13

Torture’s Confession

The real bottom line for any model is whether we find it useful and illuminating.

—MICHAEL LAVER, Private Desires, Political Action

Interrogational torture does not work. President Bush thought it worked. So did his Vice President, Dick Cheney, his Defense Secretary, Donald Rumsfeld, and his CIA Director, Robert Gates. Supreme Court Justice Anthony Scalia thinks it works. The creators of Jack Bauer and the television show 24 as well as the more recent film Zero Dark Thirty think it works. Many Americans think it works. Even many liberal college professors and others opposed to torture on other grounds think it works. For anyone who has ever had so much as a root canal, it sure seems like it should work.

But it does not. Interrogational torture generates bad information. It results in false information by innocent detainees. It results in ambiguous information of unclear value. It results in no information at all. Does saying it does not work mean it can never work? No. It can work. Under conditions that hardly ever obtain in the real world, it can work (but only if we’re willing to torture innocent detainees). Can you put out a fire with gasoline? Yes you can, but, as my brother-in-law physicist says, it is not recommended.¹

There may not be much useful information, but there will be a lot of torture. Torture of innocent detainees. Torture of detainees with information to try and get more information, whether they have more or not. And the torture will be nasty torture too. The slippery slope created by the incentives given to the interrogator means that torture will exceed even the limits and constraints placed on it by those who support its limited use. The ugliness of torture just keeps getting uglier.

We did not reach these conclusions by assuming our interrogators are a bunch of sadists. We did not reach these conclusions by assuming that interrogators are susceptible to dehumanizing social–psychological effects à la what happened at Abu Ghraib. We did not even reach these conclusions by assuming that an organizational culture pushes interrogators to adopt a “willing to do what it takes” mentality or professional incentives geared to favor the quantity of information over its quality.

We reached these conclusions by assuming that our interrogators were basically “good guys” and “good women.” They preferred not to torture, especially an innocent detainee or a detainee who has told all he knew, and were only willing to do so as a last resort. So we did not “rig the game” in advance to get a lot of torture. If anything, we rigged it to get less torture. We gave not only the Bush program but torture proponents generally the best shot they could take to justify their position. And so, from relatively weak assumptions (i.e., reasonable ones favorable to proponents), we arrive at a powerful and important conclusion about what happens when humans torture other humans for information: Good guys get bad information with ugly methods.

IMPLICATIONS

In Chapter 3 we said that if the Bush model closely approximated the ideal model espoused by pragmatic legal theorists and philosophers, then a test of that model is simultaneously a test of the ideal pragmatic model. By the end of that chapter we said that the Bush model did indeed approximate that pragmatic model. Over the course of the next nine chapters we have seen that the Bush model fails. Since the Bush model qualifies as a crucial test of the ideal principle, the pragmatically normative model justifying interrogational torture, the failure of the Bush model is a failure of that legal–theoretical and philosophical justification. Torture cannot be justified.

We cannot torture. No matter what. In real life, it is never just one person and a bomb. Interrogational torture always becomes a government program, with all that entails. Torture really does sit on a slippery slope. Innocents will be tortured, as will detainees who have information but who have provided it all. Resistant detainees will be subjected to increasingly brutal tortures, to no avail. Some information may emerge, but it will be muddled and clouded by false and misleading information. In many cases, there will be no information at all.

Some will say we must torture even if there is a small chance it will work. No. There are some things we cannot do because they run too much against the grain of our character. As Senator John McCain said on the floor of the Senate, the CIA torture program “stained our national honor.”²

We know right now we risk the death of more of our men and women in uniform by refusing to use biological and chemical weapons and flamethrowers. We permit hundreds of thousands of people a year to die—knowing we could prevent those deaths—because we hold other things dear. According to Harvard University’s School of Public Health, 467,000 Americans die from smoking every year and another 64,000 from alcohol use (Danaei et al. 2009). We do not ban cigarettes or alcohol. According to the U.S. Centers for Disease Control and Prevention, about 30,000 Americans die every year from gunshot wounds.³ We haven’t tried to change the Second Amendment to the Constitution. The United Nations World Food Program estimates that approximately 3.1 million children under five die worldwide from undernourishment.⁴ We do not radically reorient our foreign policy aid programs. Do we risk other deaths by forswearing torture? Possibly, but probably not. But we also gain by not swelling the ranks of suicide bombers and staying true to our character (Pape 2015).

