The Tyranny of the Ideal

by Gerald Gaus

  80 Popper, The Open Society and Its Enemies, vol. 1, p. 160.

  81 Elster, Logic and Society, p. 57.

  82 Ibid., pp. 57–58.

  83 Wilde, The Soul of Man under Socialism, p. 40. See also Goodwin and Taylor, The Politics of Utopia, pp. 213ff.

  84 Or, as Bellamy puts it, “ever onward and upwards.” Looking Backward, p. 1.

  85 Sidgwick, The Methods of Ethics, p. 22.

  86 Popper, The Open Society and Its Enemies, vol. 1. p. 161. Compare Bacon. We are informed by the lawgiver of New Atlantis: “And recalling into his Memory, the happy and flourishing Estate, wherein this Land then was; So as it might bee a thousand wayes altered to the worse, but scarse any one way to the better; thought nothing wanted to his Noble and Heroicall Intentions, but onely (as farr as Humane foresight miught reach) to give perpetuitie to that, which was in his time so happily established” (New Atlantis, p. 22).

  87 Simmons, “Ideal and Nonideal Theory,” p. 35.

  88 See, e.g., Rawls, Justice as Fairness, pp. 105, 115, 118n, 119, 194. Despite the repeated use of this phrase, Rawls probably does not have in mind anything quite so Platonic. “Political liberalism … does not try to fix public reason once and for all in terms of one form of favored conception of justice.” “The Idea of Public Reason Revisited,” p. 582. See further §IV.1.3.

  89 Consider how quaint the utopias of Plato, More, Bacon, or Bellamy strike us today. Of course some believe that we are at the end of history, and their theory of the ideal has glimpsed the owl of Minerva. Popper was quite right to see historicism as a complement to Platonism (The Open Society and Its Enemies, vol. 2).

  90 Mill, Chapters on Socialism, p. 737.

  91 Mill, Chapters on Socialism, p. 737.

  92 For Robert Owen’s own account, see his A New View of Society.

  93 Haworth, “Planning and Philosophy.”

  94 Kumar, “Utopian Thought and Communal Practice,” p. 18.

  95 Ibid., p. 19. A grave problem, however, is to distinguish endogenous causes of collapse (which suggest problems with how the social world was ordered) from exogenous ones, e.g., the environment in which the experiments took place. That the Ralahine experiment failed after its “proprietor, John Scott Vandeleur, gambled away his fortune in the clubs of Dublin and fled the country to escape his creditors” can hardly be said to show us much about the viability of Owenism. Haworth, “Planning and Philosophy,” p. 151.

  96 Haworth, “Planning and Philosophy,” p. 153; Kumar, “Utopian Thought and Communal Practice,” p. 18.

  97 Quoted in Haworth, “Planning and Philosophy,” p. 152.

  98 Without this prohibition the groups develop different perspectives, rather than exploring the same one. As we will see presently, and especially in the next chapter, different perspectives can be of great use in solving optimization problems, but their benefits and problems are different from models in which the teams share all the elements of a perspective, and simply explore different parts of the same optimization space.

  99 For an extended discussion of an example from management, see D’Agostino, “From the Organization to the Division of Cognitive Labor.”

  100 Although, of course, regression and other statistical techniques can allow us to draw some useful inferences from natural experiments.

  101 William Owen, Diary of William Owen, p. 129.

  102 Kumar, “Utopian Thought and Communal Practice,” p. 1.

  103 For example, before the North American Free Trade Agreement was launched extensive and varied modeling was used to predict effects; even what seemed like consensus conclusions of the models often turned out quite wrong on critical matters. See Shikher, “Predicting the Effects of NAFTA.”

  104 Page, The Difference, p. 208. Emphases in original.

  105 Surowiecki discusses the example of the search in 1968 for the lost United States submarine, the Scorpion, in which diverse predictions within a group as to its location were aggregated to arrive at group prediction that was accurate to within 225 yards. The Wisdom of Crowds, pp. xx–xxi.

  106 D’Agostino, Naturalizing Epistemology, p. 138.

  107 Ibid.

  108 Page, The Difference, p. 286.

  109 Wagner, Zhao, Schneider, and Chen, “The Wisdom of Reluctant Crowds.” See also Sunstein, Infotopia, chap. 1.

  110 D’Agostino, Naturalizing Epistemology, pp. 138–41.

  111 This subsection draws on work that I conducted with Keith Hankins. I thank him for permission to use it.

  112 See Hong and Page, “Groups of Diverse Problem Solvers Can Outperform Groups of High-Ability Problem Solvers”; Hong and Page, “Problem Solving by Heterogeneous Agents.”

