Thinking, Fast and Slow (28 page)

If subjective confidence is not to be trusted, how can we evaluate the probable validity of an intuitive judgment? When do judgments reflect true expertise? When do they display an illusion of validity? The answer comes from the two basic conditions for acquiring a skill:

• an environment that is sufficiently regular to be predictable
  
• an opportunity to learn these regularities through prolonged practice
  

When both these conditions are satisfied, intuitions are likely to be skilled. Chess is an extreme example of a regular environment, but bridge and poker also provide robust statistical regularities that can support skill. Physicians, nurses, athletes, and firefighters also face complex but fundamentally orderly situations. The accurate intuitions that Gary Klein has described are due to highly valid cues that es the expert’s System 1 has learned to use, even if System 2 has not learned to name them. In contrast, stock pickers and political scientists who make long-term forecasts operate in a zero-validity environment. Their failures reflect the basic unpredictability of the events that they try to forecast.

Some environments are worse than irregular. Robin Hogarth described “wicked” environments, in which professionals are likely to learn the wrong lessons from experience. He borrows from Lewis Thomas the example of a physician in the early twentieth century who often had intuitions about patients who were about to develop typhoid. Unfortunately, he tested his hunch by palpating the patient’s tongue, without washing his hands between patients. When patient after patient became ill, the physician developed a sense of clinical infallibility. His predictions were accurate—but not because he was exercising professional intuition!

Meehl’s clinicians were not inept and their failure was not due to lack of talent. They performed poorly because they were assigned tasks that did not have a simple solution. The clinicians’ predicament was less extreme than the zero-validity environment of long-term political forecasting, but they operated in low-validity situations that did not allow high accuracy. We know this to be the case because the best statistical algorithms, although more accurate than human judges, were never very accurate. Indeed, the studies by Meehl and his followers never produced a “smoking gun” demonstration, a case in which clinicians completely missed a highly valid cue that the algorithm detected. An extreme failure of this kind is unlikely because human learning is normally efficient. If a strong predictive cue exists, human observers will find it, given a decent opportunity to do so. Statistical algorithms greatly outdo humans in noisy environments for two reasons: they are more likely than human judges to detect weakly valid cues and much more likely to maintain a modest level of accuracy by using such cues consistently.

It is wrong to blame anyone for failing to forecast accurately in an unpredictable world. However, it seems fair to blame professionals for believing they can succeed in an impossible task. Claims for correct intuitions in an unpredictable situation are self-delusional at best, sometimes worse. In the absence of valid cues, intuitive “hits” are due either to luck or to lies. If you find this conclusion surprising, you still have a lingering belief that intuition is magic. Remember this rule: intuition cannot be trusted in the absence of stable regularities in the environment.

Feedback and Practice

Some regularities in the environment are easier to discover and apply than others. Think of how you developed your style of using the brakes on your car. As you were mastering the skill of taking curves, you gradually learned when to let go of the accelerator and when and how hard to use the brakes. Curves differ, and the variability you experienced while learning ensures that you are now ready to brake at the right time and strength for any curve you encounter. The conditions for learning this skill are ideal, because you receive immediate and unambiguous feedback every time you go around a bend: the mild reward of a comfortable turn or the mild punishment of some difficulty in handling the car if you brake either too hard or not quite hard enough. The situations that face a harbor pilot maneuvering large ships are no less regular, but skill is much more difficult to acquire by sheer experience because of the long delay between actions and their manoticeable outcomes. Whether professionals have a chance to develop intuitive expertise depends essentially on the quality and speed of feedback, as well as on sufficient opportunity to practice.

