INTERVIEW: New study finds national security officials are way (way) too confident

The dust may have cleared from the June 22 US bombing of Iran’s nuclear facilities, but the impact of the attack remains clouded by uncertainty: What was damaged? Were uranium stockpiles and centrifuges destroyed or moved? Was any radiation released? Will Iran’s nuclear program survive? What happens next?

Last week, Pentagon spokesman Sean Parnell told reporters that “we have degraded their program by one to two years. At least intel assessments inside the [Defense Department] assess that.” The same day, Iran’s President Masoud Pezeshkian ordered a suspension of the country’s cooperation with the International Atomic Energy Agency. It may be a long while before intelligence agencies, let alone the public, have any clarity about the state of Iran’s nuclear ambitions, and whether the US decision to attack was worth the risk.

Meanwhile, a large new study of US and NATO military and intelligence officers’ intuitions about risk and uncertainty makes one thing clear: Overconfidence among national security officials is nearly universal.

In a forthcoming paper to be published in the Texas National Security Review, Jeffrey Friedman details the results of a survey showing that 2,000 relatively high-ranking national security officials were consistently too sure of their assessments. “Overconfidence was so extreme that it essentially canceled out the knowledge that these individuals possessed,” he writes.

Friedman is an associate professor of political science at Dartmouth, where he researches the politics and psychology of foreign policy decision-making. The results of his latest study are, as he tells me in the following interview, alarming. But he also says there’s an easy fix.

This interview has been edited and condensed.

Thomas Gaulkin: Let’s talk about your new paper: What are the risks and uncertainties involved in assessing the assessment of risk and uncertainty?

Jeffrey Friedman: Basically, there’s wide disagreement about how well we can expect national security officials to assess uncertainty. You have a bunch of realist scholars or rational-choice modelers who say, given how important this enterprise is, we should expect national security officials to be pretty accurate at assessing uncertainty. There are psychologists who say, actually, humans are fallible, and even smart people like national security officials should be prone to overconfidence in their judgments, which is widespread in the general population. And then there’s another group that studies military culture and bureaucracy who say, actually, it might be the opposite, that many national security analysts are inclined to hedge their judgments, perhaps overly so in in ways that could make them underconfident.

And part of the reason we just don’t know the answer to this is that it’s notoriously hard to get systematic evaluations of how accurate national security officials judgments are. So, to my knowledge, this is the first study that gathers a large generalizable sample of national security officials.

Gaulkin: Who were the officials, and what did you ask them to assess?

Friedman: Roughly 2,000 national security officials took part in the study, from more than 40 NATO allies and partners. The officials all came from professional military education institutions like the US National War College or the NATO Defense College. These are relatively high-ranking people. If you’re an active-duty military officer, you generally go to these institutions once you’ve been promoted to the rank of colonel, and then these institutions all have civilian participants who come in from foreign affairs ministries or intelligence services, and they’re at similar ranks.

As part of a relationship with four of those institutions, I conducted surveys in which these national security officials would make a few dozen assessments of uncertainty. Some of those would be in response to factual questions. So we’d ask, “What are the chances that the United States defense budget is larger than the defense budget of the next 10 countries combined?”, or “What are the chances that ISIS has killed more people than Boko Haram over the last decade?” Other questions were forecasts: “What are the chances Bashar al-Assad will still be in power a year from now?” “What are the chances that next year will be the hottest year on record?”

So we have some assessments of uncertainty about current states of the world, some predictions. They made more than 60,000 assessments. And the very clear conclusion from these is that, at least within the context of what we can measure from surveys, national security officials judgments are vastly and systematically overconfident.

Just to give you some summaries of the data: When these national security officials said they thought a statement was 90 percent likely to be true, it was only true about 57 percent of the time. When they thought that a statement had only a 10 percent chance of being true, it was true 32 percent of the time. And that is not just a matter of struggling to use numbers. For example, when I ran a version of the survey in which people used qualitative statements to assess uncertainty, when they said they were almost certain the statement was true, the statement was false 32 percent of the time.

