In 1906, Francis Galton observed a weight-guessing contest at a county fair. No individual correctly guessed the weight of an ox — but the average of all guesses was within 1 pound of the correct weight. This "wisdom of crowds" effect has been replicated across thousands of studies.

The conditions for collective intelligence are specific: diversity of opinion (each person has private information), independence (opinions aren't determined by others), decentralization (no one can dictate the answer), and aggregation (a mechanism for combining individual judgments).

When these conditions hold, groups dramatically outperform even their best individual member. Prediction markets consistently beat expert forecasters. Wikipedia, despite anyone being able to edit it, is remarkably accurate on most factual topics. Averaged weather forecasts outperform any single model.

But when the conditions break down, crowds become mobs. Social media destroys independence (everyone sees everyone's opinion). Centralized algorithms destroy diversity (showing everyone the same trending content). Echo chambers destroy opinion diversity. The infrastructure for collective intelligence is being systematically undermined by the very platforms that could enable it.

Groupthink (Irving Janis, 1972): cohesive groups prioritize consensus over critical evaluation. Members self-censor dissent, share an illusion of unanimity, and collectively rationalize poor decisions. The Bay of Pigs invasion, the Challenger disaster, and countless corporate failures followed this pattern.

Information cascades: when people make decisions sequentially and can observe prior choices, early choices disproportionately influence later ones. If the first three restaurant reviewers give five stars (perhaps by chance), subsequent reviewers anchor to that rating, creating a self-reinforcing signal disconnected from quality.

Polarization: group deliberation tends to make individual positions more extreme, not more moderate. If a group leans slightly toward risk, discussion makes them riskier. If they lean slightly conservative, discussion makes them more conservative. This is why political discussions in homogeneous groups radicalize rather than moderate.

The Dunning-Kruger cascade: in groups where expertise varies, the least competent members are often the most confident, and their confidence disproportionately influences group decisions. Expertise doesn't correlate with airtime in most group settings.

The key insight is that collective intelligence is a design problem, not a character problem. The same people can produce wisdom or madness depending on the system they're in:

Prediction markets: participants bet real money on outcomes, creating accountability and incentivizing accuracy over signaling. They consistently outperform expert panels.

Deliberative polling: random citizens are given balanced information and facilitated discussion time. Their resulting opinions consistently differ from both raw polls and partisan positions — typically more nuanced and informed.

Red teaming: dedicated adversarial groups whose job is to find flaws in plans. The Catholic Church's "Devil's Advocate" during canonization. Military war-gaming. Security penetration testing. Institutionalized dissent prevents groupthink.

Delphi method: experts provide independent estimates, see anonymized group results, then revise. Iteration without social pressure preserves independence while enabling learning.

Wiki-style knowledge: anyone can contribute, anyone can edit, changes are tracked and reversible. The structure makes it self-correcting over time — errors are corrected by the community faster than they're introduced.

The common thread: successful collective intelligence systems protect independence, encourage diverse input, create accountability for accuracy, and provide mechanisms for aggregation and error correction.