12

Torture’s Garden

Wherever you go there is more pain, more torture, more blood flowing and draining through the soil … more terror and more hell.

—MIRBEAU, Le Jardin des Supplices, 1899

In Octave Mirbeau ’s disturbing fin de siècle novel, there is a garden with many different tortures. We too have found torture’s garden to be extraordinarily verdant. It contains all varieties of information. False information. Ambiguous information. Partially and fully valuable information. No information at all. Torture also blooms well there. Torture of detainees who hide information. Torture of detainees who hide no information and have given all the information they have. Torture of innocents. Torture of those from whom it cannot compel information.

We are now ready to harvest this fruit and assess the claims made by torture proponents in Chapter 3. That is, we can collect the observations, propositions, and implications over the past six chapters and compare them to the four sets of benchmarks set by torture proponents. Recall that each benchmark has two components, a predictive component that tells us what will emerge from the model as well as a normative component that tells us the standard against which we compare what actually does emerge from the model. Thus, meeting the normative benchmark depends on satisfying the predictive benchmark; a failure to achieve the prediction means that the normative standard has not been met. A failure to meet these normative standards, in turn, would mean that the pragmatic model fails to satisfy even its own criteria for justifying interrogational torture. As a result, for each set of benchmarks—information reliability, torture frequency, torture severity, and the torture justification outcome—we examine first the predictive benchmark and then use those results to assess the normative benchmark. Since we have run through all the equilibria, we start with the last benchmark, the pragmatic model’s Torture Justification Outcome.

TORTURE JUSTIFICATION OUTCOME

Recall the predictive component of Benchmark 4 from Chapter 3 stating the ideal outcome sketched by proponents of the Bush program:

BENCHMARK 4 Torture Justification Outcome (TJO): A minimum degree (severity) and amount (frequency) of torture is used against only the most resistant detainees with valuable information, who give up all, or nearly all, that valuable information. Neither cooperative detainees who have provided all their information nor innocent detainees are tortured.

In order to compare this prediction against the results of RIT, we need to first translate it into game theoretic language and the terms of the RIT model. This gives us the following:

BENCHMARK 4^′ Torture Justification Equilibrium (TJE): Supported by a reasonable set of beliefs, the following strategies will constitute an equilibrium under objective questioning in the RIT model: the Cooperative Detainee plays “information,” the Innocent Detainee plays “no information” (i.e., tells the truth), and the Pragmatic Interrogator plays “no torture” after both actions (“information” and “no information”).

Setting aside for the moment what we mean by “reasonable” beliefs, we can begin our search for an equilibrium matching the TJE by eliminating any equilibria in which no valuable information is provided. This leaves the valuable information–surprise torture equilibrium under objective questioning and the two valuable information–selective torture equilibria, one under objective and one under leading questioning. In principle, since leading questioning provides no new information and no new intelligence, assessing the claims about information made by the pragmatic model should require us to stick to equilibria under objective questioning. For the sake of argument, however, we can set this aside for the moment and what we will say applies to both questioning types.

Recall that there are two interpretations of both the surprise torture and valuable information equilibria, one in which the Cooperative Detainee genuinely has revealed all he knows (full disclosure) and one in which he is still hiding information (partial disclosure). Each interpretation applies to both equilibria; in one he is tortured (valuable information, surprise torture), whereas in the other he is not (valuable information, selective torture). The full-disclosure interpretation of the surprise torture equilibrium we should eliminate for torturing a Cooperative Detainee after having provided all his information because it conflicts with the TJO and the TJE’s prohibition on this torture. It makes no difference in what follows, however, whether we include or exclude it.

Strictly speaking, the second interpretation (partial disclosure) of the surprise torture equilibrium might also be ruled out by the TJE because the Cooperative Detainee is tortured, but only because we do not specify whether the information disclosed is full or partial in stating the TJE. It seems unfair to proponents to eliminate this equilibrium since a detainee with information has refused to provide it all and is tortured. This is a justifiable outcome on the pragmatic view.