THE LOGIC OF INTERROGATIONAL TORTURE

We have found that the pragmatic model claiming to justify interrogational torture fails. We can do more, however. We can say just why and how it fails. Doing so in this section paves the way for saying something more broadly about the role of formal modeling in political philosophy in the next section.

The logic motivating interrogational torture is a simple binary relation between torture and information. No torture, no information. Some torture, some information. Lots of torture, lots of information. Yet proponents of interrogational torture argue that it is possible to employ limited, targeted torture to get lots of valuable information. They want to have their cake and eat it too. We have seen that they cannot. From Chapters 6 to 12 we have seen precisely the opposite: torture exceeding its limits, restraints, and controls but resulting in little valuable information. Instead of minimal torture and maximal valuable information, we see the opposite: maximal torture, minimal valuable information.

Why?

Why Interrogational Torture Fails

The answer is found in the modifications we made from the BIT to the RIT models. The main difference between the RIT model of Chapter 5 and the BIT model of Chapter 4 is the addition of several uncertainties facing the Detainee and the Interrogator, namely those surrounding Detainee and Interrogator types, the clarity of the information, and the degree of information hiding. We found it necessary to include these elements missing from the BIT model in order to better represent the reality of interrogational torture. Indeed, we found that leaving them out resulted in quixotic outcomes far removed from reality.

From Chapter 8 forward we also found that including them resulted in outcomes which do, in fact, reflect reality. (I return to the problem of the no information, no torture equilibrium below.) They reflect that reality because the real world of interrogational torture is much more complex and messy than the BIT model admits.

By including the possibility of resistant and innocent detainees in addition to the cooperative type, we see how the combination of leading questioning and torture leads to ambiguous and false information. We see how innocent detainees with no information look identical to resistant detainees who never give up their information. By including the possibility of a sadistic interrogator we come to understand why even cooperative and innocent detainees who might otherwise provide information do not, fearing they will be tortured anyway. By including the possibility that the interrogator misperceives valuable information as nonvaluable, we show how detainees can be subjected to surprise torture even after they have provided valuable information. By modeling explicitly the common-sense idea that interrogators have some threshold of information disclosure that will satisfy them—a threshold unknown and unknowable by the detainee—we explain refusal by a nominally cooperative detainee to reveal information, surprise torture after providing information, and information satisfying the interrogator and no torture. Finally, by building questioning type (objective versus leading) into the moves and payoffs of the players, we demonstrate a range of problems with interrogational torture, from false and ambiguous information to the trade-offs between questioning type, information, and torture.

Inevitable Trade-off Between Information and Torture

This last example of a trade-off points to something deeper. The reason torture proponents cannot have their cake and eat it too is due to the “double maximand” problem, perhaps made most famous by the utilitarian philosopher Jeremy Bentham. Bentham thought the goal of legislation and public policy was the “greatest happiness of the greatest number” of people (Bentham 1891, p. 93). The trouble is that one cannot maximize both goals simultaneously. There are inevitable trade-offs between them. Kymlicka (2002), p. 50 provides a nice illustration of this problem. Say there are two distributions, each assigning different units of happiness via some public policy to a population of three people. Distribution A assigns 10 to each, while distribution B assigns 20 to persons 1 and 2 but nothing to person 3. Distribution A provides happiness to the greatest number (all three get something), while distribution B provides the greatest (total) happiness (of 40 as opposed to 30).

