Deliberation and Disagreement: Problem-Solving, Prediction, and Positive Dissensus. Hélène Landemore Yale University

Transcription

1 Deliberation and Disagreement: Problem-Solving, Prediction, and Positive Dissensus Hélène Landemore Yale University Scott E. Page University of Michigan and Santa Fe Institute Paper prepared for presentation at the Conference Epistemic Democracy in Practice, Yale University, New Haven, CT, October 20-22, 2011 THIS IS A DRAFT. PLEASE DO NOT QUOTE WITHOUT PERMISSION 1

2 Deliberation and Disagreement: Problem-Solving, Prediction, and Positive Dissensus Hélène Landemore and Scott Page Abstract: Consensus plays an ambiguous role in deliberative democracy. While it formed the horizon of early versions of deliberative democracy, it is currently under attack as an empirically unachievable goal and an undesirable normative ideal. We argue that this move away from consensus has so far generated more confusion than clarity and threatens to dissolve the normative appeal of deliberative democracy compared to the aggregative democracy model it was supposed to improve on. The paper further argues that, from an epistemic point of view, the normative appeal of consensus versus disagreement will vary depending on the deliberative context. We distinguish between two such contexts: problem solving and prediction. When it comes to problem solving, the Habermasian ideal of rational consensus retains a strong normative appeal as a stopping rule for deliberation. When it comes to predictive tasks, however, the search for a consensus can be epistemically counter-productive. Instead deliberators should seek to achieve or preserve a form of deliberative disagreement we call positive dissensus, which paves the way for more accurate aggregated predictions. 2

3 Deliberation and Disagreement: Problem-Solving, Prediction, and Positive Dissensus Hélène Landemore and Scott Page Deliberative democracy has long promoted the quest for a consensus on, if not truth per se, at least the better argument, to which the intrinsic force of the latter is supposed to guide reasonable participants in a deliberative process. Habermasians have thus famously promoted the search of a rational consensus as the goal of an ideal deliberative procedure, characterized as the ideal speech situation. Consensus, however, is now the object of some disillusionment and suspicion, including among deliberative democrats. The main criticism is that consensus is too utopian a goal to be worth pursuing. Others argue that building the goal of a consensus into the definition of deliberative democracy is counterproductive, empirically creating a pressure to conform conducive to group think, polarization, and generally detrimental to both the epistemic and procedural properties of deliberation. Critics of deliberative democracy finally from liberal pluralists to agonistic pluralists and difference democrats more radically argue against the artificial and even violent pacification of politics that a search for a rational consensus inevitably entails in their view. In this paper, we want to raise some skepticism towards the general goal of a rational consensus favored by deliberative democrats, although not on the grounds advanced by existing critics. In particular, while we think that there is something to the celebration of pluralism and disagreement advanced by critics of rational consensus and, specifically, agonistic pluralists and difference democrats, we will partially side with their conclusions as to the value of disagreement from within the framework of 3

4 deliberative democracy and on the basis of epistemic arguments that should, in fact, appeal to more traditional consensus-seekers. It is well known that the consensus following some deliberations need not indicate an epistemic improvement (e.g., polarization effect). This paper argues that conversely, the fact that a group ends up disagreeing more after deliberation than prior to it may, in some cases, mark a collective epistemic improvement and pave the way for a better collective decision. Our contribution in this paper is four-fold. First we argue that something has been lost in the deliberative democrats current move away from consensus and towards disagreement, absent any clear idea of what stopping rule should replace rational unanimity as the normative horizon of democratic deliberation. Second, we introduce what we think is a necessary and rarely made distinction between two main tasks that deliberators usually face: problem solving and predictive (also evaluative) tasks. Third, we show that while the Habermasian ideal of a rational consensus on an outcome retains normative appeal as a stopping rule to deliberation in the context of problem solving, it is generally less attractive, and in fact often counterproductive, for predictive/evaluative tasks. When it comes to making predictions, deliberative groups are better off fostering more of a certain type of disagreement rather than more consensus or convergence on an outcome. Our fourth contribution consists in distinguishing, within the category of deliberative disagreement (Gutmann and Thompson 1996) between the epistemically fruitful kind we shall call it positive dissensus and the epistemically useless or even harmful kind negative dissensus. 4

