Why large groups of ordinary people outguess experts — and the exact conditions under which they stop.

Core lessons

1. The crowd is smart only under conditions. Galton’s ox-weighing example illustrates how independently collected estimates can combine into a surprisingly accurate answer. It is not a claim that groups are always wise. Surowiecki’s book is about the conditions: diversity of opinion, independence of judgment, decentralization of knowledge, and a mechanism for aggregation. Remove any one and the crowd degrades into a mob, a committee, or an echo.

2. Errors cancel only when they’re uncorrelated. The statistical engine underneath everything: each person’s guess is truth plus error. If errors are independent, they cancel in aggregate and the signal remains. If everyone’s errors point the same way — because they read the same news, defer to the same expert, or watch each other guess — averaging amplifies the shared bias instead of canceling it. Independence isn’t a nicety; it’s the mechanism.

3. Diversity beats individual ability (within reason). A group of diverse, decently-informed people typically outperforms a group of the very best individuals, because the best individuals tend to think alike and share blind spots. Adding a less-expert but differently-thinking member can improve the group; adding another clone of the smartest member usually doesn’t.

4. Information cascades: how smart individuals make dumb crowds. If people decide in sequence and can see earlier choices, early movers’ opinions drown out later movers’ private information. Each rational individual concludes “all these people can’t be wrong” — and a crowd of rational people marches off a cliff. This is the anatomy of bubbles, pile-on funding rounds, and the empty restaurant next to the full one.

5. Three kinds of problems. Cognition problems (questions with answers: how much does the ox weigh, will this ship on time) — crowds excel here. Coordination problems (how do buyers and sellers find each other, which side of the road to drive on) — solved by conventions, norms, and markets. Cooperation problems (taxes, tipping, pollution) — require trust and institutions that make self-interest compatible with the group. Diagnosing which type you face tells you what machinery you need.

Key frameworks

The four conditions checklist. Before trusting any aggregate — a poll, a planning meeting, a prediction market, a feedback dashboard — audit it:

  1. Diversity: do participants bring genuinely different information and priors?
  2. Independence: did each judgment form before exposure to the others?
  3. Decentralization: can people draw on local, specific knowledge?
  4. Aggregation: is there a mechanism that actually combines the inputs (a market, a vote, a median), rather than a discussion where the loudest voice wins?

Collect first, discuss second. The most portable practice in the book: gather private, written estimates before any group discussion. Discussion is for surfacing reasons, not for forming the number.

Beware the talkative and the confident. In deliberating groups, talkativeness and status predict influence far better than accuracy does. Unstructured discussion makes groups more extreme (polarization) and more confident — often while making them less accurate.

The Columbia/NASA case. Surowiecki’s study of the Columbia shuttle disaster’s management meetings shows aggregation failing inside a hierarchy: information existed at the edges, but the meeting structure — status-driven, time-boxed, skeptical of dissent — never let it reach the decision. A crowd was present; no mechanism was.

When to reach for this book

  • When designing any system that pools opinions: forecasting, voting, feedback, prediction, pricing.
  • When a team unanimously agrees too quickly — this book explains why that’s a warning sign, not a comfort.
  • When deciding whether to trust “the market,” “the poll,” or “what everyone is saying.”
  • Pairs naturally with Thinking, Fast and Slow: Kahneman explains individual error; Surowiecki explains when groups cancel it versus compound it.

Memorable ideas and lines

The useful idea is that a crowd is not wise because it is large. It becomes wise only when people bring different information, make independent judgments, operate close to the facts, and have a clean way to aggregate what they know.

That is why the same group can be brilliant in one design and foolish in another. Change the incentives, let one voice dominate, or force consensus too early, and the crowd stops being a distributed intelligence system.

The ox-weighing example is an image of independent estimates brought together by a clear rule. The useful question for a team is which parts of that information structure its own process preserves or breaks.

A crowd is only as wise as its information structure

The book’s central test is structural. Before trusting a collective answer, inspect where information came from, whether judgments formed independently, which local knowledge remains available, and how inputs become an output. Counting more voices does not repair correlated evidence or a process that rewards conformity. Wisdom is an achievement of design, not a property of a large group.

Aggregation mechanisms embed values as well as information

A market, vote, average, jury, and deliberating committee do not merely combine the same inputs in different ways. Each defines who participates, what influence means, which information is expressible, and what outcome counts. Choosing an aggregation rule is therefore part of the substantive decision, not a neutral technical step.

Prices can combine dispersed beliefs and incentives efficiently while excluding needs that lack purchasing power and external costs that never enter the transaction. Voting gives participants formal equality while compressing intensity, information, and minority risk into a rule for counting. An average uses all numeric inputs but can be distorted by extremes; a median is robust to extremes while discarding information about magnitude.

The mechanism should match the question. Forecasts may benefit from independent probabilities or estimates. Complex value conflicts require deliberation and legitimate authority, not only prediction accuracy. Safety decisions may need vetoes or thresholds because the cost of one kind of error is asymmetric.

Surowiecki’s four conditions remain necessary but not sufficient for a just decision. A crowd can accurately predict an outcome without having the right to impose it, and a legitimate democratic choice can remain factually mistaken. Good design separates epistemic questions—what is likely true—from normative and governance questions—what should be done and who may decide—then chooses a process capable of carrying each.