From Chapter 7 forward, we have seen analogous versions of this problem in the multiple trade-offs between information value and frequency on the one hand and torture frequency and brutality on the other hand. In our case, the problem is simultaneously minimizing torture (maximizing the humanity of an inhumane system) and maximizing information.

From Implication 7.1 we know that there is a trade-off between information and torture across questioning types. All else being equal, leading questioning results in less frequent torture but poorer information; objective questioning results in more frequent torture and better-quality (new) information. It is simply not possible to limit torture’s frequency and still get valuable information as opposed to mimicked confirmation.

Implication 10.2 suggests a related, but distinct, trade-off. Everything else being equal, eliciting information is more likely only if the information becomes less valuable or the torture becomes more severe, not both. There is, in other words, a fundamental, over-arching trade-off between the predictability or likelihood of getting valuable information and minimizing torture. You can increase the reliability of information or you can minimize the brutality of torture. One or the other, but not both simultaneously.

There is a related trade-off between the likelihood of information hiding and torture. Consistent with the logic and incentives of interrogational torture, interrogators will torture if they believe that the detainee is hiding more information. Implication 10.3 tells us that, everything else being equal, eliciting information is more likely (the equilibrium region expands to a greater proportion of the parameter space), and torture of a cooperative detainee less frequent (the torture of an innocent stays the same), only if the standard of detainee cooperation is lowered, thereby increasing information hiding.

We saw all this visually in the discussion of the volume occupied by torture in Figure 12.1. Imagine you have a slider for each of the thresholds , , and . Your task is to try and maximize the volume supporting valuable information, taking one axis at a time.

Start with and imagine it’s in the middle, halfway between 0 and 1. In order to increase the volume occupied by valuable information along the q axis—that is, to increase its frequency by pushing down—torture must become more brutal (higher k) or the information less valuable (lower v). On the other hand, reducing torture severity means reducing the likelihood of information. For torture’s information reliability to go up, so must torture. (Of course we have seen above that even at its maximum possible degree of reliability, torture remains unreliable.) Proponents cannot escape the brutal pain–information logic of interrogational torture.

Now move to , keeping in mind that interrogators torture if they believe detainees are hiding information (i.e., to the left of ). If you slide to the left, you increase the space in which cooperative detainees provide information and are not tortured, but only because you’ve lowered the standard of what you consider full disclosure. This increases the chances of information hiding. In other words, you increase the chances of getting some information and not torturing, but also the chances of missing information.

Finally, grab the threshold. In order to increase the reliability of torture (i.e., expand the valuable information region), you must push it toward the back, as close as possible to its maximum of less than one-half. As you do so, you simultaneously increase the region supporting the torture of an innocent detainee. Reducing the chances of torturing an innocent requires reducing the reliability of information (in the sense the predictability of getting information).

Table 13.1 INFORMATION–TORTURE TRADE-OFFS

In short, interrogational torture is inevitably accompanied by trade-offs between information on the one hand and torture on the other hand (Table 13.1). It is simply not possible to limit torture and still get valuable information. This is why we see the same results every time, everywhere it has been practiced throughout history.

Paradoxes and Perversities

There is more. Whereas the trade-offs just discussed are consistent with the brutal pain–information calculus of interrogational torture—and inconsistent with the claims of limited torture, lots of information—there are also relations which run counter to our intuition, which are even paradoxical.

The more a detainee believes in the logic of torture, the more he believes that interrogators are pragmatic and do not torture gratuitously, the more likely is:

1. surprise torture of a cooperating detainee (Implication 8.1)

2. ambiguous information (Implication 9.1)

3. false information (Implication 9.4)

in addition to the expected valuable information.

Finally, from Implication 10.5, everything else being equal, the very assumptions driving an interrogational torture program—the most likely detainee has information and will reveal it under (the threat of [more]) torture, very few detainees cannot be “broken,” and the chances of a detainee being innocent are virtually zero—all make valuable information less, not more, likely.