5 Although we direct our claims to an audience of political scientists and democratic theorists, the distinctions made in this paper will hopefully prove useful outside of political sciences. In particular, although we do not enter here the question of the nature of political questions and their possible differences with moral, scientific or other questions, our belief is that the differences are not such that they would preclude a generalization of our claims from one domain to the others. We see for example obvious parallels between the pluralism of the scientific community, for example in its effort to make predictions about climate change (e.g., Parker 2006), and the pluralism of the political community of a given country in its assessment of what is most conducive to the common good for that country, for example on the question of budget deficit. Both communities attempt to make predictions in a situation characterized by vast uncertainty, both tasks involve factual and normative judgments, and in both cases diverse and sometimes even incommensurable models are used to support conflicting conclusions. The first section of this paper assesses the current standing of the deliberative democracy literature on the question of consensus, diagnosing a general move away from consensus, specifically rational consensus, under the pressure of empirical results that challenge both the feasibility and desirability of actual consensus and normative theories that challenge the value of rational consensus, or any kind of consensus, as a viable ideal in the first place. The second section introduces the difference between problem solving and predictive tasks that allows for this more nuanced account of the value of consensus. and argues that rational consensus retains normative appeal for the task of problem solving. The third section turns to the predictive context. Here we distinguish between four degrees of deliberation, for which the search for a consensus turns out to be 5

6 generally less attractive. We conclude by distinguishing two types of deliberative disagreement: negative dissensus and positive dissensus. 1. Deliberative Democracy and Disagreement: Moving Away from Consensus The ideal of consensus as unanimous agreement on a decision was initially at the heart of the ideal of deliberative democracy. In a much quoted passage, Joshua Cohen, one of the first theorists of deliberative democracy, writes: Outcomes are democratically legitimate if and only if they could be the object of free and reasoned agreement among equals (Cohen 1989: 22). The meaning of free and reasoned agreement among equals in this passage is generally interpreted as a requirement of unanimity on a particular outcome (e.g., Bächtiger et al. 2010: 37, Dryzeck and Niemeyer 2006: 635), by contrast in particular with aggregative approaches to democracy in which a mere majority can impose their choice, without further reasoning, on a dissenting minority. In another famous statement considered as a founding principle of deliberative democracy, Habermas defined the discourse principle as the principle according to which: Only those norms can claim to be valid that meet (or could meet) with the approval of all affected in their capacity as participants in a practical discourse (Habermas 1990a, 66, our emphasis). Here the requirement of explicit approval by all is made the stopping rule for a legitimate decision or choice. While the discourse principle was initially developed in the context of Habermas general theory of communicative action, which was not specifically designed for a political context, Habermas explicitly retains consensus as the regulative ideal of deliberation in the realm of the law and democratic decision-making as well (1996: 179). Specifically, deliberators are expected, in the ideal speech situation where there are no 6

7 time and information constraints, to reach an uncoerced agreement on the better argument. 1 John Rawls is another author who made consensus a defining stapple of deliberative democracy, even though the nature of consensus in his own theories is somewhat ambiguous and the democratic nature of deliberation arguably a late development of his ideas. In A Theory of Justice, the unanimous agreement supposed to be achieved in the stylized situation of the Original Position a hypothetical situation in which deliberators are placed under a veil of ignorance to deliberate and reach a decision about the best principles of justice is essentially an artifact of the assumption that all deliberators are rational clones bound to think alike and converge on the same conclusions. Their consensual agreement, in other words, is a foregone conclusion. Rawls later substituted for what was deemed an overly monological definition of unanimous agreement the more realistic and dialogical ideal of an overlapping consensus between actual (rather than hypothetical) reasonable individuals (1993). 2 In its latest formulations, Rawls overlapping consensus arguably came to share even more traits of the Habermasian ideal of rational consensus (Rawls 1995, see also the analysis by McGann 2006: ). The centrality of consensus in these early and influential versions of deliberative democracy explains that both proponents and critics of the theory would converge in 1 Even though Habermas also acknowledges, particularly in later writings, the legitimacy of majority rule and bargaining (a strategic rather than communicative form of action), he firmly maintains the priority of rational discourse and consensus on both. 2 The concept of an overlapping consensus is distinct in that it recognizes the irreducible plurality of incommensurable worldviews characteristic of pluralistic liberal societies and yet assumes that even from within such comprehensive worldviews, individuals can share enough core values to achieve a stable common ground of principles, which is less than a rational consensus yet more than a mere modus vivendi (a temporary compromise between positions). For example, Lockean protestants, Kantians, and Millians can all endorse the principle of religious toleration on moral grounds, albeit from within irreducible moral perspectives (Waldron 2004: 95). 7

