About Can You Fake Randomness?
Tap heads or tails as randomly as you can. Then find out how badly you failed — because almost everybody does, and the failure is statistically detectable.
Three genuine tests run on your sequence. A Wald–Wolfowitz runs test counts your unbroken blocks against the number independent fair flips would produce; humans alternate too much, so they produce too many. An exact longest-run test compares your longest streak against the true distribution of longest runs for your sequence length — computed here, not looked up — because people cut streaks short at three or four when six is ordinary. An exact binomial test checks your head count, because people balance too carefully; a real coin rarely lands on exactly half. The three p-values combine through Fisher’s method into a single number.
What that number means is stated precisely rather than dressed up: it is the confidence that a fair coin did not produce your sequence. The tests were validated by simulation before release — on fair coins they raise a false alarm about 4% of the time, as they should, and at 120 flips they catch a strong human bias about 96% of the time and a mild one only about a third. Nothing is uploaded; every calculation happens in your browser.
Go at a natural pace and do not overthink it — deliberating changes what is being measured. The sequence is hidden while you tap so you cannot audit yourself.
Every run is recorded in this browser so you can see whether practice helps. It usually does, a bit — and the way it helps is instructive: people improve mainly by forcing themselves to allow longer streaks.
Hand the device over. The second player taps the same number of flips and the two results are scored side by side.
Coins are the easy case. Tap 0–9 as randomly as you can and the same logic applies with sharper teeth, because this is exactly how people pick PINs. Two tests run: a chi-square check of whether every digit appeared equally often, and a repeat check, since people almost never press the same digit twice in a row when chance says they should about one time in ten.
How to Use the Fake Randomness Test
Choose a length, then tap heads or tails as randomly as you can manage — keyboard arrows work too and are faster. The sequence stays hidden while you go, deliberately: being able to review what you have already produced turns an instinct test into an editing exercise, and it is the instinct that is interesting. Go at a natural pace. When you finish, three genuine statistical tests run and each reports its own verdict before they combine into one.
The Three Tests, and What Each One Catches
- The runs test. A run is an unbroken block of the same outcome, so
HHTTTHhas three. For independent fair flips the number of runs has a known mean and variance, so yours converts to a z-score. Over-alternating produces too many runs, and this is the single most reliable human signature. - The longest-run test. People truncate streaks. The exact probability distribution of the longest run is computed here for your specific sequence length, so your longest streak is compared against the truth rather than against a rule of thumb.
- The balance test. An exact binomial test on your head count. Humans balance too neatly, aiming for a near-perfect split, while a real coin lands there surprisingly rarely.
The three p-values are combined with Fisher’s method, which sums their logarithms into a chi-square statistic — the standard way to pool independent tests. What comes out is a single probability, and the headline percentage is simply one minus it: the confidence that a fair coin did not produce what you produced.
Why Streaks Feel Wrong When They Are Not
The deepest reason people fail this test is a misconception about what randomness looks like. Random does not mean evenly spread; it means unpredictable, and unpredictable sequences are full of clumps. In 100 fair flips the longest run averages around seven identical outcomes. Runs of five and six are entirely routine. Yet almost nobody hand-generating a sequence writes four in a row, because by the fourth it stops feeling random.
Psychologists call this the representativeness heuristic: we judge a sequence by how closely it resembles our idea of the process, rather than by how likely the process is to produce it. It is the same instinct behind the gambler’s fallacy — the feeling that after five reds, black is somehow due. It is not. The wheel has no memory, and the run you are staring at is exactly the kind of run randomness produces constantly.
Where This Costs You Something Real
This is a party trick with a serious edge. PIN choice is the clearest case: people avoid repeated digits, avoid adjacent repeats, and lean heavily on dates. The result is that the real distribution of four-digit PINs is nothing like uniform — a small handful account for a startling share of all PINs in circulation, and the effective search space is a fraction of the ten thousand it appears to be.
The same bias shapes hand-picked lottery numbers (which does not change your odds of winning, but does change how many people you split with), hand-shuffled playlists, and passwords chosen by feel. It is why every serious security recommendation says the same thing: do not generate the secret yourself. Use a source that actually draws at random, such as our Password Generator. And if you want to see how much a hand-made password costs you in practice, the Password Strength & Crack-Time Estimator prices it against real hardware.
What This Test Can and Cannot Tell You
Worth being precise, because a tool that overclaims about statistics has no business running statistics. These tests were validated by simulation before release. On genuinely random sequences they raise a false alarm about 4% of the time, which is what a 5% threshold should do. At 120 flips they catch a strong human bias (alternating around 70% of the time) in about 96% of cases, a typical one (around 65%) in roughly three quarters, and a mild one (around 60%) only about a third of the time.
So a low score does not certify you as random. It means either your bias is genuinely small or the sequence was too short to prove it — and with a hundred-odd data points, short is the norm. Running the 200-flip version raises the power noticeably. Every other tool of this kind will happily tell you that you are 87% random; none of them can justify the number.
Prefer a test with a stopwatch rather than a coin? The Reaction Time & Focus Test is a real continuous performance task with signal-detection scoring. Browse every Fun & Novelty tool for more.
Frequently Asked Questions
Why are humans so bad at being random?
Because we confuse randomness with the absence of pattern, and then work to remove any pattern we notice. Real randomness is full of pattern — clumps, streaks and imbalances appear constantly and mean nothing. When people generate sequences by hand they alternate far more than chance, typically switching about 60 to 70 percent of the time instead of 50, and they cut runs short because four heads in a row "feels" wrong. It is not wrong; in 120 flips a run of six is entirely ordinary.
What is a runs test?
A run is an unbroken block of the same outcome, so HHTTTH contains three runs. The Wald-Wolfowitz runs test compares how many runs you produced against how many independent fair flips would produce, which has a known mean and variance. Too many runs means over-alternation, the classic human signature. Too few means unnaturally long streaks. The test converts your count into a z-score and a p-value.
How long should the longest streak be?
Longer than almost anyone expects. In 100 fair flips the longest run averages about seven identical outcomes in a row, and runs of five or six are entirely routine. Most people hand-generating a sequence stop at three, because a fourth feels like it is no longer random. This tool computes the exact distribution of the longest run for your sequence length, so the comparison is precise rather than a rule of thumb.
Does a low score mean I really am random?
Not quite, and it matters. With 120 flips the test catches a strong human bias almost every time and a typical one about three times in four, but only about a third of mild biases. A result of "no significant deviation" therefore means your bias is small or the test lacked the power to see it — not that it is absent. That limit is a property of any statistical test on a short sequence, and pretending otherwise would be exactly the kind of overclaiming this tool exists to expose.
Where does being bad at randomness actually matter?
PIN and password choice, most of all: people avoid repeated digits and adjacent repeats, which shrinks the real search space far below the theoretical one. It also shows up in lottery number picking, in shuffling a playlist by hand, and in the gambler's fallacy that a run of red makes black "due". In security terms the lesson is blunt: never generate a secret yourself. Use something that draws from a proper random source.