Despite the easy but mistaken tendency to think of interrogational torture in one-off, ticking time bomb situations, it is a government program. It is a system of torture. The unavoidable and inevitable messiness and uncertainty accompanying such a system or government program with many interrogators and detainees results in the failure of the pragmatic model. A rigorous and systematic analysis of those results reveals the inevitable trade-offs between information and torture causing that failure. It discloses perversities and paradoxes as twisted and contorted as the bodies subject to its practice.

FORMAL MODELS AND POLITICAL PHILOSOPHY

In demonstrating not only that interrogational torture cannot live up to the claims made for it by its proponents, but also why it fails to do so by interrogating its inner logic, we cash out the claim I made in Chapter 1 that applying game theory to interrogational torture is not just useful, but powerful. It is helpful at this point to step back and think about the formal argument here in the larger context of formal models and political theory, in two somewhat different ways. The first draws on an analogy to the philosopher John Stuart Mill’s four grounds supporting the “liberty of thought and expression” (Mill 1972 [1859], pp. 78-113). The second reflects on the different uses to which models are put. Both sets of considerations reinforce the value of modeling interrogational torture.

J.S. Mill and Challenging Dogma

Mill, a utilitarian philosopher like Bentham, was a staunch defender of free speech, especially speech and expression that ran contrary to popular opinion. He was motivated in part by an Aristotelian impulse that viewed the exercise of reason in public discourse as virtuous activity, as an end in itself. But he was also motivated by the conviction that effective public discourse must be rational—that is, founded on reason and truth rather than dogma and prejudice.

The pragmatic model of interrogational torture is rife with such received opinion, dogma, and prejudice, much of it based on intuition. Mill identified four ways in which free expression confronting popular opinion is a means to the goal of rational discourse and understanding. The same grounds Mill found for confronting public opinion apply to our confrontation of the proponents’ view with the analytical model.

First, we do not want to assume that what we think we know, either from observation or via intuition, is all there is to know. That would assume a sort of infallibility of observation and intuition in terms of completeness or comprehensiveness. The fact is, our intuitions and even our casual observations might be missing something.

An example of this is the surprise torture equilibrium. Though the history of, and commentary on, torture is replete with many different cases, there is little attention given as a separate category to those cases where detainees have provided all their information but continue to be tortured, whereas this is not true for the torture of innocents or other categories. Similarly, the effects of the clarity of the information to the interrogator has received no systematic treatment. In short, game theory helps us avoid what at the outset we called the dangerous seduction of intuition. Intuition tells us we already know instinctively everything there is to know about interrogational torture. Intuition is wrong.

Second, our intuitions may be only partially correct; our observations may be incomplete or tell us only part of the story. To assume otherwise is to assume a different sort of infallibility, an infallibility of accuracy about what we do know, rather than the comprehensiveness or extent of what we know. Intuition, for example, would tell us that everyone breaks. Just contemplating what it might be like to be waterboarded or put in a confinement box for hours on end tends to reaffirm that belief via intuition. We also have observations in the form of historical accounts and case studies that confirm this belief. And yet, Henri Alleg, Jeremiah Denton, and many, many others show us that real life is not so straightforward. Some people really do resist the most horrific pain and psychological disruption and never talk or instead provide intentionally misleading information. Game theory helps illuminate this outcome as well.

Moreover, some of those partially true intuitions can lead to surprising, unforeseen truths. Take, for example, leading questioning. Nearly everyone who has written about torture emphasizes false confessions and false information, especially under leading questioning. The RIT model shows not only how false confession and confirmation emerge from leading questioning, but also ambiguous information, which has not been systematically discussed.

Mill’s third ground is perhaps most directly related. He notes that even if some commonly held opinion is fully true,

unless it is suffered to be, and actually is, vigorously and earnestly contested, it will, by most of those who receive it, be held in the manner of a prejudice, with little comprehension or feeling of its rational grounds (Mill 1972 [1859], p. 112).