We might also eliminate the partial disclosure interpretation of the valuable information equilibrium because it means that a detainee got away with hiding information. We can, though, grant this too to torture proponents, for it changes nothing in what follows.

Thus, both the partial disclosure interpretation of the valuable information–surprise torture equilibrium and both interpretations of the valuable information–selective torture equilibrium come closest to the pragmatic ideal.

Closest, but not close enough. An Innocent Detainee is still tortured for telling the truth in both sets of equilibria.

The reason is that the Interrogator believes the failure to reveal information signals a Resistant—not an Innocent—Detainee (), who is then tortured. The Interrogator must torture after failing to receive valuable information in order to compel valuable information from those willing to give it up under the threat of torture. This holds for the valuable information–surprise torture equilibrium and for the two valuable information–selective torture equilibria, one under objective and one under leading questioning. We thus have the following proposition:

Proposition 12.1 (Necessary Torture of Innocents). In order for interrogational torture ever to generate any valuable information, innocent detainees must be tortured for telling the truth.

Indeed, we know from Implication 10.4 that the likelihoods of information and torture of innocents inevitably track each other. Expanding the information region expands the region supporting torture of innocents; reducing the torture of innocents means reducing the likelihood of valuable information. As a result, the TJE of Benchmark 4^′ does not exist:

Proposition 12.2 (TJE Impossibility). The set of strategies defining the pragmatic model’s torture justification equilibrium (TJE)—the Cooperative Detainee plays “information,” the Innocent Detainee plays “no information” (i.e., tells the truth), and the Pragmatic Interrogator plays “no torture” after both actions (“information” and “no information”)—does not occur in equilibrium for any set of beliefs.

Information from torture is possible (albeit under very unlikely conditions, as we will see), but only if innocents are tortured. It is possible to avoid torturing innocents, but then no valuable information is forthcoming. Not both. Moreover, from Implication 10.1 we also know that it is impossible to minimize both the frequency and brutality of torture in eliciting information, as the TJO benchmark requires: “A minimum degree (severity) and amount (frequency) of torture is used ….”

Thus, the ideal outcome of torture proponents is just that: a quixotic ideal that fails to reflect reality of interrogational torture. Interrogational torture fails to satisfy the predictive condition of Benchmark 4. Now recall the normative part of the TJO benchmark:

BENCHMARK 4 Torture Justification Outcome: Torture in interrogations is justified if and only if torture

1. is not used against cooperating detainees who have provided all their information (Cooperatives),

2. is not used against innocent detainees (Innocents),

3. does not exceed the minimum frequency (Total Frequency) and severity (Severity) “necessary,” and

4. (the threat of) torture generates all, or nearly all, the valuable information possessed by knowledgeable detainees (Information).

All four conditions must be met for the program to meet the proponents’ benchmark. A violation of any one condition renders the program a failure.

Proposition 12.2 shows that the last, Information, condition is satisfied only if the Innocents condition is violated. The latter condition cannot be satisfied if the Information condition is to be satisfied. It is not possible to both elicit information and avoid torturing innocents. The surprise torture equilibrium and the fact that it is likely (Implications 8.1 and 8.2) violate the Cooperatives condition. Proposition 12.5 shows that torture is likely to exceed restraints and guidelines on severity, violating the Severity condition. We return to the questions of torture frequency and severity in more detail below, but what we have seen thus far is sufficient to show that, using its own model and its own logic, the pragmatic model of interrogational torture fails to meet its own normative standard justifying the practice.

INFORMATION RELIABILITY

Perhaps this seems too strong to some. Even if the ideal outcome does not occur in equilibrium, (some) torture proponents may be willing to accept (many? some? a few?) innocents being tortured as long as the program provides good information overall. On this view, the program would be a success if it met the predictive and normative parts of the Information Reliability benchmark, even if it failed to achieve the ideal outcome claimed for it. Recall that the predictive part of this benchmark is:

BENCHMARK 1 Information Reliability: Most detainees have information and give up (nearly) all of it so that the ratio of clear and valuable information to all information will be high.

