Groups & Conformity, Deeper: Smarter or Dumber Is a Matter of Structure

When people who already lean the same way discuss an issue, their position does not converge toward the middle — the whole group takes one more step in the direction it was already facing. The sense of consensus and the degree of extremity rise together, and from the inside the two are indistinguishable.

Two mechanisms run in parallel. Persuasive arguments: the arguments actually voiced in the room skew toward the initial leaning and are more numerous and fresher on that side, so you hear reasons you hadn't thought of and raise your bid. Social comparison: people want to sit slightly above average on whatever the group values — and when everyone does that, the mean drifts.

The strongest evidence isn't from the lab. Sunstein and Schkade analyzed thousands of U.S. federal appellate panel decisions and found that the same judge votes markedly more extremely on a panel of three appointees of the same party than on a mixed panel. Homogeneity itself is the variable.

Exercise: before an important meeting, have everyone anonymously write down "position + confidence 0–100," then write it again afterward. Same direction, confidence broadly up = polarization, not consensus.

The more people present, the lower the odds that any given individual steps in — yet real street data shows the odds that someone steps in actually rise with crowd size. Both hold; the difference is whether you're asking about the individual or the group.

Darley and Latané's five-step model shows why action breaks so easily: all five steps must clear, and any one blocked means nothing happens.

Philpot et al. (2020, American Psychologist) coded 219 real conflicts from public CCTV footage in the Netherlands, the UK, and South Africa: someone intervened in over 90% of incidents, and the probability of intervention rose with the number of bystanders. The conditions that most reliably produce inaction in the lab — ambiguity, no real danger, strangers — are precisely the ones real emergencies least often satisfy. (Worth correcting too: the canonical "38 witnesses did nothing" account of the Kitty Genovese case was debunked by Manning et al.'s archival work in 2007; it was a journalistic construction.)

Exercise: scroll back through any work channel from last week and count how many "someone should do this" messages were never picked up. That count is your team's diffusion-of-responsibility water level.

Anonymity doesn't make people normless; it makes them switch from personal norms to the norms of the group they're currently in. Under the same anonymity, one group turns cruel and another turns generous — the variable isn't anonymity, it's the norm.

The classic account (the Zimbardo lineage) runs: anonymity + group → reduced self-awareness → impulse release. Postmes and Spears' 1998 meta-analysis found weak support for that prediction; the data fit the SIDE model better — anonymity reduces the salience of personal identity while raising social identity, so behavior tracks that group's situational norms more tightly. Archival work on the Stanford Prison Experiment (Le Texier, 2019) likewise shows guards received explicit coaching rather than spontaneously losing control (see Day 22).

Exercise: pull a few of your messages from the past week in three different groups and read them side by side. Which of these would I not say in the other two, and why?

The same group can be more accurate than its smartest member, or dumber than its most mediocre one. The dividing line isn't member quality — it's whether judgments were formed independently.

Aggregation works because individual errors point in different directions and cancel. Scott Page's diversity prediction theorem writes this as an identity: collective error = average individual error − prediction diversity. More diversity, more accuracy.

And diversity gets eaten by social influence. Lorenz et al. (2011, PNAS) had participants repeatedly estimate factual quantities while showing some groups others' estimates: the spread of answers narrowed sharply, collective accuracy did not improve, and individual confidence went up. That is exactly how groupthink is manufactured — the feeling of convergence rises while the information content falls.

Exercise: revisit your team's last three decisions and ask of each — who spoke first, and did anyone commit a judgment to writing before hearing others? The answers usually explain the outcome.