Curse of knowledge

These illustrate the pattern as researchers describe it. How strong and how general it is remains disputed (see Evidence).

“Turn where the gas station used to be”

Someone who has lived in a town for thirty years gives a visitor directions: “Go past the old mill, then turn left where the gas station used to be. You can’t miss it.”

The directions are perfectly clear to someone who remembers the gas station. The speaker knows the landmark so well that it’s hard to register that a visitor has never seen it and has no way to find it. This is the core case: your knowledge makes a message seem complete when it isn’t.

The sarcastic text

After a long, dull training session, someone texts a coworker who wasn’t there: “Well, THAT was the best three hours of my life.” The coworker, who had heard the session was good, replies: “Glad you liked it! I’ll sign up for the next one.”

The sender knows how the session went, so the sarcasm seems obvious in the words themselves. The reader has only the words. This is a subtler form, because nothing specialized is involved: the knowledge that “curses” the sender is simply their own experience and intention. Research on ambiguous messages finds that speakers tend to overestimate how well their meaning comes across.

The easy quiz

A club organizer writes a trivia quiz, looking up each answer as they go. Reviewing it, they think: “These are pretty easy. Most people will get at least eight out of ten.” The average score on the night is four.

Once you’ve just read the answer, a question feels easier than it is, and the feeling of difficulty a newcomer would have is gone. Estimating what others know after learning it yourself is one of the settings where this has been studied directly.

• When the other person really does share the knowledge. Colleagues with the same training, old friends and people who were there can reasonably be assumed to know what you know. Assuming common ground is only an error when it isn’t there.
• When you have evidence about the audience rather than your own sense of it. Test results, feedback or a trial run with real newcomers tell you what they understand, independent of what you know.
• When you’re deliberately writing for insiders. A specialist manual that assumes expertise is making a choice about its audience, not a misjudgment.
• When your own experience is the best guide available. Using how hard something was for you to judge how hard it is for others is a reasonable starting point. The trouble begins when learning has erased that experience.

The test: would this make sense to someone who knew only what they’ve been told?

Shop talk between nurses

At shift change, one nurse tells another: “Bed four, post-op day two, NPO after midnight, sats fine on two liters.”

To an outsider this is nearly incomprehensible. But both nurses share the training and the setting, so the shorthand is efficient, not a failure to imagine the listener’s position. That’s the other person really does share the knowledge condition.

The worksheet that worked last year

A teacher who finds a math worksheet trivial expects her new class to finish it in fifteen minutes. She isn’t relying on her own sense of it: last year’s class, at the same level, took fourteen minutes on average.

Her expectation matches what the curse of knowledge would predict, but it rests on evidence about students, not on how easy the work feels to her. That’s the evidence about the audience condition.

Colin Camerer, George Loewenstein and Martin Weber, who named the effect (crediting the phrase to Robin Hogarth), described it as a failure to set information aside. A well-informed person predicting a less-informed person’s judgment ought to ignore what only they know. Instead, their prediction lands somewhere between what the less-informed person would think and what the informed person knows to be true.

The idea is closely related to Hindsight bias, and some researchers treat the two as one family: hindsight bias is the curse of knowledge applied to your own earlier self, whose ignorance you can no longer reconstruct once you know how things turned out.

Researchers disagree about the mechanism:

• Anchoring on yourself. On one account, people start from their own knowledge and adjust for the other person’s position, but not far enough.
• Losing the cues. Jonathan Tullis and Brennen Feder argue that once you’ve learned something, you lose the cues, such as whether you could answer and how long it took, that would tell you how hard it is for others. In their studies, reducing people’s reliance on their own knowledge did not make their estimates more accurate.
• Being the source. Boaz Keysar and Anne Henly found that speakers overestimated how often listeners understood their ambiguous sentences, but people who overheard and also knew the intended meaning did not. That suggests part of the effect comes from being the one who means something, not from knowledge alone.