We can assess this prediction in two ways, by looking at the set of RIT equilibria and outcomes and by using the parameter space to examine the set of beliefs supporting the prediction. Starting with the former, there are nine total equilibria of the two versions of the game (leading and objective questioning), but three of these equilibria call for the same actions by the Detainee and the Interrogator. Thus, there are six behaviorally distinct equilibria. Since there are also two substantive interpretations (full and partial disclosure) of two of the six equilibria (valuable information, surprise torture and valuable information, selective torture), there are eight total outcomes.

Valuable information occurs in three of the nine total equilibria and four of the eight substantive outcomes of the RIT model. If we were to imagine for the moment that all the equilibria are equally likely, then your chances of ending up with valuable information are as low as one in three and no higher than a coin flip. Remember too that this includes the leading questioning version of the valuable information, selective torture equilibrium. Were we to exclude it since it provides no new information or intelligence, then those numbers drop further, to a low of just over two in ten equilibria. If any of these numbers were the chances your car would start on a cold morning, it’s unlikely you would tell your friends you have a reliable car.

But perhaps some equilibria are just unlikely in terms of the beliefs required to sustain them. Perhaps we should just focus on the equilibria in which valuable information is provided. We can once again deploy the parameter space “cube” to assist us. Since introducing the parameter space in Chapter 7, we have thought about the equilibrium regions in the following way. A large region is supported by a greater range of beliefs than a smaller region. This means that, for any equilibrium, an increase in its size means that a greater proportion of possible beliefs sustain it. This, in turn, suggests that it is more likely; it can occur for a wider range of conditions. Benchmark 1 predicts, then, that equilibria with valuable information should take up a large proportion of the parameter space.

As we just discussed, the only two equilibria with valuable information are valuable information, surprise torture and valuable information, selective torture. Once again, since leading questioning provides no new information, strictly speaking the leading questioning version of the valuable information, selective torture equilibrium does not satisfy reliability Benchmark 1. For the purposes of the illustration and the discussion below, we will speak primarily about the objective questioning version, but what we will say applies equally to the leading questioning version.

We have already placed these equilibria individually in the parameter space in Chapters 8 and 10. Figure 12.1 places them together in that space.

Figure 12.1 Maximum Extent, Valuable Information and Torture

Take a look first at the darker-shaded region supporting valuable information, selective torture, recalling that this region is defined by , , and . From Observation 7.1 and Propositions 7.1 and 7.2, we know that both and are less than one-half (i.e., toward the front), with approaching that value and approaching zero the more the Interrogator believes everyone breaks—as is likely.

From Proposition 7.3, we know that as the cost of torture to the Interrogator c approaches zero and the additional cost of unnecessary torture drops (as a approaches r), again, as is likely, the Interrogator’s information hiding threshold is greater than or equal to one-half (i.e., on the right side). The final threshold we cannot constrain because, as we said in Chapter 7, there are neither formal mathematical nor good empirical grounds to do so. As a result, Figure 12.1 bounds the valuable information region using what we know about and and assuming a most favorable to proponents, that is, very low so the region is about as large as it can be. It is worth noting that in order for the threshold to be this low, either the information must be of very low value (low v) or the torture must be very brutal (high k) (or both).

Now take a look at the lighter-shaded region to its left. This is the volume supporting valuable information in the surprise torture equilibrium. The only threshold differences here are the addition of , which forms the left-hand boundary of the equilibrium whereas forms the right-hand side, dividing the two equilibria. The dotted region occupying the volume to the front of is the region in which either the Cooperative or Innocent Detainee (or both) is tortured in at least one of the equilibria in the model. The blank space to the back of is occupied by no torture. We return to these last two regions shortly.

