In September 1894 a cleaner at the German embassy in Paris, who was secretly working for French army intelligence, handed over a note she had found torn into six pieces. Pieced together, it listed military documents that someone was offering to the Germans. It had no signature and no date, but its meaning was plain. Someone inside the French army was selling secrets.
Suspicion soon settled on Captain Alfred Dreyfus, an artillery officer on the General Staff. He was Jewish, at a time when much of the French army disliked Jews, and there was no real evidence against him. The minister of war called him in on an empty file. One handwriting expert compared his writing with the note and found many differences. Another man, Alphonse Bertillon, a police officer famous for a way of identifying criminals by measuring their bodies, and no expert on handwriting, came up with an answer to those differences: Dreyfus had disguised his own hand, copying himself badly on purpose. So every place where the writing didn’t match became one more sign of a clever spy.
In December 1894 seven judges found Dreyfus guilty, all seven. On 5 January 1895, in the courtyard of the military school in Paris, his badges were torn off and his sword was broken over a knee while he shouted, “Innocent, innocent! Long live France! Long live the army!” He was shipped to Devil’s Island, a bare rock off the coast of South America, to be held there for life.
In March 1896 a new head of army intelligence, Lieutenant-Colonel Georges Picquart, went through a fresh batch of papers from the same embassy. Among them was a letter-card from the German military attaché, never sent, addressed to a French officer called Major Ferdinand Walsin Esterhazy. Picquart got hold of letters Esterhazy had written, and saw that his handwriting matched the torn note. He read the secret file that had been shown to the judges in 1894, and was struck by how little was in it. Spying, it seemed, had gone on after Dreyfus was locked away.
His superiors didn’t want to know. The case had been judged, they said, and that was that. Picquart was watched, then sent away to a post in Tunisia. When the writer Émile Zola took up the case in 1898, in an open letter headed J’accuse…!, he called Picquart the one honest man in the affair, and noted that people were now saying he had forged the letter-card himself. Why would he? Zola asked, and added a detail that makes Picquart worth meeting: “The beauty of it is that he was, as it happens, an antisemite.”
Soldiers and scouts
Picquart shared the prejudice of the men around him, so he had every reason to believe what they believed. He followed the handwriting anyway. The writer Julia Galef, who told his story in a talk in 2016 and again in her book The Scout Mindset (2021), uses two old jobs in an army to explain the difference between him and his colleagues.
A soldier’s job is to defend. Something attacks his side, and he fights it off. A scout’s job is to go ahead and find out what is really there: where the river can be crossed, how many people are on the far hill, whether the bridge is still standing. A scout may hope to find a bridge in a handy place, but a scout who drew one on the map because it would be handy would get his army killed. What matters to a scout is that the map is right.
People treat their own beliefs in both ways. In soldier mode, a welcome piece of evidence is a friend to let in and an unwelcome one is an enemy to keep out. The question being asked, without anyone noticing it, is “How can I keep believing this?” Bertillon’s self-copying spy is the soldier at full stretch: a fact that pointed away from Dreyfus was turned around until it pointed at him. In scout mode the question is “What’s true here?”, and a fact counts for what it shows, whichever way that is.
None of this needs a person to be stupid or dishonest. Galef’s point, in that talk, is that soldier mode is mostly something we don’t notice ourselves doing, and that what makes a scout is less a matter of brains than of feelings: wanting to see what is there more than wanting your side to win. Psychologists have watched it happen in the laboratory. In a well-known experiment from 1979, people who were for or against the death penalty read about two studies, made up but presented as real, one finding that it deterred murder and one finding that it didn’t, with their methods described. Each side rated the study that agreed with them as the better-done one. When 51 such experiments, with more than 18,000 people, were pooled in 2018, the tilt was clear, and it was about the same size on both sides of politics. Neither side is exempt.
Looking only where you expect a yes
Soldier mode has a quieter form, which doesn’t defend a belief so much as never put it at risk.
In 1960 a psychologist at University College London, Peter Wason, gave 29 students a puzzle. He told each of them that the three numbers 2, 4, 6 followed a simple rule he had in mind. Their job was to find the rule by writing down other sets of three numbers. After each set, he would say whether it followed the rule or not. They could try as many sets as they liked, and they were to announce the rule only when they felt highly confident they had found it.
Most students thought of a rule at once, such as “even numbers going up by two”, and then tested it with sets that fit it, such as 8, 10, 12 or 20, 22, 24. Each time Wason said yes, and each yes made them surer. Then they announced their rule, and were told it was wrong. Wason’s rule was simply “any three numbers going up”. Only 6 of the 29 found it without first announcing a wrong one. Nine announced two or more wrong rules.