It differs from the False consensus effect, which is overestimating how many people share your opinions, choices and habits. The curse of knowledge concerns what people know, not what they prefer. It also points the other way from the Illusion of explanatory depth, in which people overestimate their own understanding.

• Test on a real newcomer. Have someone who doesn’t know the material follow your directions, read your message or take your quiz. Their result is evidence you can’t produce by imagining it.
• Feedback helped in one experiment. In a preregistered study by Debby Damen and colleagues, people who judged how an uninformed person would read a sarcastic message became less biased after being told how such readers actually interpreted similar messages.
• Trying harder to take the other view hasn’t reliably worked. Damen and colleagues report that explicitly directing attention to what the other person knew did not help in an earlier study. Camerer and colleagues found that neither money for accuracy nor feedback reduced the bias in individuals, although trading in a market cut it by about half.
• Spell out the steps and landmarks you’d be tempted to skip. This is common sense rather than a tested remedy.

Status: contested. Informed people have repeatedly been found to overestimate what the uninformed know, across quite different tasks. But there is no multi-lab replication or meta-analysis, and in the best-known laboratory test with adults, large replications found the effect less than half its original size or, in one preregistered series, found none. By the site’s criteria, evidence between two labels takes the less confident one.

Foundational studies.

• Camerer, Loewenstein and Weber (1989) had Wharton students predict eight companies’ 1980 earnings from investment reports. A second group, told the actual earnings, then traded assets that paid out according to the first group’s average prediction, and estimated that prediction. Their estimates were pulled toward the true earnings they knew. Financial incentives and feedback did not reduce the bias; market trading cut it by about half but did not remove it.
• The tapping study. The best-known illustration comes from an unpublished 1990 Stanford dissertation by Elizabeth Newton. As recounted by Chip and Dan Heath (2006), people tapped out the rhythm of well-known songs such as “Happy Birthday” for listeners to identify. Across 120 songs, listeners named 3 (2.5%), while tappers had predicted that listeners would get about half. Because the study was never published in a journal, it’s a vivid example rather than strong evidence.
• Keysar and Henly (2002) found that speakers asked to convey a specific meaning with an ambiguous sentence consistently overestimated how often their listeners understood it.

The false-belief task, and its replications. Birch and Bloom (2007) gave adults a story in which a girl, Vicki, puts her violin in one of four containers and leaves; her sister moves it. Participants estimated where Vicki would look first. Those who knew where the violin had been moved, when that location was plausible, gave it more probability than those who didn’t know, suggesting their knowledge crept into their judgment of Vicki’s false belief.

• Ryskin and Brown-Schmidt (2014) ran seven experiments with larger samples. The effect was in the same direction but estimated at d = 0.20 across all seven, less than half the size calculated from the original (d = 0.47), and significant in only some experiments. (Here d measures the size of a difference between groups; 0.2 is conventionally “small” and 0.5 “medium”.) By their own criterion this counted as a failure to replicate, and they suggested the effect’s real-life importance may need reevaluating.
• Farrar and Ostojić (2018), in three experiments with 283, 281 and 744 participants, found the effect in all three when adults reasoned about a person like themselves, with a pooled estimate across conditions of d = 0.31.
• Samuel (2023), in a series of experiments with a version of the same task, several of them preregistered, found no evidence that knowing the object’s specific new location biased people’s judgments. He suggested that differences in how the task is presented may explain the conflicting results.

Other tasks. Tullis and Feder (2023), in four experiments, found that studying the answers to trivia questions made people’s estimates of how many novices would know them less accurate. Damen and colleagues (2021), in a preregistered study with 142 participants, built on an earlier direct replication of a classic experiment in which readers who know a message is sarcastic expect uninformed readers to catch the sarcasm.

What remains uncertain is how large and how general the effect is, and why it happens. That knowledge sometimes distorts judgments of what others know is widely reported. How much it matters in any particular setting, and whether the laboratory false-belief version is real in adults, are not settled.