Sticking with information reliability for the moment, what does Figure 12.1 suggest about the likelihood of getting valuable information from torture?

Don’t hold your breath.

The lighter- and darker-shaded regions together are at about their maximums and they take up a relatively small proportion of the parameter space. Moreover, we are including the surprise torture equilibrium, which means that the information comes at the price of torturing both innocents and those who have provided information—maybe all of it. Nor would it help to put the leading question variant of the valuable information, selective torture equilibrium into the space. That equilibrium simply occupies the exact same space currently occupied by the two equilibria in Figure 12.1; it would not increase the total space occupied by valuable information. (And the information from leading questioning is not as valuable anyway.)

Since even at their maximum volume the two equilibria including valuable information take up a relatively small proportion of the parameter space, they occur for only narrow ranges of values on two of the three belief axes. Now recall that, to keep the model tractable, we gave the Detainee only two moves, “information” and “no information.” We gave the move “no information” three interpretations: (1) truthful, accurate, but nonvaluable information, (2) false and misleading information, and (3) no information whatsoever.

Thus, the complementary volume (everything outside the gray-shaded boxes) represents some mix of no information and false and misleading information, as Figure 12.2 more clearly illustrates. This is the minimum extent of no/false information; as rises, the white box representing information shrinks toward the top. The vacated space is colonized by more of the gray “no information.” It is also important to note that this volume does not represent the absence of information or the bad information from a Resistant Detainee, which we expect to get all the time by assumption. Nor is it even only bad information from false confirmation by an Innocent Detainee attempting to avoid torture. This is the lack of information or the elicitation of misleading information from a Cooperative Detainee—that is, a Detainee who does have valuable information and is willing to give it up under the right conditions.

Figure 12.2 Minimum Extent, False and No Information by Cooperative Detainee

In short, torture in interrogations does not predictably, reliably generate valuable information. This, together with the relative rarity of valuable information among the set of all equilibria, gives us the following proposition:

Proposition 12.3 (Information Unreliability). The ratio of valuable information to all other information under interrogational torture is low, with a mixture of no information, nonvaluable information, and false information dominating the parameter space.

Moreover, other implications support and explain why. We know from Implication 7.2, for example, that, all things being equal, interrogators are more likely to get less valuable information than highly valuable information from torture, no different from interrogations without torture. This directly contradicts proponents’ claims that torture is justified precisely because it does do better. Although increasing the brutality of torture (raising k) increases the likelihood of valuable information, we know from Implications 9.2 and 9.5 that it also increases the likelihood of ambiguous and false information elicited by the Interrogator under leading questioning. The Torture–Information Paradox in Implication 10.5 reveals that the very same assumptions motivating states to torture for information are likely to make the elicitation of valuable information less, not more, reliable.

As the Information Reliability benchmark makes clear, the normative standard for information reliability depends on interrogational torture predictably generating reliable information:

Benchmark 1 Information Reliability: Interrogational torture is successful if and only if detainees give up (nearly) all their information so that the ratio of clear and valuable information to all information is high (Information).

Proposition 12.3 as well as Implications 7.2, 9.2, 9.5, and 10.5 demonstrate that the pragmatic model fails to meet its own normative Information Reliability standard. Interrogational torture is not reliable in generating valuable information, despite the claims of proponents.

The only possible justification for interrogational torture is that it is effective in reliably generating reliable information. This is a necessary (if not necessarily sufficient) condition for the practice ever to be justified. The failure to meet this condition refutes the pragmatic justification of torture for information.

TORTURE FREQUENCY

We have just shown that the “end” part of the “end justifies the means” logic of interrogational torture fails to live up to the pragmatic model. What about the “means” part of the calculus? The pragmatic model claims that torture will be very limited, visited upon just a small fraction of detainees. More precisely, according to the predictive part of the Torture Frequency benchmark, we have the following:

Benchmark 2 Torture Frequency: Torture will be employed infrequently, just on a few particularly resistant detainees who refuse to provide information, so that:

1. the total frequency of torture is low (Total Frequency),

2. Cooperative detainees are not tortured after they have provided all their information (Cooperatives),

3. Innocent detainees are not tortured for telling the truth (Innocents).

Proposition 12.1 already showed that the Innocents condition is violated. The valuable information, surprise torture equilibrium and the fact that it is likely (Implication 8.2) together demonstrate that the Cooperatives condition is violated.