8 assessing that The core of the theory [of deliberative democracy] is that there would not be any need for an aggregation mechanism, since a rational discussion would tend to produce unanimous preferences (Elster 1986: 112). Or, more succinctly: The goal of deliberation is to arrive at consensus (Young 1996: 122). The legacy of this influential version of deliberative democracy can be seen in the title of real-life experiments in deliberative democracy, such as the famous Danish consensus conferences. Many criticisms have been raised, however, towards the ideal of consensus and specifically the Habermasian ideal of rational consensus. The first kind of criticism is that there is too much of a gap between the normative ideal and reality, to the point that the normative ideal loses relevance. In practice, as evidenced in actual deliberative experiments (e.g., List, Luskin, Fishkin, McLean 2000/2005), post-deliberative unanimity is rarely, if ever achieved. Full unanimity seems in any case utterly utopian in the case of deliberation on a mass-scale (although, there, a preemptive question is whether we can speak of deliberation in the first place). Critics suggest that it would be more useful to think up other ways to bring deliberation to a close than consensus, let alone rational consensus. 3 From a more theoretical perspective, some people are concerned that the ideal of consensus is not so much utopian as setting the wrong kind of goal for participants in a deliberation, in effect creating pressure on participants to reach an agreement. Social pressures and peoples non-rational conformism (Mackie 2006: 285) may induce people to converge on a non-rational, epistemically and even morally inferior consensus 3 This common criticism is, however, rather weak if one subscribes to the value of regulative ideals even when those can only be asymptotically reached. However descriptively unrealistic, the goal of getting closer to consensus might still retain some normative value. 8

9 (see Asch s 1951 for dispiriting evidence). 4 Other theoretical objections are based on studies that show how group polarization (a post-deliberative reinforcement of previously held beliefs in like-minded groups) tends to plague certain groups of deliberators (Sunstein 2003). These objections are not addressed to rational consensus as a goal per se, however, but more generally to the practice of democratic deliberation as potentially inferior, epistemically speaking, to mere judgments aggregation (see also Surowiecki 2004). 5 At a deeper normative level yet, many critics object to the very assumption behind the ideal of consensus, whether rational, overlapping, or otherwise, by denying that democracy has much to do, if anything at all, with communicative action. In the view of these critics, pluralism, specifically value pluralism, is a fundamental fact of politics and cannot be tamed through deliberation. Ian Shapiro thus provocatively argues that politics is only about power and interests so that calling for deliberation and consensus is at best a rhetorical device masking self-serving strategies, at worst an entirely useless and utopian endeavor (Shapiro 1999). In a distinct vein, agonistic pluralists or difference democrats (more radical versions of early liberal pluralists) see in deliberative democracy itself an instrument of domination imposing the communicative norms of powerful white men dispassionate, rational arguments on what should be an open forum to all forms of expression, including story-telling, greetings, and non-rational narratives (Mouffe 4 Again, it is questionable whether the fallibility of human beings is enough to refute the normative validity of rational consensus per se. The fact that people will tend to achieve an a-rational, conformist consensus instead of the normatively desirable rational consensus does not make the ideal less desirable. One may argue that there are institutional ways to remedy for the human tendency to deviate from the goal rational consensus so that the ideal need not necessarily be abandoned. 5 Such claims arguably refute only a certain type of communication, namely discussion among like-minded people (see Manin 2005 and Landemore and Mercier 2010 for the argument that this kind of exchanges should not count as deliberation per se). Again, like the objections above, it is in any case unclear what they prove as to the normative desirability of consensus, specifically rational consensus. 9

10 1999, Sanders 1997, Young 2001 and 2002). From this point of view, the ideal of rational consensus is seen as hegemonic and dangerous. Instead, agonistic pluralists and difference democrats propose to celebrate differences and disagreement. As a result of these battle-lines between being drawn and redrawn between advocates of consensus and advocates of disagreement, within and outside deliberative democracy, the current position of deliberative democracy as a field on the ideal of consensus is somewhat confusing. For recent commentators, we ought to distinguish two camps : on the one hand, the camp of Type I deliberation theorists, who stick to the Habermasian rational consensus as a valuable ideal, although they do so at some normative and empirical costs; and on the other hand, Type II deliberation theorists who have engaged into a more empirical program rid of rational consensus a program that has, however, several blindspots of its own (Bätchiger et al. 2010). What is for sure is that many deliberative theorists have taken their distance from the ideal of consensus, particularly rational consensus, and come to re-assess the value of disagreement. From a necessary evil, disagreement is now touted as an often desirable good (e.g., Gutmann and Thompson 1996; Thompson 2008: 508). Theorists have come to endorse the full legitimacy of stopping rules for deliberation that used to be considered as second-best of consensus, such as majority rule or other forms of judgment aggregation, or even the kind of non communicative and thus non- rational (in a Habermasian sense) agreement reached through bargaining. Type II deliberative democrats led by Jane Mansbridge for example argue that when interests and values conflict irreconcilably, deliberation ideally ends not in consensus but in a clarification of conflict and structuring of disagreement, which sets the stage for a decision by non-deliberative methods, such as aggregation or negotiation among cooperative organisms (Mansbridge et al. 2010: 68, our 10