The trouble was where they looked. Every set they tried was one their own rule said would pass, so it couldn’t tell their rule apart from Wason’s, which said the same about it. The only sets that could have taught them anything were ones their rule said should fail: 1, 2, 3, or 5, 10, 200. The students who found the rule at their first try were the ones who had tried many more such sets.
There’s a fair point on the other side. Testing the cases you expect to fit is often a perfectly sensible way to check an idea, and later researchers argued that in real life it often serves people well. It goes badly wrong in one situation in particular: when your idea is narrower than the truth, as “up by two” was narrower than “going up”. Then every test you think of comes back yes, and you never find out.
The same thing happens on a phone. Someone wondering whether a remedy works types “does it cure diabetes” and finds pages that say it does. The scout’s move is to search for what would show it doesn’t work, and look at what the people who tested it found. The question that unlocks Wason’s puzzle works everywhere: if I were wrong, what would I expect to see, and have I looked?
Wrong, and glad to know it
Charles Darwin had a habit he called a golden rule. For many years, whenever he came across a published fact, an observation or a thought that went against his general results, he made a note of it “without fail and at once”. He had found that such facts were “far more apt to escape from the memory than favourable ones.” So he wrote down exactly the things his mind was keen to lose. Owing to this habit, he said, very few objections were raised against his views that he hadn’t at least noticed and tried to answer.
That’s the scout’s attitude to being wrong. To a soldier, being wrong is a defeat: ground lost to the enemy. To a scout it is news. If the map showed a bridge and there isn’t one, the most useful thing in the world is to find out before the army arrives. Every mistake noticed is a place where your map has just got better.
The opposite habit has been watched too. Over nearly twenty years, the psychologist Philip Tetlock collected tens of thousands of forecasts about world events from a few hundred experts in politics, economics and the news, and checked them against what happened. The experts who explained everything by one big idea, and said so with great confidence, did worse than those who drew on many small ones. And when their forecasts failed, the experts seldom said their theory had been wrong. They said the thing had almost happened, or would happen yet, or would have happened but for something nobody could have foreseen.
How good is your 80%?
Being a scout about the outside world is one thing. The harder map to keep honest is the one of how much you yourself know. A belief can be held at 60 or at 95, and the only way to find out whether your 95 deserves the name is to count: of all the times you said 95, how often were you right? The twenty questions of How Sure?, after Chapter 2, kept a record of exactly that, in this browser.
Twenty answers spread over six levels of sure are only a few at each, so this record is a first glimpse, not a verdict. Weather forecasters are judged on thousands of forecasts. But even a glimpse can surprise: a dot below the line at 90 or 100 means that your sure was surer than your knowledge. Experiments with questions of general knowledge have often found that pattern, and the useful thing about it is that it can be put right.
How do we know it can be put right? From a contest. In September 2011 a research agency of the American government, working to improve the country’s intelligence, set five university teams competing to forecast world events, such as “Will Russia test an intercontinental missile by this date?” or whether Italy’s prime minister would leave office before 2012. One team, the Good Judgment Project, led by Tetlock and Barbara Mellers, recruited thousands of volunteers through blogs, alumni groups and word of mouth. To join, they needed a university degree and had to sit about two hours of tests.
Each forecast was a chance, from 0 to 100, and each was scored when the question closed. The scoring was harsh on overconfidence. Say 90% and be right, and you lose almost nothing; say 90% and be wrong, and you lose about eighty times as much. A forecaster can’t do well by sounding sure, only by being sure when it’s deserved.
And the team ran experiments inside the contest, choosing at random which volunteers got what. Some were given a short course on reasoning with chances: start from how often things of this kind usually happen, average several estimates where you have them, beware of overconfidence and of seeking only agreement. Some were put in teams that shared what they had found and argued over it, while others worked alone. In the team’s report on the first two years, both helped: the trained beat the untrained, and teams beat people working alone. A later paper looked back over all four years of the contest and put the training’s gain at about 6 to 11 percent, from a course lasting less than an hour: a small change from a small dose.
The two-year report had one more result. At the end of the first year it took the best 2 percent of its forecasters, sixty people, and put them together in teams of their own. If they had only been lucky, they should have sunk back toward the average the next year, as lucky streaks do. They didn’t. They beat every other group by a wide margin, a strong sign that forecasting is partly a skill. And they changed their minds a great deal. The ordinary forecasters gave about one and a half forecasts per question. These gave nearly eight, nudging their numbers again and again.