Gaulkin: So what does this tell us about what’s going wrong when these assessments are made, and what’s the upshot for national security?

Friedman: I would say that this is a really good method for getting lots and lots of data on how national security officials assess uncertainty. The downside with it is these are surveys to which people are responding fairly quickly. But what I think we can say is: This measures national security officials’ intuitions, the cognitive first steps that they take when they address a decision problem. We have really good evidence that those intuitions then anchor decision processes and may even get reinforced by decision processes. Of course, we can’t say this is exactly the same thing as, say, evaluating intelligence estimates, but at the very least it gives it gives us a very clear window, I think, into national security officials’ psychology when it comes to assessing uncertainty.

And it’s just very clear that whether they’re using numbers or words, people are assigning much too much certainty to their judgments. And this is not unique to national security officials. When conducting similar exercises with students at Dartmouth or participants in executive education programs, almost everyone who has not received structured feedback on their assessments of uncertainty is overconfident. So this is totally common, and I think the result of this for national security decision making in particular is that people are liable to overestimate the effectiveness of military actions. If you think that a military action is likely to succeed, you will probably be overestimating that likelihood. And it can also lead people to underestimate risks—if you think some downside risk is unlikely to materialize, you’re probably too certain about that also, and not assigning enough credibility to that picture.

And just to emphasize how consistent these patterns are: 97 percent of national security officials in the study would have made more accurate judgments if they assigned less certainty to every single one of the judgments they made. So this is not something that’s driven by outliers. It appears to be a very generalizable bias within this community, which is consistent with what we see in almost any community of people who haven’t received this kind of feedback before.

I should say, I’m certainly not the first person to think about this. There’s been a lot of good qualitative work on case studies here. The war in Iraq is an obvious example where decision makers were overconfident about effectiveness and underplayed downturn risk. There’s a terrific book on this by a political psychologist named Dominic Johnson. I should note that he thinks that some degree of overconfidence is productive. But again, what this article does is helps us to quantify just how big and how consistent these patterns are. I don’t think he’s talking about, you know, this kind of massive overconfidence. He’s pretty clear in his book that a healthy dose of overconfidence is good; but when the data show that when people are completely certain, they’re wrong 25% of the time, I think that’s pretty alarming.

Gaulkin: Presumably, part of being a qualified national security official is that you have the knowledge and experience to support your intuitions. What goes into being a national security person who’s good at assessing risk—and why were they so bad at it in this scenario?

Friedman: Everybody who participated in this study is extremely impressive. We’re talking about people almost all with nearly two decades of experience or more. National War College in the United States is arguably the most prestigious place that you can be sent for this level of education. I mean, these are future chiefs of staff, combat and command. These are really impressive people. So they obviously know a lot.

But I think part of the challenge with assessing uncertainty is most people go through life without getting clear feedback on how accurate their judgments are. And part of the problem with understanding those capabilities is that they’re inherently abstract. Like, if I say something has a 70% chance of happening and it doesn’t happen, maybe my judgment was wrong, or maybe I just got unlucky. It’s sort of hard to know. You can really only get clear feedback on how well or how poorly you’re doing if you take the time to gather the data. And for better or for worse, it just turns out that most national security institutions don’t gather those data, and don’t provide personnel with the kind of structured feedback that you would need in order to calibrate your assessments of uncertainty.

They’re certainly not alone in this. I mean, most organizations don’t provide that kind of feedback. Does the Bulletin of Atomic Scientists do that? Probably not. Does Dartmouth’s government department do it? No. Does Goldman Sachs do it? No. I mean, I just think is hard to do, and it’s unusual.