Expertise is not a single skill; it is a collection of skills, and the same professional may be highly expert in some of the tasks in her domain while remaining a novice in others. By the time chess players become experts, they have “seen everything” (or almost everything), but chess is an exception in this regard. Surgeons can be much more proficient in some operations than in others. Furthermore, some aspects of any professional’s tasks are much easier to learn than others. Psychotherapists have many opportunities to observe the immediate reactions of patients to what they say. The feedback enables them to develop the intuitive skill to find the words and the tone that will calm anger, forge confidence, or focus the patient’s attention. On the other hand, therapists do not have a chance to identify which general treatment approach is most suitable for different patients. The feedback they receive from their patients’ long-term outcomes is sparse, delayed, or (usually) nonexistent, and in any case too ambiguous to support learning from experience.

Among medical specialties, anesthesiologists benefit from good feedback, because the effects of their actions are likely to be quickly evident. In contrast, radiologists obtain little information about the accuracy of the diagnoses they make and about the pathologies they fail to detect. Anesthesiologists are therefore in a better position to develop useful intuitive skills. If an anesthesiologist says, “I have a feeling something is wrong,” everyone in the operating room should be prepared for an emergency.

Here again, as in the case of subjective confidence, the experts may not know the limits of their expertise. An experienced psychotherapist knows that she is skilled in working out what is going on in her patient’s mind and that she has good intuitions about what the patient will say next. It is tempting for her to conclude that she can also anticipate how well the patient will do next year, but this conclusion is not equally justified. Short-term anticipation and long-term forecasting are different tasks, and the therapist has had adequate opportunity to learn one but not the other. Similarly, a financial expert may have skills in many aspects of his trade but not in picking stocks, and an expert in the Middle East knows many things but not the future. The clinical psychologist, the stock picker, and the pundit do have intuitive skills in some of their tasks, but they have not learned to identify the situations and the tasks in which intuition will betray them. The unrecognized limits of professional skill help explain why experts are often overconfident.

Evaluating Validity

At the end of our journey, Gary Klein and I agreed on a general answer to our initial question: When can you trust an experienced professional who claims to have an intuition? Our conclusion was that for the most part it is possible to distinguish intuitions that are likely to be valid from those that are likely to be bogus. As in the judgment of whether a work of art is genuine or a fake, you will usually do better by focusing on its provenance than by looking at the piece itself. If the environment is sufficiently regular and if the judge has had a chance to learn its regularities, the associative machinery will recognize situations and generate quick and accurate predictions and decisions. You can trust someone’s intuitions if these conditions are met.

Unfortunately, associativentu memory also generates subjectively compelling intuitions that are false. Anyone who has watched the chess progress of a talented youngster knows well that skill does not become perfect all at once, and that on the way to near perfection some mistakes are made with great confidence. When evaluating expert intuition you should always consider whether there was an adequate opportunity to learn the cues, even in a regular environment.

In a less regular, or low-validity, environment, the heuristics of judgment are invoked. System 1 is often able to produce quick answers to difficult questions by substitution, creating coherence where there is none. The question that is answered is not the one that was intended, but the answer is produced quickly and may be sufficiently plausible to pass the lax and lenient review of System 2. You may want to forecast the commercial future of a company, for example, and believe that this is what you are judging, while in fact your evaluation is dominated by your impressions of the energy and competence of its current executives. Because substitution occurs automatically, you often do not know the origin of a judgment that you (your System 2) endorse and adopt. If it is the only one that comes to mind, it may be subjectively undistinguishable from valid judgments that you make with expert confidence. This is why subjective confidence is not a good diagnostic of accuracy: judgments that answer the wrong question can also be made with high confidence.

You may be asking, Why didn’t Gary Klein and I come up immediately with the idea of evaluating an expert’s intuition by assessing the regularity of the environment and the expert’s learning history—mostly setting aside the expert’s confidence? And what did we think the answer could be? These are good questions because the contours of the solution were apparent from the beginning. We knew at the outset that fireground commanders and pediatric nurses would end up on one side of the boundary of valid intuitions and that the specialties studied by Meehl would be on the other, along with stock pickers and pundits.

It is difficult to reconstruct what it was that took us years, long hours of discussion, endless exchanges of draft s and hundreds of e-mails negotiating over words, and more than once almost giving up. But this is what always happens when a project ends reasonably well: once you understand the main conclusion, it seems it was always obvious.