There is already far too much prejudice on all sides of the torture debate, with precious “little comprehension.” Game theory supplies the rational grounds or firmament for the intuitions which do withstand scrutiny and a systematic, logical explanation for what we observe.

We see, for example, how a cooperative detainee’s willingness to divulge information can be represented as the ratio of the value of the information to the costs of torture. This simple mathematical expression is derived mathematically, logically, deductively from the model but also corresponds to our intuitions about how both the value of information and pain of torture are related to the detainee’s decision whether or not to talk. Those intuitions are informed by real cases and also guide the logic of the proponents’ pragmatic model. The mathematical expression also lays bare the interrogator’s incentive to keep torturing to drive down the threshold and make the detainee willing to talk.

Finally with respect to Mill, there is the danger that received opinion left unchallenged may become “dogma,” “a mere formal profession, inefficacious for good, but cumbering the ground” (Mill 1972 [1859], p. 112). We see this in the case of torture, with opponents claiming that torture never works and proponents saying it always does, each ignoring the anecdotes advanced by the other. The formal RIT model provides clarification and explication of the conditions under which such claims are (not) true.

Using—Not Believing—Models

A second approach to considering the value of modeling interrogational torture examines the way we use them. As Laver (1997), p. 7 puts it in the epigraph to this chapter, “[t]he real bottom line for any model is whether we find it useful and illuminating.” In other words, models are used for various purposes and should be assessed by how well they achieve those purposes (Clarke and Primo 2007, p. 742). The closer the correspondence to the object they model, the better they will do. Clarke and Primo (2007), pp. 742–744) identify five uses of models, four of which are relevant here: structural, generative, explicative, and predictive. We will also identify an additional use not included in their typology.⁵

The first use is structural, according to which the purpose is to gather known but distinct empirical regularities within one analytical framework. The idea is that there are disparate, disconnected findings or facts which can be generated from and explained in terms of (i.e., structured by) one coherent and unifying logical account. The RIT model is structural insofar as it explains how we get the variety of outcomes we see in the real world of interrogational torture from a single set of assumptions about detainees and interrogators and what each knows and believes.

Several, but not all, of the outcomes in Chapters 8 to 11 have been discussed widely by historians, philosophers, and legal scholars, each emphasizing a different reason or cause. Some have mentioned many of the outcomes, also referring to different causes. None has explained all of them via one consistent, logical framework. Aristotle and Ulpian discuss both resistant prisoners who never reveal information and innocents who falsely confess, as does Beccaria (Aristotle 350, pp. I, 15; Pennington 2008; Beccaria 1764). The Catholic saint Augustine, in contrast, ignores the former case and mentions the danger of the latter, but says the successful cases of confession justify the practice (Augustine 1984, p. 860). The RIT model organizes and explains these diverse outcomes in terms of different combinations of Detainee types and different values for important parameters in the model: questioning type, clarity of information, and others.

The purpose of a generative model is to “generate interesting and nonobvious statements,” “counterintuitive results which are unanticipated prior to the model being solved” (Clarke and Primo 2007, p. 744; also Laver 1997, p. 7). In other words, whereas structural models help us understand better the wide and disparate array of things we already know, generative models give us new perspective on, and insight into, existing problems or phenomena. The RIT model serves this purpose as well.

In fact, one needs only review the short passages on torture from Aristotle and Augustine to Beccaria to see that those reflecting on torture tend to focus on just three outcomes: Innocents who provide false confirmation of questions or false confessions, those who say nothing and hold out, and those who provide valuable information or a confession. There is little discussion of surprise torture or ambiguous information, both generated by RIT. Other examples of new insights include Implication 8.1—the more willing a Cooperative Detainee is to provide information, the more likely he is to be tortured after having provided it—and Implication 8.3—the more important the information (in terms of unknown unknowns), the more likely is “unnecessary” torture (from the proponents’ point of view) because the value of information is less well understood.

