In a room of 23 people the chance that two of them share a birthday is slightly better than even. It feels wrong because 23 is a small number against 365, and the intuition is comparing the wrong things.
Twenty-three people do not make 23 comparisons. They make 253 — every possible pair — and 253 chances at a one-in-365 event is where the probability comes from.
Why is it counted in pairs?
Because a match is a property of a pair rather than of a person. Each new arrival is compared against everyone already there, so the comparisons grow much faster than the headcount.
| People | Pairs | Chance of a match |
|---|---|---|
| 10 | 45 | about 12% |
| 23 | 253 | about 51% |
| 30 | 435 | about 71% |
| 50 | 1,225 | about 97% |
| 57 | 1,596 | about 99% |
| 70 | 2,415 | about 99.9% |
The pair count is the number of ways to choose two from the group, which the combinations article covers as a general rule. Here it is the whole explanation.
The probability is easier to compute backwards. Work out the chance that everyone has a different birthday — each person having to avoid all the dates already taken — and subtract from one.
What is the question people think they asked?
Whether someone shares a birthday with them specifically, which is a completely different problem with a completely different answer.
That version compares one person against everyone else rather than everyone against everyone, so it grows in a straight line rather than in pairs. Reaching an even chance takes 253 other people — the same number that appears as the pair count above, which is a pleasing coincidence and not a deep one.
Almost every argument about this result is the two questions being asked past each other. Both answers are correct for their own question.
Where does this actually matter?
Anywhere a collision is a problem and the number of items is much larger than it feels. Hash tables, short identifiers, randomly generated codes and file checksums all live under the same arithmetic.
The practical rule is that collisions start becoming likely at roughly the square root of the number of possible values. A million possible codes means collisions appear around a thousand codes issued, not around a million.
That is why short random codes need to be longer than intuition suggests, and why an identifier space is chosen against the square root of the count you expect rather than against the count itself.
What does the real calculation assume?
That birthdays are spread evenly across the year and that nobody in the room is a twin. Neither is quite true, and both push in the same direction.
Real birth dates cluster — by season, by day of the week, and around dates that hospitals schedule around. Any clustering makes matches more likely rather than less, so 23 is a slight underestimate rather than an overestimate.
The leap day is usually left out, which changes the answer by a fraction of a percentage point and is not the reason anyone finds the result surprising.
Why does intuition fail here?
Because people estimate the probability of a specific event and the question asks about any event. The mind reaches for "my birthday" because that is the concrete version, and the abstract version has far more ways to come true.
It is the same error behind being surprised at meeting an acquaintance abroad. The chance of meeting that particular person is tiny; the chance of meeting any of the several hundred people you would recognise is not.
Recognising which of the two a question is asking is most of the skill. Once the pairs are counted the arithmetic is ordinary.
How is the probability actually computed?
By multiplying the chances of successive misses. The second person has 364 of 365 days available to avoid the first, the third has 363, and so on — so the chance of no match at all is that product, and the answer to the original question is one minus it.
Computing it that way is why the result climbs so fast. Each new person multiplies in a slightly smaller fraction, and a product of many fractions just below one falls away much more quickly than adding them would suggest.
It also explains why the curve flattens at the top. Past fifty people the chance of no match is already tiny, so further arrivals change very little.
Questions people ask
Why is it called a paradox? Only because the answer conflicts with intuition. Nothing about the arithmetic is contradictory.
Does the same logic apply to lottery numbers? To repeated draws, yes. Two identical draws in a long history are far likelier than any specific draw repeating.
Does it need to be a random group? Any group not selected by birthday. A room of people born in June is a different problem.
How many for certainty? 366, or 367 allowing for the leap day. Certainty needs every date used up.
Does this apply to shared initials? Yes, and it happens sooner, since there are far fewer common initials than birthdays.
Is 23 exactly the crossing point? It is the first count above fifty per cent. At 22 the chance is just under.
Count the pairs, then the odds. The probability calculator handles the arithmetic for a single event and its repeats, and the combination and permutation calculators give the pair counts that drive the whole result.