Why does everyone suddenly agree at once?

Picture a plain room at the University of Virginia. Seventy-two economics students walk to the front, one at a time, and each draws a single ball from a covered urn, either type-a or type-b, then sets it down again, unseen by the rest of the room. Two urns sit behind the table, and only the person running the session knows which one is live: urn A holds two type-a balls and one type-b ball, urn B holds two type-b balls and one type-a ball, so a single private draw points at the truth about two times out of three. Each student, having looked at their own ball alone, calls out which urn they believe it is. Two dollars ride on a correct guess, and the session runs fifteen rounds: a 1997 study by Anderson and Holt, published in the American Economic Review.

Now walk into period nine of one particular session. The urn in play is B, but the first two students to speak both happen to draw type-a balls, the rarer signal, and both call urn A. The third student’s own private draw points at B, the truth, correctly. Rational as anything, that student does the math anyway: two people spoke before them, and two agreeing votes outweigh one private ball. The third calls A. So does the fourth. So do the fifth and sixth. Finally, the bottom row shows a reverse cascade in which urn B was used, but the first two decision makers saw a signals and predicted urn A. All four subsequent decision makers followed this pattern, despite their private b draws. Four students in a row watched their own evidence point one way and called the other, and every one of them was doing exactly what the math said to do.

Here is the part that trips people up: this behavior is entirely rational. Bikhchandani, Hirshleifer and Welch wrote the paper that named it, in 1992, and their whole point runs against the idea of a crowd going mindless. They called the pattern an informational cascade, and their own definition is exact: An informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individual without regard to his own information. Optimal is the word doing the work in that sentence: a deliberate piece of reasoning, using real information about what other people did, arrived at on purpose every time.

Nobody in their model needs to like conformity, or care what the room thinks of them. a reasoning process that takes into account the decisions of others is entirely rational even if individuals place no value on conformity for its own sake, they wrote a few pages later. The failure, when a cascade does fail, never lives in any one person’s head. It lives in the room: once enough people have gone quiet about what they privately saw, the information stops moving, and everyone after inherits that same silence.

You might picture a cascade needing a crowd first, some critical mass before individual judgment gets swamped. Two people are enough. Anderson and Holt spell out the arithmetic behind their own urn game: the first two decisions can start a cascade in which the third and subsequent decision makers ignore their own private information, and cascade behavior actually showed up in the lab in forty-one of the fifty-six rounds where the theory said it should. Two aligned voices, and the third person’s own eyes stop mattering.

Here is the model in miniature: eight identical items, the same eight every time, sent into three separate worlds. Every world begins from the same quality scores and no other advantage. Turn the dial and watch how much each world’s winner rides on copying its own first movers, rather than on the item itself.

At zero, all three worlds crown the same item, every single time you press New world: quality alone is doing the choosing, and three separate tries land on the same answer because there is a real answer to find. Carry the dial past roughly a third of the way and the worlds start to split. The filled bar drifts to a different item in each panel, out of the same eight starting scores, because each world is now weighing its neighbors’ early noise as though it were evidence. Same eight items, same starting scores, three different winners, for no reason greater than which noise crossed which world first.

The toy version fits in a browser tab. The real one needed something closer to a small city: published in Science in 2006 by Salganik, Dodds and Watts, it is the study people now know as the MusicLab experiment. 14,341 people downloaded previously unknown songs, every song starting at zero downloads, every participant working either from their own ears alone or from a running count of downloads so far. The ones working from a running count got split again, into eight separate worlds, each accumulating its own downloads, each blind to every other world’s chart. Same forty-eight songs, eight parallel histories. Salganik and his coauthors are careful with the family tree: they call this a cousin of the informational cascade rather than a rerun of it, related through the same engine of copied choices, sharing a shelf with Bikhchandani, Hirshleifer and Welch’s own paper in their own footnotes, but built and measured on its own terms.

The finding, in the paper’s own words: Increasing the strength of social influence increased both inequality and unpredictability of success. Eight worlds, the same forty-eight songs, and different songs became the runaway hit depending only on which world you were standing in. Quality still mattered, some: the best songs never do very badly, and the worst songs never do extremely well, but between those two edges almost any outcome was on the table, decided by whichever early download count a listener happened to see first.

You meet a miniature version of this every time you open a tab full of trending charts, bestseller badges, and popularity sort orders. A ranking built from other people’s visible choices carries the same DNA as a MusicLab world: once enough of a ranking accumulates from early momentum rather than from your own taste, later visitors inherit that momentum as if it were a verdict. Some rankings get an extra push from money changing hands, and that is its own story. But even an honest chart, with zero funny business anywhere in it, can crown a winner for no better reason than which choice arrived first and got copied. The fix costs one habit, the same habit this whole guide has been building: treat a chart position as a lead worth checking, never as a verdict already reached, and go looking for the people who tried the thing on their own terms before the chart existed. That is a different kind of evidence entirely, the kind a chart position can never quite fake.