11 emphasis). 6 In this approach, the ideal termination of deliberation is not agreement but disagreement, followed by a non-deliberative decision rule. The problem with these new developments is that they tend to blur the line between deliberative democracy and the original aggregative model of democracy on which deliberative democracy was supposed to be an improvement. If majority rule rather than consensus is the only realistic way to reach collective decisions, if disagreement, self-interests, and bargaining are re-introduced in the ideal of deliberative democracy or, conversely, the principle of deliberation introduced into the practices of bargaining (as in Mansbridge s (2009) concept of deliberative negotiation ), what is, then, the remaining specificity and advantage of this paradigm? The recent literature in deliberative democracy has, however, also produced some constructive distinctions, which do seem to preserve some specificity for the deliberative ideal and can help us move forward. It will be useful to briefly introduce them here. Guttmann and Thompson have thus proposed a distinction between deliberative disagreement and non-deliberative disagreement (Gutmann and Thompson 1996). Deliberative disagreement is, roughly, a disagreement that subsists even as the deliberation satisfies some valuable procedural standards: people listened to each other, expressed norms of reciprocity, equality, and respect and recognized the legitimacy of their opponents positions. By contrast, non-deliberative disagreements of the kind present in aggregative models of democracy simply express the brute confrontation of different judgments or interests without any efforts being made at convincing, 6 Mansbridge et al. (2010) not only give up on rational consensus as the normative ideal of deliberation but reintroduce in deliberation interests, power, and forms of strategic action such as bargaining. 11

12 understanding, or otherwise rationally engaging the opponent. Cass Sunstein has also usefully introduced the notion of incompletely theorized agreements (Sunstein 1995), in which the agreement is on a general principle but not necessarily on any supporting rationale or theory, or on each and every of its particular implications. Originally introduced in the field of legal theory, this concept nicely extend to deliberative democracy as well, where it offers a more realistic and generalizable goal than rational consensus or even overlapping consensus. Unlike Habermas rational consensus, an incompletely theorized agreement does not demand consensus all the way up and down, so to speak, on all the reasons supporting a particular decision and all the implications of this decision. Unlike Rawls overlapping consensus, which refers to the common core of values and principles concerning the basic structure of a just society shared by otherwise diverse and sometimes conflicting comprehensive worldviews, an incompletely theorized agreement can apply to any subject. The notions of meta-agreement or meta-consensus have also come to enrich the vocabulary of deliberative democracy. Christian List thus contrasts the ideal of substantive agreement a post-deliberative consensus on a particular outcome with what he analyses as the more useful category of meta-agreement. Meta-agreement is a post-deliberative consensus on how a problem should be conceptualized (List 2003). 7 For List, deliberation should aim at reaching meta-agreement rather than substantive agreement, both because meta-agreement is a more realistic goal and because it is also likely to facilitate further non-deliberative democratic decision-making through majority rule. List indeed suggests that in the two distinct contexts in which meta-agreement 7 Incidentally, for List, a meta-agreement is a more general case of what Sunstein calls imcompletely theorized agreement. 12

13 applies preference aggregation and judgment aggregation, meta-agreement may offer attractive escape-routes from the paradoxes and impossibility theorems of social choice theory and thus render possible the emergence of a meaningful collective decision. 8 John Dryzeck and Simon Niemeyer, in a similar vein, have developed the ideal of meta-consensus on values, beliefs, and preferences or a combination of them and proposed to use this ideal as a new foundational principle of deliberative democracy meant to replace Joshua Cohen s and other early deliberative democrats focus on strict unanimity on all three dimensions (Dryzeck and Niemeyer 2006: REF). Dryzeck and Niemeyer convincingly argue that it makes no sense to be for pluralism, against pluralism, for consensus, or against consensus (Dryzeck and Niemeyer 2006: 647) and that the ideal of meta-consensus can reconcile both camps. In other words, Dryzeck and Niemeyer embrace pluralism and consensus but at different levels: pluralism at the level of first-order values, beliefs, and preferences and consensus on either or all of these dimensions only at a second-order or meta-level. Finally, at a more empirical and descriptive level, researchers have recently identified a new category of consensus, which differs from strict unanimity and somewhat blurs the line with voting (Urfalino 2011, Steiner REF). Urfalino labels it apparent consensus (Urfalino 2012) or also, and more accurately, decision by nonopposition (Urfalino 2011). While unanimity rule is a well-known stopping rule for deliberation, generally contrasted with voting, decision by non-opposition is a rarely noted, yet empirically frequent decision-procedure (representing a third of observed 8 In the context of preference aggregation, meta-agreement may thus imply preference single-peakedness and the existence of a Condorcet winner. In the context of judgment aggregation, meta-agreement may similarly imply unidimensional alignment (List 2003) and the emergence of the median voter s judgment as a coherent collective decision in a proposition-by-proposition majority voting (List 2003: 12 and List 2010). 13