So the people who did best were not those with the most confident opinions, or those who never made mistakes. They were people who wrote their beliefs down as numbers, kept score, worked with others who could catch their errors, and kept updating. Some of it was talent: the team’s own verdict was that its best forecasters were partly found and partly made. But the training, the teams and the keeping score made forecasters better on average, and anyone can take those up.
Judging yourself
If people misjudge how sure to be about Cairo and Kathmandu, they can certainly misjudge themselves. One study about this has become famous far beyond psychology, and it deserves a careful look, both for what it found and for what has been made of it.
In 1999 Justin Kruger and David Dunning, at Cornell University in the United States, gave students tests of humor (rating jokes against the ratings of professional comedians), logical reasoning and English grammar, then asked each of them how well they thought they had done. Then they sorted the students into four groups by their real scores. Students in the bottom group scored, on average, better than only 12 percent of their classmates, but placed themselves above 62 percent of them. The top group did the opposite and rated themselves a little lower than they were. Kruger and Dunning’s explanation was that the skill you need to do a task well is often the same skill you need to see that you’ve done it badly. A student who doesn’t know a grammar rule can’t spot where they broke it.
This became the Dunning–Kruger effect, and it’s usually shown as a curve: confidence shooting up to a peak among beginners, crashing as they learn more, then rising slowly with real skill. That curve isn’t in the paper. Its graphs show only the four groups, each with its real place in the class and the place it guessed.
And those graphs raise a question. Look at one of the tests, on logical reasoning, from 20 questions. Afterward the students guessed how many they had got right.
Both groups guessed about 14. Suppose nobody could judge their own score at all, and everyone simply guessed a bit above the middle. The weakest students would then seem to overrate themselves hugely and the best to underrate themselves, without anyone being more blind than anyone else. Those at the bottom have a long way to fall short of an ordinary guess, and those at the top a long way to rise above it. This drift toward the middle is Chapter 19’s regression to the mean, and the arithmetic of it alone draws a picture very like the famous one. Researchers have since made the same pattern out of pure random numbers, where by design nobody knows anything about themselves.
So is the effect only arithmetic? Here the experts genuinely disagree. Kruger and Dunning saw this objection coming and answered it in the same paper: the bottom group missed by more than the top group did, which, they argued, the drift toward the middle couldn’t produce on its own. Three researchers in the United States repeated two of the studies in 2021 with about 4,000 people each, far more than the first study, and found that weaker performers really were worse at telling which of their own answers were right. But a 2020 study of nearly a thousand people, comparing how clever they thought they were with an intelligence test, found none of the patterns the effect should leave, and its authors concluded that it may be much smaller than reported. Others find that people’s view of their own skill, measured carefully, mostly does track their real skill, with experts judging themselves better than beginners.
What holds up is this. People are often poor judges of how good they are, the weakest overrate themselves most, and the best are a little too modest. How much of the pattern is a real blind spot at the bottom, and how much is the arithmetic of noisy guesses, is still argued about, so the next time someone uses the effect to sneer at people who disagree with them, it’s fair to say so. One finding from the original paper is the hopeful one. Students in the bottom group who were then given a short lesson in the logic they had got wrong became much better at judging how badly they had done, though they still rated themselves too high.
Which brings the chapter back to the forecasters. The cure for a poor map of yourself isn’t to feel humble in general. It’s to make calls you can check, write down how sure you were, and count.
As for Dreyfus, his story ended slowly. Esterhazy was tried early in 1898 and acquitted after two days. Picquart was arrested, and later charged with forging the letter-card. Then in August 1898 the officer who had forged a letter to shore up the case against Dreyfus confessed, and killed himself the next day. Dreyfus was brought back from Devil’s Island in 1899 and convicted again, then pardoned. Only on 12 July 1906 did France’s highest court cancel the verdict for good. Dreyfus went back into the army as a major. Picquart, made a general, became minister of war that October.
Nothing in Picquart made him a likely hero. He didn’t like Jews, he was loyal to his army, and the evidence he found hurt both. What he had was a scout’s attachment to the map: when the handwriting pointed somewhere he didn’t expect, he looked where it pointed, and kept looking when he was told to stop. Most of us will never face a choice that hard. We face small ones every day, in a forwarded message, a family argument, a test we thought we had done well in. Each time the question is the same. Not “how can I keep believing this?”, but “what is really there?”
Found a mistake? Please tell us.