As the title of our article suggests, Klein and I disagreed less than we had expected and accepted joint solutions of almost all the substantive issues that were raised. However, we also found that our early differences were more than an intellectual disagreement. We had different attitudes, emotions, and tastes, and those changed remarkably little over the years. This is most obvious in the facts that we find amusing and interesting. Klein still winces when the word
bias
is mentioned, and he still enjoys stories in which algorithms or formal procedures lead to obviously absurd decisions. I tend to view the occasional failures of algorithms as opportunities to improve them. On the other hand,
I find more pleasure than Klein does in the come-uppance of arrogant experts who claim intuitive powers in zero-validity situations. In the long run, however, finding as much intellectual agreement as we did is surely more important than the persistent emotional differences that remained.

Speaking of Expert Intuition

A few years after my collaboration with Amos began, I convinced some officials in the Israeli Ministry of Education of the need for a curriculum to teach judgment and decision making in high schools. The team that I assembled to design the curriculum and write a textbook for it included several experienced teachers, some of my psychology students, and Seymour Fox, then dean of the Hebrew University’s School of Education, who was an expert in curriculum development.

After meeting every Friday afternoon for about a year, we had constructed a detailed outline of the syllabus, had written a couple of chapters, and had run a few sample lessons in the classroom. We all felt that we had made good progress. One day, as we were discussing procedures for estimating uncertain quantities, the idea of conducting an exercise occurred to me. I asked everyone to write down an estimate of how long it would take us to submit a finished draft of the textbook to the Ministry of Education. I was following a procedure that we already planned to incorporate into our curriculum: the proper way to elicit information from a group is not by starting with a public discussion but by confidentially collecting each person’s judgment. This procedure makes better use of the knowledge available to members of the group than the common practice of open discussion. I collected the estimates and jotted the results on the blackboard. They were narrowly centered around two years; the low end was one and a half, the high end two and a half years.

Then I had another idea. I turned to Seymour, our curriculum expert, and asked whether he could think of other teams similar to ours that had developed a curriculum from scratch. This was a time when several pedagogical innovations like “new math” had been introduced, and Seymour said he could think of quite a few. I then asked whether he knew the history of these teams in some detail, and it turned out that he was familiar with several. I asked him to think of these teams when they had made as much progress as we had. How long, from that point, did it take them to finish their textbook projects?

He fell silent. When he finally spoke, it seemed to me that he was blushing, embarrassed by his own answer: “You know, I never realized this before, but in fact not all the teams at a stage comparable to ours ever did complete their task. A substantial fraction of the teams ended up failing to finish the job.”

This was worrisome; we had never considered the possibility that we might fail. My anxiety rising, I asked how large he estimated that fraction was. Rw l剢 sidering t20;About 40%,” he answered. By now, a pall of gloom was falling over the room. The next question was obvious: “Those who finished,” I asked. “How long did it take them?” “I cannot think of any group that finished in less than seven years,” he replied, “nor any that took more than ten.”

I grasped at a straw: “When you compare our skills and resources to those of the other groups, how good are we? How would you rank us in comparison with these teams?” Seymour did not hesitate long this time. “We’re below average,” he said, “but not by much.” This came as a complete surprise to all of us—including Seymour, whose prior estimate had been well within the optimistic consensus of the group. Until I prompted him, there was no connection in his mind between his knowledge of the history of other teams and his forecast of our future.