How about the Total Frequency condition? It is admittedly a bit vague, but then so are proponents and their claims. Presumably there should be few detainees who are tortured in both absolute and relative terms. In other words, if torture is a last resort for hardened terrorists, then we should expect not just that the total number of detainees tortured should be small, but also that this number should be a small proportion of all detainees questioned. After all, many detainees are not hardened and cooperate. There will even be innocents who should be weeded out of the program before they are ever tortured.

So the Total Frequency condition requires that the proportion of those tortured in this reference population should be very small. In terms of the model, this suggests that torture should take up a very small proportion of the parameter space and should figure in few of the model’s equilibria. So how does this prediction fare?

Not very well.

Take another look at Figure 12.1 and the region filled with dots. This is the volume to the front of the interrogator’s innocent detainee recognition threshold . As we said above when introducing the figure, there is some form of “unnecessary” torture in this region, either of an Innocent detainee or of a Cooperative Detainee who has provided all his information, or both. This is why, for example, it covers the shaded regions where valuable information is provided. Even in the valuable information, selective torture equilibrium, where the Cooperative Detainee providing information is not tortured, the Innocent Detainee is tortured for failing to answer the Interrogator’s questions satisfactorily.

Recall from Proposition 7.1 and the discussion thereafter that . This means that torture takes up just under half of the entire parameter space. This is far greater than the tiny fraction that one would expect from the claims of proponents and the Total Frequency condition of the Torture Frequency benchmark. Also remember that the entire parameter space includes the cases where the Interrogator is in general willing to torture if Detainees don’t talk but lets them go without having provided information because she thinks they are innocent (i.e., ). If we restrict our attention to the region in which Interrogators think a Detainee providing no information is not innocent, as would seem likely, then this “unnecessary” torture takes up the entire space. In other words, the choice of reference population makes no difference. In both cases the prediction in the Total Frequency condition is incorrect; torture occurs with greater frequency than predicted by proponents.

What if we look not at the thresholds and range of beliefs represented by the parameter space but instead at the range of equilibria and outcomes in which there is “unnecessary” or “unjustifiable” torture? As a reminder, “unnecessary” and “unjustifiable” mean here unnecessary and unjustifiable in the eyes of torture proponents, not opponents. The benchmark requires that there should be few, if any, equilibria supporting such “unnecessary” or “unjustifiable” torture. How accurate is this prediction?

Not very.

In seven of the eight substantive outcomes the Resistant Detainee is tortured for failing to provide information, even though it has no effect on him. This does not count against the pragmatic model, however, for on that view such torture is justified. How about what proponents themselves consider “unjustifiable” torture?

Five of the eight substantive outcomes entail such “unjustifiable” torture on the pragmatic view. Knowledgeable Detainees are tortured, for example, after they have provided all the information they possess, contra the Cooperatives condition. Innocent Detainees are tortured for telling the truth, contra the Innocents condition. In two of the remaining three outcomes, the Innocent Detainee falsely confesses under leading questioning and the threat of torture.

Moreover, 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 (Implication 8.3). Even under leading questioning, the more important that confirmation is to the Interrogator, the more likely is torture of an Innocent Detainee (Implication 9.3). There is only one outcome in which neither a Cooperative nor an Innocent Detainee is tortured unjustifiably or falsely confesses (the no information, no torture equilibrium), but in that equilibrium both the Cooperative and Resistant Detainees get away without providing any information. Indeed, this is the only outcome of the eight in which there is no torture at all.