14 decision procedures in Steiner REF). It does not have the stringent requirement of unanimity and often takes place in the shadow of voting (Urfalino 2011, Novak 2010). Whereas the unanimity rule requires that all involved explicitly approve of a decision, which gives a veto power to those who refuse to, decision by non-opposition is a decision procedure by which deliberation ends not when all approve but when few enough people voice an objection and the rest remains silent, expressing consent, although not necessarily approval. The reasons why they express consent may be varied, including that they think they would be outnumbered if a vote were to be taken The main problem with the current state of the theory is that, because of the growing disaffection for consensus, combined with the concept stretching (Steiner 2008) that this occasionally leads to, there is, well, no consensus on what should take the place of consensus as the natural stopping rule for deliberation. This might not be a bad thing if this disagreement paved the way for a nuanced typology of when and where consensus or disagreement should be privileged, but this is missing as well. All the while, majority rule remains seen with some suspicion by many deliberative democrats who fear conceding too much to aggregative democrats. Indeed, if majority rule is all we can aspire to, why not just skip the deliberative phase, aggregate pre-deliberative judgments and call it a day? If the point of deliberation is not to bring us closer to consensus, why engage in it at all? 9 9 In their advocacy of alternative outcomes for deliberation, Dryzeck and Niemeyer, who present metaconsensus and intersubjective rationality (rather than truth) as substitute ideals to the original ideal of rational consensus (Niemeyer and Dryzeck 2007: 499) recognize that that these alternative outcomes do not qualify as collective decisions per se (Dryzeck and Niemyer 2006: REF), i.e., ways to bring deliberation to an end. Thus they acknowledge that Our emphasis on meta-consensus might seem to leave open the question of how collective decisions get made. But whatever the mechanism used (short of dictatorship), be it majority rule, approval voting, unanimity, bargained resolution, or agreements that majorities support and minorities can live with, meta-consensus makes collective choice more tractable. Sometimes tractability will enable agreement on what is to be done, but other kinds of outcome are possible (REF). List, for his 14

15 Another gap in the current state of the theory is that even those who celebrate disagreement and offer useful distinctions between different types of disagreement have little to say about the epistemic value of these different types of disagreements. In particular, assuming that disagreements can have epistemic value, in the sense of increasing the odds that the group reaches a more enlightened collective decision, is it always the case that deliberative disagreements, i.e., the kind that satisfies certain procedural criteria deemed valuable by deliberative democrats, have such epistemic value while non-deliberative disagreements don t? Similarly, is it necessary for an epistemically superior judgment to be supported by a meta-agreement or meta-consensus on some higher order set of preferences or judgments? We suspect that there is no necessary alignment of procedural and epistemic properties. It is more likely, on the contrary, that sometimes (although not necessarily often) a disagreement that violates norms of equality, respect, reciprocity is epistemically more fruitful than a deliberative disagreement that satisfies all these procedural criteria. A heated exchange between political adversaries might ultimately do more to change their views than a polite conversation that fails to prod deep enough into the sources of disagreement. Conversely, it is possible that a procedurally satisfying deliberative agreement could pave the way for terrible epistemic results. part, helpfully proposes that deliberation can be fruitful even if it falls short from yielding a consensus if the remaining post-deliberative disagreement corresponds to a meta-consensus on preferences or judgments that has the properties of preparing the way for meaningful collective decisions through majority rule. He warns, however, that this possibility depends on highly contingent empirical circumstances. While it has been empirically demonstrated that deliberation, at least in the context of preference aggregation, tends to bring preferences closer to single-peakedness (List 2010 and Farrar et al. 2010), this existing empirical research does not positively prove that deliberation actually reduces the likelihood of Condorcet cycles, that is can turn non-single peaked preferences into single-peaked preferences. Thank you to Sean Ingham (Harvard University) for the latter, important point. 15

16 The relation between epistemic properties of a given disagreement and the presence or absence of meta-agreement or consensus might be trickier to figure out and it would seem plausible that there be some correlation between the existence of a meta-consensus and the epistemic properties of the corresponding disagreement. But even if List (2010) is right to hypothesize that meta-agreements on preferences or judgments allow us to escape the paradoxes and impossibility theorems of social choice theory (respectively Arrow s impossibility theorem for preferences and the general version of the so-called Discursive Dilemma for judgments), the rationality of a collective choice according to the criteria of social choice theory is itself a distinct problem from the question of the epistemic value of that collective choice, which is our main focus in this paper. In any case, we think that the answer to these questions cannot be taken for granted. This paper does not ambition to answer all the questions at once. Our effort here will simply be to address the following: Is consensus, understood as rational, unanimous agreement on an outcome, ever epistemically desirable? If so, when and where? Conversely, when and where is disagreement, as the lack of rational and unanimous consensus on an outcome, epistemically preferable? We limit our inquiry by focusing on the epistemic properties of disagreement in different deliberative contexts Why Rational Consensus Retains Normative Appeal in Pure Problem Solving Contexts Let us introduce briefly the distinction between the two types of tasks that we argue deliberators can apply themselves to: problem solving and predictions. A prediction is 10 In the terms of Christian List, we place ourselves in the context of judgment aggregation and assume a general background consensus on preferences where necessary (which may be reached through deliberation as well). 16

