When I started studying collective intelligence, I believed something that now embarrasses me slightly. I believed that more minds were better than fewer, that diversity was a reliable good, and that the main obstacle to smarter groups was simply getting enough of the right people into the room. Aggregate enough perspectives, I thought, and wisdom would emerge more or less on its own.
A few years of study and fieldwork have corrected most of this. Not all at once, and not comfortably. Here is what I got wrong, and what replaced it.
I Thought Aggregation Was the Mechanism
The textbook version of collective intelligence rests on a clean idea: if individuals make independent errors, averaging across many judgments cancels those errors out, and the group lands closer to the truth than most of its members. This is real. The mathematics is sound. I took it as the general case.
It is the special case. The mathematics depends on independence, and independence is exactly what real groups lack. People hear each other. They watch who holds status. They feel the pull to agree once a few others have. The errors stop being random and start pointing the same way. What I had treated as the default mechanism of collective intelligence turns out to require conditions that most groups violate the moment they start talking to one another.
What replaced the belief: groups do not usually aggregate. They negotiate. The outcome reflects who spoke first, who carried authority, and whose knowledge was treated as relevant, at least as much as it reflects the quality of the available information.
I Thought Diversity Was Always Good
Diversity is central to every argument for collective intelligence, and I treated it as an unqualified asset. More cognitive diversity, better outcomes. It seemed obvious.
The fieldwork complicated it. Diversity only helps when the group has a way to surface and use the different perspectives it contains. A diverse group with no mechanism for dissent, or with steep power differences, performs worse than its composition suggests, because the diversity is present but inaccessible. The varied perspectives exist in the room and never make it into the decision. I had been counting diversity as if merely having it was enough. What matters is whether the structure of the group lets that diversity do any work.
This shifted where I look for the problem: away from who is in the group and toward whether the group is built to hear them.
I Thought It Was About Intelligence
This is the one that took longest to see. The phrase “collective intelligence” had me looking for cognitive explanations. Better reasoning, better information, better aggregation of judgment. I framed failures as breakdowns in the group’s thinking.
Most of the failures I actually encountered were social and political rather than cognitive. A farming community that rejected a sound recommendation was weighing trust in the institution making the recommendation, memory of past schemes that had failed, and the social cost of cooperating with neighbours they did not fully trust. The relevant knowledge was about power, history, and risk, held by people the formal process had not thought to ask, and none of it fits in a model.
Calling this a problem of collective intelligence almost misses the point. The intelligence was there. What shaped the outcome was the distribution of voice and authority, which is a question about politics, not cognition.
What I Kept
I did not end up disillusioned with the field. If anything, the corrections made it more interesting. The clean mathematical version is still useful as a baseline, a description of what groups could achieve if the social conditions allowed it. The gap between that baseline and what groups actually do is where the real questions live.
What I kept is the conviction that groups can be smarter than their members, and that this is worth pursuing. What I lost is the assumption that it happens automatically. Collective intelligence is not a property that emerges whenever you assemble enough people. It is something a group either builds or fails to build, through the structures it uses to surface knowledge, distribute voice, and let dissent survive contact with consensus.
Getting this wrong first was probably necessary. The clean version is what drew me in. The messy version is what kept me.