Drawing on our observations about both the proportion of the parameter space taken up by torture as well as “unjustified” torture’s frequency among the set of outcomes, we can state the following proposition:

Proposition 12.4 (Torture’s Slippery Slope 1). Once torture is admitted as an interrogation technique, it will exceed the limits and controls imposed on it and become more frequent than proponents expect.

Recall now the normative part of the Torture Frequency benchmark:

Benchmark 2 Torture Frequency: Interrogational torture is successful if and only if torture is not employed too frequently:

1. the total frequency of torture is low (Total Frequency),

2. Cooperative detainees are not tortured after they have provided all their information (Cooperatives),

3. Innocent detainees are not tortured for telling the truth (Innocents).

Proposition 12.4 demonstrates that the Total Frequency condition is violated as well. This, along with having demonstrated that the Cooperatives and Innocents conditions are also violated, shows that interrogational torture fails to satisfy the pragmatic model’s normative standard for Torture Frequency. Interrogational torture results in more torture than even proponents anticipate and are willing to accept.

TORTURE SEVERITY

The pragmatic model places restrictions not just on who is tortured, but also on how those who are tortured are tortured. In particular, there are restrictions on the range of acceptable techniques as well as on the intensity or severity of those techniques. The result is the predictive component of the Torture Severity benchmark:

Benchmark 3 Torture Severity: When torture is employed, its severity will approximate the minimum degree necessary to compel valuable information.

Now proponents might, again with some justification within their pragmatic model, say “minimum degree” cannot be interpreted too literally or it becomes a strawman easily taken apart. Fair enough, but neither can it be so loose as to be a Jello-man impervious to examination because it is too slippery to nail down. This will not matter if the model comes down clearly one way or the other. Let’s see.

Although the RIT model does not make the severity of torture one of the moves by the Interrogator, it is still possible to assess Benchmark 3 by examining another feature of the model: the Detainee’s information revelation threshold in Figure 12.1. This threshold along the vertical q axis is the point at which the Detainee switches from “no information” (below the threshold) to “information” (above the threshold). Recall that the threshold, mathematically derived from the model but also intuitively compelling, is the ratio of the value of information v to the costs of torture k, or (both to the Detainee). Once again, it takes little familiarity with mathematics to see that this fraction (and so the threshold) increases as the value of information increases and/or the torture costs decrease. As the threshold increases, the space in which valuable information is provided to the Interrogator, the bottom surface of the darkest-shaded valuable information region in Figure 12.1, shrinks toward the top of the cube.

The job of an Interrogator using torture is to lower the threshold, to drive it down so that the Detainee talks “earlier” and the valuable information region moves toward the bottom, increasing its volume in absolute terms and relatively in terms of the total parameter space. How can the Interrogator do this? She has no control over the value of information, but she does have control over the torture costs to the Detainee.

In fact that’s all she has control over. She knows different human beings respond differently to the same torture so she will not know the value of k for any particular technique on any particular detainee. But she does know (“believe” is more accurate) that increasing the severity (i.e., ratcheting up k) of that particular technique for that particular detainee will lower the threshold. If she does it enough, she may tip the Detainee over the threshold to “information.” This is the inescapable and brutal logic of torture anytime and everywhere it is practiced. In other words we have:

Proposition 12.5 (Torture’s Slippery Slope 2). Once torture is admitted as an interrogation technique, the strategic incentives facing the interrogator result in increasingly brutal forms of torture.

In all likelihood, then, even a tightly controlled and regulated interrogational torture program will fail to match the predictive benchmark for torture severity and thus also fail to match the pragmatic model’s normative benchmark on limits on torture:

Benchmark 3 Torture Severity: The program succeeds if and only if torture is not employed too severely—well beyond the minimum degree necessary to compel valuable information.

Interrogational torture will result not only in too much torture, but severity exceeding limits and restrictions and thus reaching the level of abuse as defined by proponents themselves.