But there’s a lot of good psychological research suggesting that in the absence of that feedback, people think they are better at assessing uncertainty then they really are. They tend to wave away inaccurate judgments, saying they just got unlucky or they were close, or it didn’t really count while taking full credit for their successes. This creates what could be called an illusion of skill, or an illusion of efficacy. And—particularly if you’re a high-flying national security official and you’ve been promoted rapidly through the hierarchy, and you have lots of reasons to believe that you are really good at most aspects of your job—I think it’s pretty easy to come away with the impression that you’re better at assessing uncertainty than you really are. And that leads naturally to overconfidence.

Gaulkin: On the other hand, you say in the paper that just two minutes of training significantly improved officials’ assessments. What kind of training is that effective?

Friedman: The gold standard on this is what decision scientists call calibration training. So the way that I do this in my class at Dartmouth is I give my students a set of assessments to make. I show them how overconfident they are. I tell them that, like just about everybody else, when they think statements are 90 percent likely to be true, they’re actually true more like 60 percent of the time. And then we do the exercise over again and we can see rapid improvement. Just shattering that illusion of efficacy really helps people to correct biases that they genuinely didn’t know they had.

There are a number of techniques for doing this. You can teach entire classes on this. You can give weeks of instruction at national security academies on this. But one of the things that the paper shows is that you can also correct a lot of these biases, at least on the margins, in two minutes. What I did was, at the start of a random subset of surveys, I just provided national security officials with data on how their predecessors had performed on the survey. It showed that almost everyone who had previously taken the survey had been overconfident, told them about the degree of that bias that I’ve described to you, and then let them answer the same questions the way everyone else did. And that two-minute training alone improved performance by a quarter of a standard deviation. That’s quite a bit for just two minutes.

And I think part of what this is showing is that I don’t think national security officials are hard-wired for overconfidence. It’s not that they’re committed to being overconfident. I think like everybody else, they genuinely want to get it right. They just also are genuinely unaware of how biased their intuitions are. So even a very brief indicator that they’re liable to be overconfident can have pretty significant impact on pushing them in a better direction.

And one of the broader implications from this study is that it indicates that national security bureaucracies could harness these gains at scale. You know, if just two minutes of training can improve by a quarter of a standard deviation—and there are many other studies that show there that document that similar improvements can be permanent—I think I could make a good case that this is something that would be worth deploying widely in the national security establishment.

Gaulkin: What’s the mechanism behind that fast improvement—after reading that information about overconfidence and bias, is it just sort of taking a beat and then thinking, “Oh, ok, I might be fallible too,” something like that?

Friedman: The central insight—and this is the title of the paper—is that the world is more uncertain than you think. And it’s a simple point. It’s not complicated. It’s just something most people do not understand. If you’ve gone through two decades of a successful, fast-rising career in national security, you are highly aware of how capable you are, and you can lose sight of how much you don’t know or how challenging it is to assess uncertainty in international politics. And it’s hard to appreciate just how uncertain the world is. It can be quite shocking to see just how much more uncertain the world is than you think. But I don’t think it’s a complicated insight. That’s why once the fact is made clear to reasonable people, they can adapt.

Gaulkin: We’re talking now about a week after the US bombed Iran’s nuclear facilities, and we’ve seen various degrees of confidence expressed about the impact. Starting with President Trump, who said the sites were “obliterated,” and the director of the CIA who said they were “severely damaged,” and the Defense Intelligence Agency said—with “low confidence”—that Iran’s nuclear program was only set back a few months or something.

The prospect of the US bombing Iran has obviously been pondered and sung about for a long time; now that it has been, what seems most certain at the moment is just how uncertain officials are about the outcome. Have you seen anything that suggests how well the risks and uncertainties were assessed before the US attacked?

Friedman: I think one of the risks of overconfidence that’s relevant here is that national security officials and presidents are constantly confronted by ambiguity. There’s almost never a clear right answer, particularly in the short run, as to how you evaluate complicated things like, you know, whether Bin Laden’s in Abbottabad, or whether an attack on Fordow will succeed. So in those circumstances, people have to use their judgment and their intuition. The way you resolve that uncertainty is inherently subjective. That’s what my first book War and Chance was all about. If we believe that people are inclined to resolve that kind of ambiguity and deploy their intuitions in ways that are overconfident, that’s going to mean that people are going to generally be overoptimistic about the chances for things they think are likely to be true, and they’re likely to be overly dismissive of risks that they think are unlikely to materialize.