17 some statement made about the future. It is an estimate of some future state of the world, for example Will the Egyptian demonstrations lead to a regime change? or How high will the unemployment rate be in May 2011? Predictions can thus be qualitative or quantitative. Quantifiable predictions are often referred to as forecasts. These include economic forecasts of unemployment and growth and meteorological forecasts of tomorrow s high temperature. Problem solving, by contrast, consists of finding or constructing a solution to a problem. It is a less clean-cut task than prediction. In some instances of problem solving such as how to send a person to the moon or Hilbert s fifth problem no answer exists at the time the problem is posed. In other instances, many putative solutions exist and the challenge is to select from among them. In this latter case, if the values of these solutions (i.e., how they rank compared to each other and in absolute terms) are not known, problem solving appears to be similar to prediction. To draw a bright line that separates the processes, we introduce the following formal definitions: A prediction consists of an object or event of unknown future value (for example an unemployment rate or the size of an uprising). Each predictor has a model that she uses to make an estimate of that future value. Problem solving involves finding the best from among a set of possible solutions. Individual problem solvers possess perspectives and heuristics that enable them to search part, but not all of the set of solutions (Page 2007). People will therefore differ in the solutions they propose. The will not differ, though, in how they evaluate proposed 17

18 solutions. In a problem solving context, any proposed solution can be evaluated accurately and at a minimal cost in time and effort. 11 In problem solving contexts, the process of deliberation can follow two scenarii: it can either allow a group to leverage their diverse representations and search heuristics to find the solution of a given problem or it can generate one from scratch by production of new ideas and recombination of old ones. Imagine first that the French government must choose a city to experiment with a new program. Three députés deliberate this choice, one from Calvados, one from Pas de Calais, one from Corrèze. They are aware of different possible solutions, which have each a different objective value for the experiment. The highest possible value is ten, which is achieved only by Caen (a solution initially not considered by the deliberators). Initially, each député applies her perspective and heuristics to the problem and arrives at a proposed solution. In the language of mathematics, these would be called local optima. The goal is for the group to find the global optimum, that is, the city with the highest objective value (Caen). Prior to the meeting, we can assume that each député has thought about the problem and formulated a solution (a city). Assume that these solutions, which are local optima are as follows (we have written their values in parenthesis): Calvados: Marseille (7) Corrèze: Paris (8) Pas de Calais: Grenoble (9) 11 A canonical problem given this construction would be the famous traveling salesperson problem, in which one seeks a minimal route between a set of cities. Finding the best solution to a traveling salesperson problems with even one hundred cities is not easy, but determining the length of any particular route takes little time or effort. 18

19 Initially, the député from Calvados proposes Marseille because her way of representing the problem led her to think of Southern coastal cities. However, the député from Corrèze pushes Paris, which all three immediately agree to be the better choice. The député from Pas de Calais then suggests Grenoble, which all three again accept as an improvement. If the deliberative process were to stop here, deliberation would create no synergies. It would be merely a process of picking the best from a collection of proposals. Our point is that the deliberative process need not stop here, and typically won t. The reason is that the other two députés, those that did not propose Grenoble, may now have new ideas. They may not, for example, have considered medium sized cities. So, if in proposing Grenoble, the député from Pas de Calais mentions the expense of running the project in a big city, he might well spur the député from Calvados to think of other medium sized cities, including Caen (value 10) that are even better choices than Grenoble. The logic that underlies these iterative improvements rests on the diversity of perspectives. Neither Caen nor Grenoble initially entered the mind of the député from Calvados because she was comparing only large cities. It is not that she could not imagine a medium size city, cleary she could, only that her way of solving the problem led her to be searching larger cities. Once a moderate sized city, in this case Grenoble, was proposed, she was then able to think of other moderate sized cities, including Caen. Steven Johnson (2010), in writing about innovation, refers to the opening of these new 19

20 potentials as the adjacent possible. In our formulation, deliberation enlarges each deliberator s adjacent possible and leads to more and often better solutions. Deliberation among those three people has epistemic properties that deliberation among less cognitively diverse people would lack. The pool of information was enlarged, as the député from Calvados, who only knew about one local peak (Marseille), ends up knowing about the qualities of Paris and Grenoble as well. The députés from Corrèze and the député from Pas de Calais similarly learn about two other local peaks than the one they could think of (respectively Marseille and Grenoble, and Marseille and Paris). Notice that even if the information gained is sometimes of lesser objective quality than that which the person already held, nonetheless, only by acquiring it can the members of the group reach the highest local optimum with certainty. The député of Calvados might never have considered an option she actually knew about, her own city of Caen (10), if she had not been spurred away from her initial choice (of value 7) by the other two députés who offered better but still suboptimal solutions (of respective values 8 and 9). Deliberation also allowed the group to weed out the good arguments from the bad. While it seemed at first a good argument to look for a big city (Marseille, Paris), it turns out that it was better to look into moderate sized cities (Grenoble, Caen). Finally, deliberation led to a consensus on the best solution, namely the solution that allowed the group to reach the optimum of 10, when the pre-deliberative beliefs about the best solution could have been respectively 7, 8, and 9. By contrast, if all three députés were thinking exactly alike say like the député of Calvados who thinks of her lower local optimum first no matter how long they deliberated, their group would stay stuck on the local optimum of Marseille (7) and 20

