The term “illusory correlation” was coined by psychologist Loren J. Chapman in 1967, but the concept itself is easier to understand than the name suggests. Chapman noticed something strange happening inside clinical psychology: trained professionals were reporting connections between patients’ behavior on projective tests, e.g., Rorschach and ink blot tests, and their diagnoses. These were connections that, statistically, didn’t actually exist.

In one of the most cited early experiments, Chapman and his wife and colleague Jean Chapman asked clinical psychologists to review drawings made by patients as part of a “Draw-A-Person” test. Some clinicians reported seeing a strong link between patients who drew people with large eyes and those diagnosed with paranoia. When the researchers analyzed the actual data, no link actually existed. The clinicians had completely invented the pattern.

What explains that? Two things, mostly.

First, our brains are wired to notice what’s unusual. When something rare or distinctive happens at the same time as something else rare or distinctive, it feels significant, as if there must be a reason for it. This is sometimes called the distinctiveness effect. That co-occurrence grabs our attention in a way that hundreds of unremarkable pairings never would.

Second, we tend to remember the hits and forget the misses. Once we form a belief that two things are connected, we unconsciously start cataloging the moments that confirm it. But all those times that didn’t fit, we conveniently let those times slip by.

In 1976, David Hamilton and Robert Gifford took this a step further. They ran a series of experiments showing participants descriptions of behaviors attributed to two fictional groups. One group was larger, the other group was smaller. The ratio of positive to negative behaviors was identical across both groups. But because negative behaviors were rarer overall, and the smaller group was already less common, participants consistently overestimated how often the smaller group behaved negatively. The two rare things got mentally linked, even though the data showed no actual difference.

This is worth taking a second to think about. Hamilton and Gifford didn’t find that the bias was driven by prejudice or conscious intention. It came from the way human memory processes frequency and importance. Rare things get stored differently because they feel more important. And when two rare things show up together, the brain files them away as connected.

The famous duo Daniel Kahneman and Amos Tversky’s work in the 1970s on cognitive heuristics added another interesting layer. Their research on the availability heuristic, that is, the idea that we judge probability based on how easily examples come to mind, helps explain why illusory correlations stick. The co-occurrence that felt notable is the one we can recall most easily, which makes it seem even more common and meaningful than it is.

The result is a human brain that’s constantly connecting dots, even when those dots belong to completely different pictures.

Illusory correlation is the bias most likely to send a product team in the wrong direction without anybody noticing. And that’s because it can feel like insight. It can sound like a signal. But it’s almost always built on a handful of vivid cases plus the human brain’s urge to connect them.

Start with how teams read their own metrics. A feature ships on Monday, weekly active users tick up on Wednesday, and by Thursday, the team starts referring to the launch as a “win.” But marketing ran a campaign on Tuesday, a competitor’s website was down on Wednesday, and traffic always lifts midweek.

The launch is the most available explanation, so it gets the credit. Future roadmap arguments now lean on a causal claim that the data never supported to begin with.

A/B testing makes the problem mechanical. Ron Kohavi, who ran experimentation at Microsoft and Amazon and basically wrote the book on A/B testing, has found that roughly one in four “statistically significant winners” at big tech companies fail to replicate. Teams that peek early, run underpowered tests, or try many variants without correction will see patterns that are pure noise. The dashboards look convincing because of the clustering illusion, the random data’s tendency to clump together in ways that look meaningful. If you want to go deeper on that one, we’ve got a whole episode on it. But the short version is that clustering looks identical to a real signal until you actually check the math.

The bias also distorts how teams think about people. Three angry support tickets from healthcare customers, and suddenly, “enterprise healthcare is a tough segment” becomes received wisdom. The 200 quiet, satisfied healthcare customers don’t generate tickets, so they don’t enter the team’s mental sample. That’s Hamilton-Gifford’s findings in production: rare group plus rare event, encoded with too much weight, hardening into a stereotype that drives prioritization.

User research gets the same treatment. Six interviews with power users, two mentions of keyboard shortcuts, and a persona narrative crystallizes. From there, every quote that fits the story gets remembered; quotes that contradict it slip out of the synthesis deck. The Chapmans showed this happens even when the underlying pairings are random, so it will definitely happen when the sample is only six.

Engineering teams build their own folk correlations, too. The “Tuesday curse.” The intern who “broke prod.” The exec demo that “always” triggers an outage. Two distinctive events occur concurrently: the team encodes a vivid memory, and a superstition becomes a working assumption that shapes deployment schedules, code review intensity, and how junior engineers get treated.

The team-dynamics damage compounds. Decisions feel data-driven because someone can name a specific anecdote. Disagreement gets framed as “not paying attention to what’s obvious.” Quiet skeptics who ask for the base rate get labeled obstructionists. The team loses calibration without losing confidence, which is the worst combination.

The fix isn’t more data, but better contingency thinking. How often does the outcome happen when the suspected cause is absent? How often does the suspected cause happen with no outcome? Without that habit, teams will keep mistaking pattern recognition for understanding.

The fix isn’t to distrust every observation. It’s to ask a boring but powerful question before acting on a perceived pattern: how often did the thing happen without the other thing? That’s the data point illusory correlation conveniently forgets.