The Crowd Needs Strangers

An ox at Plymouth

At the West of England Fat Stock and Poultry Exhibition in Plymouth, a fat ox stood in a pen, and visitors paid sixpence each to guess what it would weigh once it had been slaughtered and dressed. Francis Galton borrowed the tickets afterward. He threw out thirteen he could not read and kept 787. Many came from butchers and farmers, and many from people who had never weighed anything heavier than a parcel. Nobody got it exactly right. The middle guess was 1,207 pounds. The average, worked out a few weeks later, was 1,197. The ox dressed at 1,198.

Galton published it in Nature in March 1907 under the title Vox Populi, the voice of the people. He had not expected much of the people. He admitted that the result reflected better on democratic judgment than he would have guessed.

Why the crowd was right

The crowd was not wise. Each ticket held the true weight plus a mistake. Some people guessed high because the ox looked enormous from where they stood. Some guessed low because they were thinking of last year’s animal, or of the price of beef. When you add up 787 mistakes that point in different directions, they eat each other. What is left is the ox.

That is the whole principle, and the formula at the top is only its bookkeeping. Every guess carries two kinds of error. The private kind is yours alone: your angle, your memory, your mood that morning. The shared kind is the error the whole crowd has in common: the same misleading sign on the pen, the same rumor in the beer tent, the same source. Add more people and the private errors cancel, faster and faster. The shared error does not cancel at all. It is the same in every guess, so the average keeps all of it.

So every crowd has a floor. However large it grows, it can never be more accurate than its shared error allows. The Marquis de Condorcet found a version of this for juries in 1785: if each juror is a little better than a coin toss, a majority becomes nearly certain as the jury grows. The word doing all the work is independent. Remove it and a jury of a thousand is a jury of one.

A crowd is never wiser than its mistakes are different.

We have known for some time how fragile this is. In 2011 a team at ETH Zurich asked people simple factual questions and let some of them see what the others had answered. Seeing the crowd narrowed everyone’s guesses without making the average any better, and it made people more confident. Agreement had been manufactured, and it felt exactly like knowledge.

What AI changes

AI changes the one term in the formula nobody used to worry about. For most of history the shared error stayed small, because people’s sources were scattered. Even broadcast television only told us what to look at. It never finished our sentences. Now a large share of the people in any room will have asked one of a handful of models before they walk in. The model may well be better than any of them alone. That is not the point. The point is that its mistake is now inside every guess.

The research has begun to show this from both sides. In 2024 two economists gave some writers story ideas from a language model. Readers judged those stories better, especially the ones from the weakest writers. The stories also came out more like each other. Three years earlier, Jon Kleinberg and Manish Raghavan had proved something sharper: when competing decision-makers all adopt the same algorithm, their collective decisions can get worse even when that algorithm is more accurate than each of them alone. Better parts, worse whole.

Here is the consequence most people miss. The danger is not that AI is wrong. It is that AI is wrong in the same place for everyone at once, and the crowd can no longer tell. A room that used to disagree and land near the truth now agrees and lands wherever the model landed. The quiet cancelling machine that made markets, juries, elections and committees work simply stops, and nothing about the room looks broken. It looks more confident than ever.

It also moves value somewhere unexpected. When everyone has a brilliant assistant, intelligence is no longer the scarce thing in the room. The scarce thing is a mistake nobody else is making. The most useful person becomes the stranger: the one whose errors come from somewhere else.

One clarification, because it sounds like it contradicts what I wrote this morning. In The Cost Is in the Pairs I argued that agents working together should share one record. That is coordination, and coordination wants shared state. This is judgment, and judgment wants separate minds. Share the record. Never share the guess.

Five predictions

1. “Write first, then ask” becomes a rule.

   Teams that make estimates, forecasts and hiring calls will adopt one simple habit: everyone commits a private answer before anyone consults an AI or a colleague. Planning poker and the Delphi method already do this for people. It will become ordinary hygiene for anything that matters, the way washing hands became ordinary before surgery.

2. A second opinion must come from a different mind.

   A second reading by the same model is not a second opinion. Serious decisions in medicine, lending, law and insurance will start to require that the check come from a different model family, a different team, or a person who never saw the first answer. “How many independent minds touched this?” becomes an audit question.

3. Correlated mistakes become the new crash.

   Markets, hiring and admissions will be more accurate on ordinary days and wrong together on rare ones. The failure will look like a stampede: thousands of delegates selling the same stock, rejecting the same applicant, booking the same restaurant, all in the same hour. Expect circuit breakers built for agreement, not only for speed.

4. Personal AI beats the company assistant at judgment.

   A company that hands every employee the same assistant with the same instructions has bought a crowd of one. A company whose people bring their own AI, each shaped by its owner’s memory, notes and history, keeps the private errors that make collective judgment work. By the early 2030s the firms with the best forecasts will be the ones with the most different delegates, not the single smartest one.

5. The stranger gets paid.

   The person who did not ask, who read the original instead of the summary, who brings the odd book and the unpopular hunch, becomes a role people hire for. Not because that person is right more often. Because when they are wrong, they are wrong somewhere else.

Why I tell my students to bring their own

I have argued for a long time that every business should have a Bring Your Own AI policy, and I usually make the case on ownership and portability. This is the deeper reason. My delegate has read my notes, my books and my mistakes. It carries my memory and my skills. It is a little strange in exactly the ways I am strange. When I tell it to decide, it decides like me, not like everyone. That is not a defect to train out of it. It is the reason it is worth asking. So when I tell my students to bring their own AI, I am really telling them to stay strangers. A room full of strangers can weigh an ox. A room full of copies can only agree.

The claim ledger

Already true

Errors cancel only when they are independent; that is arithmetic, not opinion. A large share of people now consult a small number of models for estimates, drafts and decisions. Experiments already show AI help making individual work better and collective work more alike, and shared information narrowing a crowd without making it more accurate.

What has to happen

Model answers have to stay correlated across vendors and users, rather than drifting apart as personal memory and personal instructions grow. And institutions have to keep deciding by pooling many judgments, so that lost independence actually shows up in outcomes instead of hiding in them.

Where I am probably wrong

Personal AI may fix this by itself. If delegates shaped by their owners’ lives turn out as different from each other as their owners are, the shared error shrinks and the crowd heals without anyone trying. If by 2031 crowds of AI-assisted people forecast as well as unassisted crowds did on the same kinds of questions, the floor I am worried about never rose.

Back to the table on Tuesday. Five notes, one number. The fix costs nothing. Write your own number first, in your own hand, before you ask anyone, human or machine. Then ask, and put your note on the table beside everyone else’s. The average of five strangers is worth more than the confidence of five copies.

Be the stranger in the room.

Background

• Francis Galton, “Vox Populi,” Nature 75 (1907), 450–451, and his follow-up letter “The Ballot-Box”
• Marquis de Condorcet, Essai sur l’application de l’analyse à la probabilité des décisions rendues à la pluralité des voix (1785), the source of the jury theorem
• Jan Lorenz, Heiko Rauhut, Frank Schweitzer and Dirk Helbing, “How social influence can undermine the wisdom of crowd effect,” PNAS (2011)
• Jon Kleinberg and Manish Raghavan, “Algorithmic monoculture and social welfare,” PNAS (2021)
• Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances (2024)
• James Surowiecki, The Wisdom of Crowds (2004)
• The Common Denominator
• The Signature Is the Product
• Better Anyway