Our state of mind when we heard Seymour is not well described by stating what we “knew.” Surely all of us “knew” that a minimum of seven years and a 40% chance of failure was a more plausible forecast of the fate of our project than the numbers we had written on our slips of paper a few minutes earlier. But we did not acknowledge what we knew. The new forecast still seemed unreal, because we could not imagine how it could take so long to finish a project that looked so manageable. No crystal ball was available to tell us the strange sequence of unlikely events that were in our future. All we could see was a reasonable plan that should produce a book in about two years, conflicting with statistics indicating that other teams had failed or had taken an absurdly long time to complete their mission. What we had heard was base-rate information, from which we should have inferred a causal story: if so many teams failed, and if those that succeeded took so long, writing a curriculum was surely much harder than we had thought. But such an inference would have conflicted with our direct experience of the good progress we had been making. The statistics that Seymour provided were treated as base rates normally are—noted and promptly set aside.

We should have quit that day. None of us was willing to invest six more years of work in a project with a 40% chance of failure. Although we must have sensed that persevering was not reasonable, the warning did not provide an immediately compelling reason to quit. After a few minutes of desultory debate, we gathered ourselves together and carried on as if nothing had happened. The book was eventually completed eight(!) years later. By that time I was no longer living in Israel and had long since ceased to be part of the team, which completed the task after many unpredictable vicissitudes. The initial enthusiasm for the idea in the Ministry of Education had waned by the time the text was delivered and it was never used.

This embarrassing episode remains one of the most instructive experiences of my professional life. I eventually learned three lessons from it. The first was immediately apparent: I had stumbled onto a distinction between two profoundly different approaches to forecasting, which Amos and I later labeled the inside view and the outside view. The second lesson was that our initial forecasts of about two years for the completion of the project exhibited a planning fallacy. Our estimates were closer to a best-case scenario than to a realistic assessment. I was slower to accept the third lesson, which I call irrational perseverance: the folly we displayed that day in failing to abandon the project. Facing a choice, we gave up rationality rather than give up the enterprise.

Drawn to the Inside View

On that long-ago Friday, our curriculum expert made two judgments about the same problem and arrived at very different answers. The
inside view
is the one that all of us, including Seymour, spontaneously adopted to assess the future of our project. We focused on our specific circumstances and searched for evidence in our own experiences. We had a sketchy plan: we knew how many chapters we were going to write, and we had an idea of how long it had taken us to write the two that we had already done. The more cautious among us probably added a few months to their estimate as a margin of error.

Extrapolating was a mistake. We were forecasting based on the information in front of us—WYSIATI—but the chapters we wrote first were probably easier than others, and our commitment to the project was probably then at its peak. But the main problem was that we failed to allow for what Donald Rumsfeld famously called the “unknown unknowns.” There was no way for us to foresee, that day, the succession of events that would cause the project to drag out for so long. The divorces, the illnesses, the crises of coordination with bureaucracies that delayed the work could not be anticipated. Such events not only cause the writing of chapters to slow down, they also produce long periods during which little or no progress is made at all. The same must have been true, of course, for the other teams that Seymour knew about. The members of those teams were also unable to imagine the events that would cause them to spend seven years to finish, or ultimately fail to finish, a project that they evidently had thought was very feasible. Like us, they did not know the odds they were facing. There are many ways for any plan to fail, and although most of them are too improbable to be anticipated, the likelihood that
something
will go wrong in a big project is high.

The second question I asked Seymour directed his attention away from us and toward a class of similar cases. Seymour estimated the base rate of success in that reference class: 40% failure and seven to ten years for completion. His informal survey was surely not up to scientific standards of evidence, but it provided a reasonable basis for a baseline prediction: the prediction you make about a case if you know nothing except the category to which it belongs. As we saw earlier, the
baseline prediction
should be the anchor for further adjustments. If you are asked to guess the height of a woman about whom you know only that she lives in New York City, your baseline prediction is your best guess of the average height of women in the city. If you are now given case-specific information, for example that the woman’s son is the starting center of his high school basketball team, you will adjust your estimate away from the mean in the appropriate direction. Seymour’s comparison of our team to others suggested that the forecast of our outcome was slightly worse than the baseline prediction, which was already grim.