21 would never be able to reach the higher local optimum of Caen (10) because nothing would have spurred them away from that low local optimum. Thus, the extent of deliberation as well as the potential benefits may well increase with the diversity of problem solving approaches. Let us now turn to a second example of more creative problem solving. Here deliberation does not simply reveal a pre-existing solution but builds it from scratch on the basis of shared information, arguments, and ideas. A neighborhood faces a recurrent safety issue on a dark bridge that separates it from the city s downtown. Muggings occur with sufficient regularity that people have become scared to walk home after dark. All members of a committee formed to solve this problem share the same objective: to stop the muggings but to do so a minimal costs. Deliberation here can play three roles: it can produce ideas, it can refine ideas, and it can combine ideas. Prior to the meeting it could be that no one has any concrete proposals. Through dialogue three solutions to the problem might emerge. A: Have the neighbors walk each other home and organize watches; B: Station a police car near the bridge after dark; C: Install public lighting on the bridge. Once these ideas are on the table, deliberation can refine and improve the solutions. When is the police car stationed? How do the hours change depending on the time of year? Can someone call to ask for the police car? Keep in mind that we assume that proposals can be evaluated accurately, so we have no disagreement about whether or not an idea is good. Finally, ideas can be combined. Suppose that proposal C has the lowest cost but, unlike the other two, it won t reduce muggings to zero. The existence of proposal B, the costly police car option, might lead someone to propose option D to 21

22 complement the lighting with an emergency phone. Proposal D might then be the option selected. The role of deliberation in pure problem solving contexts then is to get ideas on the table, improve them, and sometime combine them to create ever more ideas. Disagreement would mean to ignore another s proposed solution, even when its value is known to be positive. Given that everyone shares the same values and that all solutions can be evaluated, disagreement would be irrational and epistemically disastrous. There is simply no reason not to reach a consensus. Of course, one might object that our characterization of deliberation in problemsolving contexts assumes an oracle, that is a machine, person, or internal intuition that can tell us the value of proposed solutions. One might think that only for rare problems, and these would tend to be more engineering based problems, would such an oracle exist. We disagree. Often the values of competing solutions are relatively self-evident, at least in relative if not absolute terms. For example, in the example developed here, the solution Improve public lightning on the bridge and put in an emergency phone is straightforwardly better by intuitive and obvious standards such as simplicity, cost, and long-term efficiency than alternatives such as having neighbors take care of the problem or have a police car permanently stationed by the bridge after dark. We assume that the self-evident nature of some solutions correspond to what Habermas called the unforced force of the better argument. The unforced force of the better argument is, indeed, a form of oracle. Another objection would remark that in many situations the problem solving context is impure. Supposethat we have an oracle but that outcomes are 22

23 multidimensional. Specifically, consider a status quo policy A that we want to improve on. Assume that each proposed policy produces some level of equality and efficiency. To keep this simple suppose that equality and efficiency each have scores from 1 to 10 where 10 is better. A can be characterized as: A: (4,8) After a thoughtful deliberation in which we have used our various perspectives and heuristics, we come up with an alternative policy B, characterized as: B: (6,5) Obviously, B is better than A on the equality dimension but worse than A on the efficiency dimension. If we have preference diversity among the deliberators, we won t reach a consensus. If there is third option on the menu, say C (5, 7), we could even cycle as a group: we could prefer A to B to C and then C to A. This scenario won t allow us to find a peak. We will just roam around. The objection is right to remark that in real life problem solving will tend to be mixed rather than pure. In fact, the objection invites us to distinguish between two types of consensus. PREMISE CONSENSUS: do we agree on the mapping F? PREFERENCE CONSENSUS: do we agree on our rankings of outcomes in Y? It is certainly useful to distinguish analytically between the epistemic dimension of problem solving figuring out a peak or agreeing on the mapping F and the alignment of background preferences problem solving also requires agreeing on our ranking of outcomes in Y, that is reaching a meta-consensus on the dimensions to be 23

