Why Humans Are Bad at Being Random
Asked to be random, people avoid repeats, alternate too often, and dodge the edges.
Ask someone to write down a hundred random coin flips without flipping anything, and the result is detectable. People produce too few long runs, alternate too often, and avoid the edges.
The errors are consistent enough that a statistician can separate a fabricated sequence from a real one with high accuracy, which is a striking fact about a task everyone believes they can do.
This matters practically, because it means a human drawing names from a hat or picking a number is not producing what they think they are producing.
The three consistent errors
Too much alternation is the most reliable. Asked to produce random binary sequences, people switch outcomes about sixty percent of the time rather than the fifty percent a fair coin gives. The result is a sequence that looks more balanced locally than randomness actually is.
Too few long runs follows from the same tendency. A genuine hundred-flip sequence contains a run of six or seven; a fabricated one usually maxes out at four, because a longer run feels like it is going wrong and the writer breaks it.
Avoiding the edges shows up in number tasks. Asked for a random number from one to ten, people disproportionately pick numbers in the middle, and seven is chosen far more often than chance would predict — the extremes feel less random than the middle.
- Alternation rate around 60% rather than 50%.
- Longest run typically 4 where a real sequence gives 6 or 7.
- Numbers cluster away from the endpoints of a range.
- In a 1-to-10 task, 7 is chosen far more often than chance predicts.
Why we do it
The standard explanation is representativeness: people judge a sequence to be random if it resembles their mental image of a random sequence, and that image is of local balance. A run of six violates the image, so it gets edited out.
There is also a simple mechanical explanation for the alternation bias. Generating a genuinely independent sequence requires ignoring what you just wrote, and people cannot — the previous entry is right there, and avoiding repetition is an almost automatic response.
Neither explanation is a deficiency exactly. Local balance is a reasonable expectation for most real-world processes, which do have memory. Independence is the unusual case, and our intuitions are tuned for the common one.
What it costs in practice
A human drawing names from a hat is not producing a uniform draw. The physical arrangement matters — slips near the top or the edges are more likely — and the drawer's hand movements are not independent of what they can feel.
A human choosing a winner "at random" without a mechanism is producing something considerably worse: an unmeasured preference dressed as a draw. That is the case where the failure has consequences, because it is indistinguishable from favouritism and cannot be defended afterwards.
Human-generated numbers are also predictable, which matters for anything that needs to be unguessable. PINs chosen by people cluster heavily on dates and repeated digits, which is why a small fraction of possible four-digit PINs covers a large fraction of those actually in use.
The mirror image: seeing patterns in noise
The complementary error is finding structure in genuinely random data. Because randomness produces clusters and runs, and because our image of randomness does not include them, real random data looks non-random to us constantly.
This is why a picker that returns the same name twice in a row generates complaints, and why a shuffled playlist that plays two tracks by the same artist reads as broken. In both cases the tool is behaving correctly and the observer's model is wrong.
The two errors together are difficult to argue against, because a person who under-produces runs when generating will also over-detect them when observing. The intuition is consistent in both directions and wrong in both.
What to do about it
Use a mechanism. That is the whole practical response: anything that removes the human from the selection removes both errors at once, whether it is a die, a hat that is genuinely shaken, or a generator.
Where the objection is social rather than statistical, make the mechanism visible. A spun wheel is not fairer than a silent draw and it is considerably more convincing, because the audience can see that no person made the choice.
And where people complain about a run, compute the expected rate of that run rather than explaining independence. Explaining independence does not work; a number sometimes does.
Frequently asked questions
How badly do people fail at generating random sequences?
Consistently enough to be detectable. Fabricated binary sequences alternate about 60% of the time rather than 50%, and their longest runs are typically about half what a real sequence gives.
Why do people alternate too much?
Because we judge randomness by local balance, and because ignoring what you just wrote is genuinely difficult. Avoiding repetition is nearly automatic.
Is picking numbers affected too?
Yes. People avoid the endpoints of a range and over-select the middle — 7 is chosen far more often than chance predicts in a 1-to-10 task.
Is drawing from a hat actually random?
Not uniformly. Slip position matters, and the drawer's hand is not independent of what they can feel. It is far better than choosing without a mechanism.
Why do people also see patterns in random data?
Because randomness produces clusters and our image of randomness does not include them. The two errors are the same intuition running in both directions.
Does knowing about the bias fix it?
Barely. The effect persists in people who know about it, which is why the practical answer is a mechanism rather than more care.
Why does a visible spin convince people more?
Because the objection is social rather than statistical. A visible mechanism shows that no person made the choice, which is what the audience actually doubts.
What is the best response to a complaint about a run?
Compute how often that run should occur. A number sometimes persuades where an explanation of independence does not.
Tools that use this
Related guides
The Gambler's Fallacy
The coin does not owe you a tails. Why the intuition is so strong anyway.
Coin Flip Streaks and What They Mean
How long a run of heads you should expect, as a function of how many times you flip.
The Law of Large Numbers, Practically
What 'it evens out' actually promises, on what timescale, and what it never promises.