The spectacular accuracy of the outside-view forecast in our problem was surely a fluke and should not count as evidence for the validity of the
outside view
. The argument for the outside view should be made on general grounds: if the reference class is properly chosen, the outside view will give an indication of where the ballpark is, and it may suggest, as it did in our case, that the inside-view forecasts are not even close to it.

For a psychologist, the discrepancy between Seymour’s two judgments is striking. He had in his head all the knowledge required to estimate the statistics of an appropriate reference class, but he reached his initial estimate without ever using that knowledge. Seymour’s forecast from his insidethaa view was not an adjustment from the baseline prediction, which had not come to his mind. It was based on the particular circumstances of our efforts. Like the participants in the Tom W experiment, Seymour knew the relevant base rate but did not think of applying it.

Unlike Seymour, the rest of us did not have access to the outside view and could not have produced a reasonable baseline prediction. It is noteworthy, however, that we did not feel we needed information about other teams to make our guesses. My request for the outside view surprised all of us, including me! This is a common pattern: people who have information about an individual case rarely feel the need to know the statistics of the class to which the case belongs.

When we were eventually exposed to the outside view, we collectively ignored it. We can recognize what happened to us; it is similar to the experiment that suggested the futility of teaching psychology. When they made predictions about individual cases about which they had a little information (a brief and bland interview), Nisbett and Borgida’s students completely neglected the global results they had just learned. “Pallid” statistical information is routinely discarded when it is incompatible with one’s personal impressions of a case. In the competition with the inside view, the outside view doesn’t stand a chance.

The preference for the inside view sometimes carries moral overtones. I once asked my cousin, a distinguished lawyer, a question about a reference class: “What is the probability of the defendant winning in cases like this one?” His sharp answer that “every case is unique” was accompanied by a look that made it clear he found my question inappropriate and superficial. A proud emphasis on the uniqueness of cases is also common in medicine, in spite of recent advances in evidence-based medicine that point the other way. Medical statistics and baseline predictions come up with increasing frequency in conversations between patients and physicians. However, the remaining ambivalence about the outside view in the medical profession is expressed in concerns about the impersonality of procedures that are guided by statistics and checklists.

The Planning Fallacy

In light of both the outside-view forecast and the eventual outcome, the original estimates we made that Friday afternoon appear almost delusional. This should not come as a surprise: overly optimistic forecasts of the outcome of projects are found everywhere. Amos and I coined the term
planning fallacy
to describe plans and forecasts that

• are unrealistically close to best-case scenarios
  
• could be improved by consulting the statistics of similar cases
  

Examples of the planning fallacy abound in the experiences of individuals, governments, and businesses. The list of horror stories is endless.

• In July 1997, the proposed new Scottish Parliament building in Edinburgh was estimated to cost up to £40 million. By June 1999, the budget for the building was £109 million. In April 2000, legislators imposed a £195 million “cap on costs.” By November 2001, they demanded an estimate of “final cost,” which was set at £241 million. That estimated final cost rose twice in 2002, ending the year at £294.6 million. It rose three times more in 2003, reaching £375.8 million by June. The building was finally comanspleted in 2004 at an ultimate cost of roughly £431 million.
  
• A 2005 study examined rail projects undertaken worldwide between 1969 and 1998. In more than 90% of the cases, the number of passengers projected to use the system was overestimated. Even though these passenger shortfalls were widely publicized, forecasts did not improve over those thirty years; on average, planners overestimated how many people would use the new rail projects by 106%, and the average cost overrun was 45%. As more evidence accumulated, the experts did not become more reliant on it.
  
• In 2002, a survey of American homeowners who had remodeled their kitchens found that, on average, they had expected the job to cost $18,658; in fact, they ended up paying an average of $38,769.
  

The optimism of planners and decision makers is not the only cause of overruns. Contractors of kitchen renovations and of weapon systems readily admit (though not to their clients) that they routinely make most of their profit on additions to the original plan. The failures of forecasting in these cases reflect the customers’ inability to imagine how much their wishes will escalate over time. They end up paying much more than they would if they had made a realistic plan and stuck to it.