24 privileged when ranking outcomes. Our point in this paper, however, is limited to showing what deliberation can do for the latter case. Even if we assume a meta-consensus on preferences (a preference consensus) and the existence of an oracle, there are also situations in which time constraints simply do not allow for the emergence of a solution. In such cases, the solutions have to be evaluated in some way and the task facing deliberators also involves figuring out which of the available options is the most likely to succeed. This means that problem solving often bleeds into prediction. Consensus on the better solution or the unforced force of the better argument cannot always be expected to end the process. If so, some other way to end the deliberative process than consensus as unanimous agreement must be agreed on. Alternative stopping rules for deliberation that is, stopping rules that produce a decision are then either what Urfalino calls decision by non-opposition i.e., an apparent consensus in which those who disagree with the majority refrain from expressing disapproval, or a vote, in which everyone can voice their judgments but settles the issue by a majoritarian decision. It is generally the case that decision by non-opposition takes place in the shadow of the vote. In other words, the reason by dissenters refrain from voicing their disagreement is because they know they would be outnumbered if a vote were to be taken and they may have reasons not to want to be exposed as dissenters if nothing is going to come out of it (see Novak 2009). If there is uncertainty as which group would win or the dissenters refuse to surrender without an actual count of their voices, the group can go right away for the aggregation mechanism. What happens then? For example, assume that the group couldn t reach a unanimous, rational agreement or even an apparent consensus on one 24

25 solution but is equally split between two rival ones, say A and C. In this case, straightforward majority rule can be, under some assumptions, a fairly reliable way to generate an accurate collective prediction (see CJT but also, e.g., Page 2007, Landemore 2007, and Lupia and McCubbins 1998). There are several problems, however, with such a move. A first problem is that, in many cases, there are more than two options to choose between, which raises social choice theoretical problems such as the possibility of cycling majorities. For 3 options such as A, B, and C, it could be the case that in pairwise comparisons A beats B, but C beats A, and B beats C, so that no clear outcome can be determined. What is needed to avoid such problems might be more deliberation: ideally the group would deliberate until they reduce their options to two or at least considerably increase the single-peakedness of their preferences, if it can indeed be assumed (as suggested by Miller 1992, Knight and Johnson 1994, Dryzek and List 2003, List 2010) that an increase in single-peakedness correspondly decreases the probability of cycling majorities in situations where there does not exist a Condorcet winner. Second, even if deliberators manage to reduce the options to two or to somehow minimize the risks of cycles in the case of three options or more, it could be that the raw individual predictive models are not terribly accurate, despite all the problem-solving deliberation that went on before, and could be improved on by more deliberation of a different nature. In other words, instead of ending deliberation after the problem-solving phase, the group might be better off starting a new phase, which would consists in exchanging fresh reasons as to why such or such solution is likely, in each person s view, to perform better. 25

26 In our example, participants in the deliberation no longer confront the question What could help us solve the mugging problem on Court Bridge? but they are asked to answer a different one: Given that the group is divided between options A and C, which do you think is most likely to get the job done? The exchange of arguments is likely of course to overlap somewhat with the deliberative content of the problem-solving phase (after all, the deliberation here still consists in trying to give arguments for A or C) but it might also bring about completely new arguments. Should deliberators in the phase concerned with predictions seek to reach some kind of rational consensus on the better prediction, in the same way that deliberators in the problem-solving phase aim to converge on a common solution? Before we turn to the analysis of deliberation in predictive contexts, let us emphasize the following point. In many, perhaps most cases of interest, deliberation involves both problem solving and prediction. In our examples, as in any political example involving the formulation of a policy, the task involves both problem solving the crafting of new potential solutions and predicting or forecasting the effects of these solutions. Policy-making in general, or so we argue, involves both problem solving and prediction. We now turn to prediction. 3. Why Disagreement Can Pay Off in Predictive Contexts Recall that the predictive context involves for deliberators the task of making an estimate of the future value of an object based. In our construction, we presume that individuals bring to this task some amount of information as well as a cognitive model of the situation at hand. Before we turn to the question of what deliberation can do to improve 26

27 individual and collective accuracy, let us posit that the accuracy of aggregated individual predictions depends on two equally important factors: individual accuracy and collective diversity (Hong and Page 2009, Page 2007). Specifically, the accuracy of an individual equals the distance between her prediction and the actual value. A priori the accuracy of individuals won t be known exante. We can nonetheless assume that an individual has an expected accuracy, which we can assess based on historical data documenting her past predictive accuracy or on an exante analysis of the quantitative or qualitative model that she uses. For example, we can ask: What is the past predictive record of a Fox News pundit? Or, What kind of regression equation does this analyst use to predict unemployment rate and does his choice of variables seem plausible? A collection of predictors will exhibit some degree of heterogeneity or diversity. We can measure this at the level of their models do they use the same variables and categories and similar functional forms? or through the statistical properties of their predictions. If we have data on past predictions, correlation serves as a useful proxy for diversity. More diverse predictors make less correlated predictions. On any given instance the accuracy of the equally weighted collective prediction depends on the accuracy of the individual predictions and on their diversity according to the following formula: Average Collective Error = Average Individual Error Diversity of Predictions 27