CONCLUSION

The outcomes derived from following the proponents’ own logic of interrogational torture fail to match their predictions. As a result, interrogational torture fails to meet the four normative benchmarks identified by torture proponents as justifying the practice. The information generated by torture is unreliable, but torture will be more frequent and brutal than even proponents envision and would accept.

This completes the argument:

1. EITs are torture and the effectiveness of interrogational torture is an open question. (Chapter 2)

2. The Bush program approximates closely the ideal model of interrogational torture and includes limits on torture; the Bush and ideal models provide benchmarks for comparison with the game theory models to come. (Chapter 3)

3. The Bush model generates strange, quixotic outcomes. (Chapter 4)

4. The Bush interrogational torture program is more realistically modeled as objective and leading question variants of an incomplete information game, with three types of detainees, two types of interrogators, and uncertainty about the amount and value of information provided. (Chapter 5)

5. By positing a set of Detainee strategies, calculating the Interrogator’s expected utility using Bayes’ Rule to identify her best response, and checking for incentives to deviate by any of the Detainee types, it is possible to derive a perfect Bayesian equilibrium in which a Detainee is tortured after providing information. (Chapter 6)

6. The RIT model generates nine perfect Bayesian equilibria, the formal and empirical characteristics of which generate important observations, propositions, and implications, including:

(a) The Interrogator’s thresholds for believing that a Detainee is Innocent after “no information” are less than one-half, with close to one-half and closer to zero.

(b) The Interrogator’s information hiding threshold under objective questioning is greater than or equal to one-half, whereas her information hiding threshold under leading questioning as well as the Detainee’s version of are a little less than one-half.

(c) Objective questioning (potentially) provides better information, but is necessarily accompanied by more torture, than leading questioning, which, however, provides less valuable information.

(d) All things being equal, Interrogators are more likely to get less valuable information than highly valuable information. (Chapter 7)

7. Surprise torture of a Cooperative Detainee—even if he has provided all his information—is not only likely, but perversely more likely

(a) the more willing the Detainee is to divulge information and

(b) the more important the information is in terms of “unknown unknowns.” (Chapter 8)

8. The perversity under objective questioning persists under leading questioning, namely:

(a) The more the Innocent Detainee believes Interrogator promises of no torture in exchange for confirmation, the more likely is ambiguous and false information.

(b) The more brutal the torture, the more likely is ambiguous and false information.

(c) The more important the confirmation is to the Interrogator and the more difficult it is for the Detainee to understand what is being asked, the greater the likelihood of torture.

(d) Leading questioning will not eliminate torture; the higher the interrogator’s demand for confirmation, the more torture. (Chapter 9)

9. The valuable information, selective torture equilibrium reveals four further trade-offs and a paradox:

(a) Everything else being equal, eliciting information

i. requires either more frequent or more brutal torture,

ii. is more likely when the information is less valuable or the torture is more severe, but not both,

iii. is more likely when the standard of cooperation is lowered, even though it increases the likelihood of information hiding,

iv. is necessarily accompanied by innocent torture; decreasing the likelihood of innocent torture requires decreasing the likelihood of information.

(b) The very assumptions justifying interrogational torture paradoxically make it less reliable in terms of the likelihood of getting valuable information. (Chapter 10)

10. From the two equilibria in which no Detainee provides information we learn that:

(a) Everything else being equal, torture becomes less likely (the equilibrium region shrinks to a smaller proportion of the parameter space) after “no information” only if the information becomes less valuable or the torture becomes more severe, not both.

(b) It is difficult to find real-world cases in which detainees subjected to the threat of torture in interrogations were not tortured after failing to provide information. (Chapter 11)

11. The results of the model violate the pragmatic model’s necessary conditions for justifying the practice—the ideal outcome never obtains, information is unreliable in both senses of the term, and torture will be both more frequent and more brutal than proponents expect and are willing to accept—thus refuting the pragmatic argument for interrogational torture. (Chapter 12)