Explicative models, as the name suggests, explicate or “explore the putative causal mechanisms underlying phenomena of interest” (Clarke and Primo 2007, p. 744). Substituting in brackets the relevant factors for interrogational torture, a causal mechanism is “a process in which a causal variable of interest … [torture] … influences an outcome [revealing information]. The identification of a causal mechanism requires the specification of an intermediate variable or a mediator [e.g., the detainee’s belief that the interrogator will not torture after providing information] that lies on the causal pathway between the [causal] and outcome variables” (Imai et al. 2011, p. 765).

The RIT model pursues this purpose as well, demonstrating, for example, how the ratio of information value to the cost of torture determines a Cooperative Detainee’s decision to provide information or not. Another example is how a difference between the Detainee’s and the Interrogator’s information hiding thresholds leads to surprise torture.

The last model in the Clarke and Primo (2007), p. 743 typology is predictive, models used to “[f]orecast events or outcomes.” The power of a predictive model is thus based on the range and accuracy of its predictions. The more powerful the model, the closer the correspondence between its predictions and what we see in the world, that is, empirical data. It won’t be missing anything and won’t get anything wrong. The traditional method for assessing this correspondence is data collection and analysis.

The RIT model is predictive in a slightly different way. As we discovered in Chapter 1, the data just do not exist to test systematically the RIT (or any other) model’s predictions in this traditional way. Even so, there should be a minimal sort of correspondence between the model and world. The predictions of the model should not conflict with what we do know about interrogational torture from the dribs and drabs of data over the millennia. Whatever predictions made by the model we should find in the world, and there should be no important outcomes in the world missing from the model.

To this degree, Howes (2012), pp. 20, 23 and other critics miss the point when they say that rational choice models tell us nothing new and “simply formalize what others have already said.” Presumably it would count against a model if it did not generate the outcomes we commonly see. A useful model accounts for all the cases we observe in the real world, does not miss any cases in the real world, and does not generate implausible outcomes we fail to observe. And, indeed, this is the case for the RIT model.

With one exception. You may recall that we found it difficult to identify a real-world case corresponding to the no information, no torture equilibrium. What generates this outcome? The credit we extended to the proponent model that interrogators sometimes believe detainees are innocent after the move “no information” (). We made that assumption in order to both better reflect reality (innocents are swept up in torture programs) and give the benefit of the doubt to proponents that our interrogators do not torture willy-nilly and especially do not want to torture innocents.

It might be, then, that our assumption about interrogators is wrong. In the real world perhaps they really are willing to torture innocents. Maybe. There is another, more charitable and, in my view, more likely explanation. It is that the entire system of interrogational torture—the incentives facing interrogators, their own understandable motivation to root out the bad guys, the organizational culture, the pressure from higher-ups, psychological biases—together constitute a perfect storm making it extremely unlikely a detainee swept up in the system could claim innocence and be believed. Thus the model helps us gain new insight into why we apparently do not see this outcome very frequently in reality. As Clarke and Primo (2012), p. 95 put it, “it is sometimes necessary to write down models that produce outcomes that are at odds with empirical experience in order to better understand that experience.”

But RIT is also predictive in a different way. It generates predictions (equilibrium outcomes) to be assessed not just against what we see in the real world, but also against the predictions of the normative, pragmatic model of interrogational torture advanced by proponents. This is our final, fifth use of models, one we might add to Clarke and Primo (2007)’s list. This use we might call empirically substitutive. The pragmatic model of interrogational torture depends on twin empirical claims (it is effective in generating valuable information while minimizing torture) which cannot be verified empirically. The predictions of the RIT model substitute for those empirical tests, permitting us to assess the pragmatic model in the absence of data.

To this degree, RIT, like other formal models, is a helpful “tool for probing reality” (Brady 2004, p. 297). Our probing of the reality of interrogational torture has provided little support for the claims and justifications of proponents. Interrogational torture necessarily results in increasingly frequent and brutal torture, including of innocents, but fails to reliably yield valuable information. Torture games have no winners.