Why do coincidences feel impossible? “Random” is the wrong comparison.

Image: Five six-sided dice. Pierre-Selim Huard / Wikimedia Commons · CC BY 3.0.

You think of someone you have not spoken to in months. Before the thought has properly left, their name appears on your phone.

It does not feel merely unlikely. It feels addressed.

There are two familiar ways to ruin this moment. One is to declare it a sign. The other is to say “apophenia”, as though giving the pattern-detection mistake a Greek name has calculated anything.

Both answers arrive too early.

A coincidence can be exactly what chance produces when it receives enough attempts. It can also be a clue that the events were never independent, that a shared cause is hiding upstream or that the model called “chance” is simply the wrong one. The goosebumps cannot tell those possibilities apart.

The probability of this exact event is usually the wrong number

Start with a shuffle of playing cards. The chance of receiving that exact order is about one in 8 followed by 67 zeros. Yet somebody has to receive an order, and every possible order is equally preposterous when described card by card. Tiny probability alone does not make an outcome interesting.

The useful calculation depends on what you would have noticed. If you would have been startled by a matching birthday, a shared hometown, an unusual surname, the same song, a near miss by one day or any of twelve other connections, the target was not one microscopic point. It was a large and blurry family of possible hits.

The birthday problem exposes how quickly those opportunities multiply. In a group of 23 people, the probability that at least two share a birthday is just over 50% under the usual simplifying assumptions. The comparison is not one person against a 365-day calendar. Twenty-three people create 253 pairs.

Now replace 23 birthdays with years of people, thoughts, messages, journeys, names, dates and conversations. Add approximate matches. Add the ability to decide afterwards which detail mattered. The denominator becomes enormous, mostly without being recorded.

Statisticians Persi Diaconis and Frederick Mosteller made this point formal in their 1989 paper on coincidences. Their methods extend beyond the tidy birthday puzzle to unequal event rates, dependence and “almost” matches—the features real stories usually contain. In physics the related problem is called the look-elsewhere effect: a bump found at one preselected location is not statistically equivalent to the most exciting bump found after scanning an entire range.

That is the first answer. Many coincidences look impossible because we count the hit and misplace the search.

Random does not mean politely mixed

There is another trap. People often expect randomness to alternate: head, tail, head, tail, with only the occasional small run permitted under supervision. Real random sequences are less considerate. Independent events naturally clump.

Ruma Falk and Clifford Konold reviewed evidence that people generating supposedly random sequences tend to switch too often. In three experiments, participants’ judgements of randomness were better predicted by how difficult a sequence was to encode than by its objective randomness. A run such as HHHHH is effortless to describe, so it feels designed. A jagged sequence that resists compression feels random.

But a fair coin can produce five heads. Given enough flips, it must produce many striking local runs. Randomness describes the generating process, not a visual requirement that every short section look evenly scrambled.

This helps explain why precision has a particular psychological texture. “A sequence of 100 coin flips” is a mass of detail. “Five heads in a row” is a short sentence. The pattern feels like something because the mind can package it as something.

That connection is directly demonstrated for artificial sequences. Extending it to the full emotional force of “I thought of her and she called” is a plausible bridge, not a completed experiment. Personal relevance, memory and interpretation add layers that a row of coin flips does not have.

A coincidence can be evidence without being proof

Here the standard debunking begins to fail. If people only hallucinated order, coincidences would be cognitive litter. Yet noticing an unlikely alignment is also how hidden causes are found.

Imagine two laboratory instruments producing the same peculiar spike at the same second. Under a model in which their errors are independent, that alignment may be unlikely. Under a model with a shared electrical fault, it may be much more likely. The coincidence gives you a reason to investigate the power supply.

Psychologists Thomas Griffiths and Joshua Tenenbaum formalised this idea using Bayesian causal inference. In their account, a coincidence supports an alternative to the currently favoured causal theory without necessarily providing enough support to accept it. Across a series of experiments, people’s judgements tracked the statistical strength of coincidences surprisingly well.

The distinction is easy to lose. Evidence asks whether the event is more expected under one model than another. Belief also has to include how plausible those models were beforehand.

A message arriving during a thought may be somewhat more likely if the sender and thinker share a recent prompt than if their mental and messaging streams are entirely unrelated. That supports looking for a prompt. It does not place telepathy and “we both saw the same birthday reminder” on equal starting ground.

The researchers’ more generous conclusion was that people can sometimes assess coincidence strength accurately, then reach irrational conclusions by overestimating the prior plausibility of a novel causal force. The alarm can be appropriately sensitive while the suspect list is badly ranked.

The alternatives to chance are usually ordinary

When two events align, “pure coincidence” and “mysterious connection” are not the only options. The events may share a cause you have not noticed.

