Conditions for Crowd Wisdom

Groups can make better decisions when people contribute useful information from different sources. Private first answers can help preserve those differences, but people may still share the same mistaken evidence. Try this: Collect private estimates and ask what evidence each person used before discussing the results.

For this lesson, use an editorial checklist with four checks: diversity, meaning different relevant information or perspectives; independence, meaning judgments are not simply copied from others; decentralization, meaning people can contribute local knowledge; and aggregation, meaning a rule for combining judgments. Treat these as practical guidance, not as conditions that are always required or that guarantee a correct answer.

Combining judgments can help when individual mistakes offset one another. But correlation—errors tending to move together—can limit that benefit. People answering privately may still rely on the same faulty report. The paper discusses a classical majority-vote result that assumes statistically independent guesses, each with a probability greater than one-half of being correct. This is a specific theoretical assumption, not something established merely by collecting answers separately. [1]

Kao and Couzin examine theoretical models in which individuals use environmental cues, meaning signals that provide information, and a majority vote selects between two options. In some modeled environments, small groups outperform larger ones because random variation in which cues members use and what they observe can sometimes prevent shared information from controlling the decision. Results depend on cue reliability, correlation and voting strategy; the models establish no universal best group size. These are model results, not human field evidence for the four checks. [1]

For a hypothetical project forecast, ask engineers privately for completion estimates and their reasons. An online poll that hides responses until submission records initial answers before participants see others' responses. If estimates concern the same scope and use the same units, an average can provide a starting summary. Check shared assumptions and relevant expertise before relying on it: averaging alone does not establish accuracy.

Private first answers need not mean permanent secrecy. The practical collection procedure suggested here is to record an initial answer before showing anyone else's views, then make any later discussion a separate step. This is an editorial suggestion for keeping track of how judgments change, not a result showing that a particular forecasting method is universally best.

• Theory: Kao and Couzin model collective choices using environmental cues and majority voting. Whether accuracy keeps improving with group size depends on cue reliability, information correlation and the strategy individuals use to select cues. [1]

• Model mechanism: In parts of the multiple-cue model, random variation in small groups can reduce the control of correlated information and improve accuracy relative to larger groups. This advantage is conditional, not a general recommendation to add randomness. [1]

• Scope and evidence: The second model, involving fluctuations in cue reliability, often favors either a solitary individual or a very large group within the parameter space examined. The excerpt also discusses earlier animal laboratory experiments; its theoretical findings do not constitute human field validation of the four-part checklist. [1]

The supplied evidence concerns simplified models and discussion of earlier studies. It does not establish four necessary-and-sufficient conditions, a universal group-size cutoff, or a ranking of hidden polls, Delphi rounds and markets. Private responses prevent exposure to others' answers during collection, but do not ensure statistically independent information. Averaging is a practical option for comparable numerical estimates; its usefulness depends on information quality, shared errors and the task. The project forecast is a hypothetical teaching example, not a reported experiment.