The short version

Take people who lean the same way, mildly. Put them in a room and let them discuss it properly — no pressure, no bullying, nobody silenced.

They come out leaning the same way, further. Not the average of what they walked in with. Past the end of it. The person who arrived most convinced typically leaves more convinced still, and the moderates have moved out to where the convinced ones used to be.

Where the name comes from

It was found by accident and named wrongly first, which is a decent guide to how counterintuitive it was.

In 1961 a graduate student, James Stoner, was checking the then-obvious assumption that groups are more cautious than individuals — committees hedge, that’s what committees are for. He found the opposite: groups reliably endorsed riskier options than their members had chosen alone. For most of a decade this was studied as the risky shift.

Then the shifts started coming back the other way. On questions where individuals leaned cautious to begin with, groups came out more cautious. The effect was never about risk at all. Moscovici and Zavalloni named what was actually happening in 1969: the group is a polarizer. Discussion moves people further along whichever way they were already facing.

Why it happens

Two accounts, both with real support, and in practice they run together.

You hear a lopsided sample of the arguments. In a room leaning one way, most of the reasons that get spoken favour that way — and some of them will be good ones you hadn’t thought of. You genuinely learn something, and what you learn is systematically one-sided. Nobody has to be persuaded dishonestly. The persuasion is real; it’s the sample that’s skewed.

You find out where the middle is, and you don’t want to be behind it. Before the discussion you had no idea how your position compared. Afterwards you do, and being slightly more committed than average is a comfortable place to stand in a room that values the position. Everybody adjusting slightly outward moves the whole distribution outward.

This is why the effect is strongest in groups that formed around the shared view — a campaign, a fandom, a channel, a team hired for fit. The sorting happens first, and then the discussion, which everyone experiences as the free exchange of ideas, does the rest.

This is not group polarization

• Being persuaded by good arguments. If genuinely new and genuinely strong reasons arrived and you moved, that’s learning, and it’s what discussion is for. The problem is the systematic skew in which arguments turn up, not the fact that arguments work.
• Groupthink. Groupthink is a group suppressing real doubts to protect a consensus it values. Here there’s no dissent to suppress and nobody is protecting anything — the group agreed on the way in and simply moves. Groupthink explains bad agreement; polarization explains extreme agreement, which is a different failure.
• The Abilene paradox. There the group does something nobody wanted, through misread preferences. Here everybody genuinely wants it, more than they did an hour ago.
• A group reaching a strong conclusion that happens to be right. Polarization is a claim about movement, not about truth. A room can polarize toward a correct position. What it can’t tell you is whether it did, because the mechanism runs identically either way.

The test is about where the room landed relative to where it started: did the group end up further out than the most extreme person who walked in? Averaging is what discussion is supposed to do. Overshooting the maximum is the tell.

How to head it off

Everything useful here is about the composition of the room and the order of events, because once a like-minded group is discussing freely, the mechanism is already running.

• Get positions in writing before the discussion. You then have a record of where people actually started, which is the only way anyone will notice the shift afterwards — including you.
• Put somebody in the room who disagrees, and mean it. Not a token. The persuasive-arguments half of this is a sampling problem, and the only real fix for a skewed sample is a less skewed one.
• Assign the opposite case to a person, and give them time to prepare it. A devil’s advocate improvising for ninety seconds supplies weak arguments, which makes things worse — the room defeats them and comes out more certain than before.
• Be suspicious of a meeting where everybody left more confident and nobody changed their mind. That is the exact signature, and it is normally reported afterwards as a really good meeting.

The uncomfortable part is that none of this looks like a failure while it’s happening. There’s no villain, no silencing, no bad faith. Everyone is reasoning, and the reasoning is mostly sound. The room is a fair sample of nothing, and it feels like a fair sample of everything.