative attacks in 2018 solved several text schemes at rates from a few per cent to over seventy. Telling machine-written text apart has gone the other way, and faster. Clark and colleagues found untrained readers identifying GPT-3 output at close to chance, around 50%, with brief training helping little; Jakesch's group got the same for self-descriptions in 2023. One of these tests you should expect to win. The other you should expect to lose, and recently.
E. Bursztein et al., IEEE Symposium on Security and Privacy, 2010; G. Ye et al., ACM CCS, 2018; E. Clark et al., ACL, 2021; M. Jakesch et al., PNAS 120(11), 2023.The bars are powers of ten: how long one elementary step takes, counted in nanoseconds. Each step to the right is ten times the one before it, so the distance between an addition in a processor and a sum done in the head is not something practice closes.
They were obtained in a laboratory. A dim room, a fixed viewing distance, a hardware button, a participant who had slept and had done the practice trials. Here you have a browser, a display that refreshes when it feels like it, a phone that may have decided to install something, and a page opened between two other things. Every latency measured on this site is larger than the published one, and the difference is equipment rather than you.
The spread between people is wider than the difference between the two columns above in most rows. Simple reaction time varies by well over a hundred milliseconds across healthy adults of the same age, and by a comparable amount within one person across a single day. One run tells you about that run.
None of this is a diagnosis. These tests are demonstrations with a counter attached: not calibrated, not normed on any population, not administered by anyone. If a result worries you, what it is evidence for is that the question belongs to somebody qualified to answer it — not that you have anything.
And a ceiling is not a verdict on the species. Every figure here describes an unaided person doing one task in a few seconds. People also use paper, instruments, each other and centuries. The machine column, in turn, is what a browser tab does, and it is not the state of anything either.