But I actually haven’t seen great reporting yet about decisions for the operation against Iran itself. I’m not sure if that’s out there yet. It would be reasonable to think that overconfident decision makers would overestimate the prospects for terminally damaging Iran’s nuclear facilities and perhaps underestimate some of the risks of blowback that that could cause. And that doesn’t necessarily mean they make the wrong decision, but I do think the results of this survey just provide some caution that our natural intuitions for resolving uncertainty have some pretty stark biases in them. And particularly in circumstances where it’s very hard to get ground truth or objective answers, that’s exactly where you want to be worried about those cognitive pathways leading you awry.

Gaulkin: Great. So, one last topic: insurance. I don’t know if this comes up in in your scholarship sometimes, but thinking about insurance actuaries, they’re obviously very invested in assessing risk as accurately as possible for their companies’ bottom line. And one thing I saw just a couple months ago was an example from the airline industry, where—due to everything that’s developed around Ukraine primarily, and with the help of security experts—some insurance groups are shifting how they assess risk so planes will continue to be insured to fly even after a nuclear blast.

So, is there a way that insurance also provides a model for improving risk assessment among national security officials?

Friedman: Insurance companies really do this for a living. Particularly when dealing with risks that occur frequently. We’re talking about floods, fires, stuff like that. They have really highly calibrated models for figuring out how to think about risk, and they specialize in that. We should expect them to do better than ordinary people, I think.

We also know that insurance companies, like everybody else, struggle with these so-called Black Swan events. Something like the risk of nuclear attack, something that has never happened outside of World War II, has never happened between two nuclear armed powers. You know, that’s something where you cannot develop a large dataset. So there, it is not obvious that insurers working on that would necessarily be better than anybody else.

But there are a bunch of ways that you can nevertheless try to get the handle on that. Bond markets often price in catastrophic risk. There are prediction markets, including for any number of events surrounding the Iran attacks. I follow those really closely. I cannot personally beat Polymarket, so I tend to trust what they say, for better or for worse. There are prediction polls. There’s an organization called the Good Judgment Project, which has developed a crowdsourcing algorithm in which they ask hundreds or thousands of people to estimate probabilities, and they have developed very sophisticated ways of combining and weighting those judgments in order to create what looks like a potentially valid output. There are many companies, like Metaculus, that do this.

So there are lots and lots of methods for trying to get a handle on these things. But those are a little different from the kinds of intuitive judgments [by national security officials] that that my paper picks up on. I think part of why those structured methods exist is because the authors of those methods are very skeptical of intuitive judgment, and I would say as somebody who’s been playing political prediction markets for 15 years and just keeps losing money year after year, I can understand why those exist.

Gaulkin: I assume AI is also going to change a lot of how this is accomplished in the future.

Friedman: Yes, it’s true. And AI is quite good at these prediction polls. The super cutting-edge work, including stuff that’s being sponsored by the US intelligence community, is how to create AI hybrid forecasts, so-called “centaurs,” in which you have a human who interprets the output of an AI. There’s a really good paper on this by a scholar named John Lindsay that came out in the journal International Security about how he thinks AI will actually increase the importance of the human factor in decision-making.

I honestly don’t think we know. And again, unfortunately for a lot of things that the Bulletin cares about, like atomic risk, you do not have large data sets to draw on— it’s a place where there is no big data [for AI to be trained on]. The Large Language Model can do its synthetic reasoning thing, but the neural networks are primarily useful when you’ve got these bonkers datasets with impossible numbers of interactions, and that’s just less well-suited to the hyper-rare but super-consequential things, for which we must inevitably rely on our intuition. And unfortunately, our intuitions are quite, quite inaccurate.