Two friends mention the same obscure film because a streaming service has just promoted it. Two travellers meet twice because both are moving through the same conference schedule. A person you were thinking about calls because an anniversary, news item or mutual friend prompted both of you. None of these explanations requires direct contact between the two events.

Social networks also make the independent-stranger model particularly unreliable. Duncan Watts and Steven Strogatz showed that small-world networks can be highly clustered while maintaining surprisingly short paths between distant members. That paper does not calculate the odds of your airport encounter. It establishes the relevant structural warning: real networks are neither a bag of independent strangers nor a perfectly regular grid.

Geography, routines, algorithms and shared contacts concentrate opportunities. Once those dependencies are included, an encounter can remain delightful while becoming statistically less exotic.

Do some people lower the pattern threshold?

Pattern detection always involves a trade-off. A smoke alarm set to maximum sensitivity catches faint fires and complains about toast. A cautious alarm stays quiet more often and misses some smoke.

Signal-detection theory separates two things that everyday language bundles together: the ability to distinguish signal from noise, and the threshold for saying a signal is present. In a 2023 study of 723 people judging noisy images of faces and houses, stronger paranormal belief was associated with a more liberal response threshold and lower perceptual sensitivity. The pattern was less clear for conspiracy beliefs, and the belief associations were modest beside other sources of variation.

The study was correlational. It cannot tell us whether belief changes perception, repeated interpretation changes belief or another factor influences both. It also cannot turn one participant’s score into a diagnosis of how they interpret their life.

A popular extension claims that losing control makes people see patterns. A prominent 2008 paper reported that effect across six experiments. That sounds like the answer. It is not quite.

In 2018, Michiel van Elk and Peter Lodder reported seven experiments using measures including magical thinking, conspiracy belief, paranormal belief and agent detection. They found no effect of experimentally induced loss of control, and their Bayesian analyses supported the null. When they redid an earlier meta-analysis, they found strong signs of publication bias and a smaller adjusted effect.

The careful conclusion is not that control can never matter. It is that a neat causal claim escaped the evidence ahead of schedule.

Then skepticism found a pattern that was not there

The most useful warning comes from basketball.

In 1985, a famous study concluded that players, coaches and fans were mistaking random clusters of successful shots for a “hot hand”. The result became a textbook example of faulty pattern perception.

Three decades later, Joshua Miller and Adam Sanjurjo found a subtle bias in the statistic used to test it. In a finite sequence of independent hits and misses, the measured success rate after a streak is expected to fall below the underlying hit rate. A genuinely constant shooter can therefore look slightly cold after a run. Treating that biased value as though it should be 50% tilted the test against finding a hot hand.

After correction, the canonical conclusion reversed. In a 2024 analysis of controlled shooting datasets, the same researchers reported strong evidence of hot-hand performance across datasets and within individuals over time. Expert observers could also predict which shooters would become hottest.

This does not make every streak meaningful. Random shooting still produces runs, in-game defences change shot difficulty, and effects differ among players. It shows something more uncomfortable: “you are seeing patterns in randomness” can itself be an artefact of using the wrong model of randomness.

Pattern believers can count too few chances. Pattern debunkers can build a null model that erases structure. Neither side gets to skip the calculation.

How to audit an uncanny coincidence

You do not need to drain the pleasure from a coincidence to examine it. Five questions do most of the useful work:

1. What would have counted? Define the match honestly. Would one day either side, a related name or any old friend have produced the same feeling?
2. How many chances were there? Count people, days, thoughts, messages, places and interpretations—including the ones that did not match.
3. Were the events independent? Look for shared routines, networks, news, algorithms and other common causes.
4. Which alternative predicts this specifically? “Something connected them” explains almost anything. A useful hypothesis risks being wrong.
5. Does it work prospectively? A relation written down before the next event, repeated across independent cases, is far stronger than a story assembled afterwards.

The same questions are useful when a coincidence opens an irresistible information gap. As the research on why not knowing can itch suggests, a sharply outlined missing explanation is exactly the kind of gap that recruits investigation.

The feeling is a question, not an answer

Why do coincidences feel too precise to be random?

Sometimes because the event is compact, personal and selected from more opportunities than the mind has counted. Sometimes because random processes clump. Sometimes because “random” quietly assumed independence where a shared cause or network had already linked the events.

And sometimes the surprise is doing useful intellectual work. It is telling you that the current model may not explain the observation as well as another one. That is how coincidences become discoveries—but only when the alternative is specified, the denominator is recovered and the prediction survives outside the original story.

The stranger truth is not that people are foolish for noticing coincidences. It is that the same feeling can accompany an inevitable fluke, an ordinary hidden cause and the first clue to real structure.

So do not ask only, “What were the odds of that?” Ask: “Under which process—and compared with what?”