Behaviour / Analysis
Think alone, then talk: what nine doctors who never met can teach every team
Combine the answers of nine doctors who never met, and they diagnose far better than the typical doctor alone. Put people in a room together, and the picture gets more complicated — and more interesting.
The short version
- Combining the separate answers of nine clinicians raised diagnostic accuracy from 62.5% for the typical individual to 85.6% on teaching cases.
- It works best when people are of similar skill and don’t share the same blind spots; otherwise a crowd can be confidently wrong.
- Open discussion can lead groups astray — yet a huge experiment found that thinking alone first, then talking briefly in small groups, made a crowd more accurate still.
Nine doctors who never met
Human Dx is an online platform where doctors try to solve medical puzzles. A case appears — symptoms, history, test results — and each clinician, working alone, types in what they think is wrong.
Researchers looked at more than 1,500 of these cases, solved by over 2,000 users, and asked a simple question. What if, instead of trusting one doctor’s answer, you combined several?
They didn’t form teams. They took the separate answers people had already given and combined them afterwards, in groups of two, three, and up to nine. The typical clinician got the diagnosis right 62.5% of the time. Nine combined answers got it right 85.6% of the time. Even two beat one, and the combined groups also beat individual specialists working in their own field.
These were teaching cases, not real patients, and the authors disclosed conflicts of interest. But the size of the jump is hard to ignore — and so is the fact that these nine people never spoke.
It is worth being precise here, because two very different ideas are easy to confuse. A hivemind is minds leaning on each other’s judgment until they converge. What happened on Human Dx was the opposite: isolated judgments, kept apart, combined only at the end.
The crowd works best when nobody stands out
There is a catch, and a second study found it.
A team led by Ralf Kurvers took more than 20,000 real breast and skin cancer diagnoses made by over 140 doctors, and simulated what would happen if their verdicts were combined. When the doctors were of fairly similar skill, the group beat even its best member.
But when one doctor was clearly better than the rest, combining everyone’s view dragged the answer down. In those groups, simply following the strongest doctor did better.
A crowd seems wise when its members are about equally good. When one person really knows more, the crowd may drown them out.
When the crowd is confidently wrong
There is a second condition, and it matters even more. Combining answers only helps when people are more often right than wrong, and when their mistakes are their own.
If most experts share the same blind spot — the same training, the same habits, the same missing information — combining them doesn’t cancel the error. It averages it, and makes it look more certain. On the hardest cases, where most people lean the wrong way, a crowd can be confidently wrong.
So pooling is not magic. It works when people are independent, roughly matched, and not all wrong in the same direction.
What happens when the jury goes into the room
Most real teams don’t pool. They talk. And talking, the research suggests, can change the answer in ways nobody intended.
In a study of more than 500 mock juries, over 3,000 people each wrote down how much a defendant should pay in punitive damages — before deliberating. Then the juries discussed and agreed on an amount. If discussion simply averaged people’s views, the jury’s figure would land near the middle. It didn’t. Awards after discussion were systematically higher than what the typical juror had wanted beforehand.
Punitive damages have no “correct” answer, so this shows a shift rather than an error. But it is a warning: a conversation can move a group away from where its members started.
The fact that never gets mentioned
Imagine a team where each person knows something the others don’t — a detail from a home visit, a note from a teacher, a line in an old report. The best decision depends on putting those pieces together.
Researchers have built puzzles like this in the lab for decades. A review of 65 studies, covering more than 3,000 groups, found a stubborn pattern: groups spend most of their time on what everyone already knows. When the right answer depends on a fact only one member holds, they rarely find it.
And in experiments in Denmark, Iran and China, people tended to give a partner’s opinion about equal weight to their own — whether or not the partner was better at the task, even with feedback and money at stake.
Then 5,180 people talked — and got better
So should groups never talk? A remarkable experiment suggests that is the wrong lesson.
In 2018, researchers asked a live audience of 5,180 people general-knowledge questions, such as the height of the Eiffel Tower. Everyone answered alone first. Then they split into groups of five, discussed briefly, and agreed on a group answer. Combining the answers of just a handful of these small debates beat averaging the whole crowd’s individual guesses.
Older work points the same way for some kinds of problems. Groups of three to five beat even their best member on logic puzzles — the kind where, once someone finds the right answer, everyone else can see it is right.
The lesson may not be “don’t talk”. It may be “think alone first, then talk in small groups, then combine”.
There is a limit, and it matters for real casework. Talking helps most when a correct answer is recognisable once someone says it. Many decisions about families or patients aren’t like that: the right answer may not be obvious even when it’s on the table.
So what is a case conference really doing?
None of these studies looked at child welfare. Online puzzles, mock juries and quiz questions are not case conferences, and we don’t want to overstate the link.
But Denmark offers a hint. In 2002, researchers studied how child and youth caseworkers assessed cases. In 2024, another team repeated the study and found assessments had become more uniform — shaped, they said, by collective discussion and more systematic use of knowledge.
That could be very good news. The trouble is that the study recorded each group’s shared answer, not what individuals thought first. So it can’t tell us whether the groups became more accurate, or simply more alike.
What would change our mind: the study nobody has run
The test is easy to describe. Take a set of real case files and try three designs: caseworkers judging alone with their answers combined; teams talking the cases through; and caseworkers judging alone first, then discussing in small groups, then combining. Compare all three with a careful expert review.
If open team discussion did as well as the other two, most of this article would matter much less. As far as we know, that comparison has never been made.
Sources
- Barnett, Boddupalli, Nundy et al. (2019). Comparative accuracy of diagnosis by collective intelligence of multiple physicians vs individual physicians. JAMA Network Open 2(3):e190096.
- Kurvers, Herzog, Hertwig et al. (2016). Boosting medical diagnostics by pooling independent judgments. PNAS 113(31):8777–8782.
- Schkade, Sunstein & Kahneman (2000). Deliberating about dollars: the severity shift. Columbia Law Review 100:1139.
- Lu, Yuan & McLeod (2012). Twenty-five years of hidden profiles in group decision making: a meta-analysis. Personality and Social Psychology Review.
- Mahmoodi, Bang, Olsen et al. (2015). Equality bias impairs collective decision-making across cultures. PNAS 112(12):3835–3840.
- Navajas, Niella, Garbulsky, Bahrami & Sigman (2018). Aggregated knowledge from a small number of debates outperforms the wisdom of large crowds. Nature Human Behaviour 2(2):126–132.
- Laughlin, Hatch, Silver & Boh (2006). Groups perform better than the best individuals on letters-to-numbers problems: effects of group size. Journal of Personality and Social Psychology 90(4).
- Bengtsson, Kamsteen & Møller (2025). Fra vilkårlige vurderinger til ensartet faglighed. Økonomi og Politik.