  113 On the other hand, even this very modest degree of diversity can lead to problems of communication, a worry that will occupy us in the next chapter. If we deeply disagree about how to measure similarity (or distance) between social worlds, a modification to some relevant feature of the world that I consider to be relatively minor might appear quite radical to you. For instance, I might ask you to imagine a world that is otherwise like ours, but in which people are slightly more equal, though at the cost of being slightly less free, and I might judge that world to be superior to our own. If you have a different conception of what counts as slightly less free, though, you might imagine an entirely different world—one which you, reasonably, might think is much less just than our own—and in this case, it is almost inevitable that we will find ourselves talking past one another.

  114 See appendix A, point (iii).

  CHAPTER III

  The Fractured Ideal

  Searching with Diverse Perspectives

  And they said, Go to, let us build us a city, and a tower, whose top may reach unto heaven. …

  And the LORD said, Behold, the people is one, and they have all one language; and this they begin to do. …

  Go to, let us go down, and there confound their language, that they may not understand one another’s speech.

  —GENESIS 11

  1 ATTAINING THE IDEAL THROUGH PERSPECTIVAL DIVERSITY

  1.1 From Full to Partial Normalization

  IN THE PREVIOUS CHAPTER WE SAW THAT A SINGLE, FULLY NORMALIZED perspective confronts The Choice: it must choose between local improvements in justice and the pursuit of the ideal. The necessity for The Choice follows from the very core of ideal theory, that the Social Realizations and Orientation Conditions can give different answers as to what social states are “closer” to ideal justice. The Social Realizations Condition measures the justice of a social world in terms of its inherent justice (so, in one sense, the more just a social world, the closer to the ideal), but the Orientation Condition measures proximity to the ideal in terms of the similarity of the underlying structure of the social world, and so a reform that moves us closer to the ideal on the Social Realizations Condition can lead away from it given the conception of proximity expressing the Orientation Condition. The Choice is made much more troubling by the Neighborhood Constraint. If we had comprehensive knowledge of the entire landscape of justice, we would at least know, when we turn our backs on local improvements, just where the ideal lies, and just how ideal it really is. Perhaps the main worry then would be the feasibility of getting to the ideal (§II.1.3).1 But the Neighborhood Constraint implies that we almost certainly do not have such knowledge; we know near social worlds better than far-off ones. If the ideal is not in our neighborhood, no single perspective can be very confident just where, or what, it is.

  As I have noted, all too often ideal theorists respond to this with sheer denial. We do, they assert, have comprehensive knowledge of far-off social worlds; we know how utopia would work as well as how America in 2016 does. I join Popper in dismissing this as sheer delusion; the case for deep uncertainty in our understanding of the workings of far-off worlds is overwhelming. However, there is a much more sophisticated response available to the ideal theorist. A proponent of the ideal may acknowledge that we do not have comprehensive knowledge of th
e landscape of justice—no single perspective on justice could ever have a complete knowledge of the justice of all social worlds. Every perspective is always learning, searching the landscape, trying to find better optima—better utopias. At the close of the last chapter we saw, though, that a diversity of perspectives can mitigate some of the constraints faced by a single perspective in searching the optimization landscape (§§II.4.2–3). This insight has been elegantly and powerfully developed by Lu Hong and Scott E. Page, who demonstrate that—under what, prima facie, seems to be a modest set of conditions—a certain group of perspectives that agree on some parts of perspective Σ but disagree on others will locate the ideal.2 Specifically, what the work of Hong and Page suggests is that such a group of perspectives is one that concurs on the core elements of Σ’s normative theory of justice, but disagree on how they understand the similarity of social worlds in the domain {X}. Let us call such a group of perspectives evaluation-normalized versions of Σ. These perspectives all concur on the evaluative standards (ES), justice-relevant features of the worlds to be evaluated (WF), and the mapping function (MP). They disagree, however, in the similarity ordering of the worlds (SO) and the distance metric (DM), which determine the similarity of any two social worlds in {X}. Recall that ES, WF, and MP jointly satisfy the Social Realizations Condition, which requires that social worlds must be compared in terms of their inherent justice (§II.1.1); these have a sound claim to being deemed the core value elements of a perspective. Those who agree on these three elements agree on what each social world looks like (its justice-relevant features) and the overall justice of each world. Looking at some world, they would always agree on its features (WF) and on how just it is (ES, MP). Thus core normative agreement is secured by agreement on these three parts of a perspective. Let us denote all perspectives that share these evaluative elements of Σ, ΣV. Perspectives that include ΣV can have radically different similarity orderings of the domain of worlds to be evaluated and distance metrics; in terms of our basic landscape model, all perspectives in this evaluative core of perspective Σ agree precisely on the justice score of every world, so they concur about the placement of every world on the y-axis; but they may disagree about any world’s location on the x-axis. Given this, Hong and Page’s analysis would appear to show that members of ΣV, working together, will outperform any single member of the ΣV family in locating the ideally just social world, as all members of ΣV understand it. Thus an ideal theory that refuses to fully normalize its perspective, but opts instead for only evaluation normalization (ΣV), seems able to find the ideal without making claims that any single perspective (one composed of all five elements) has comprehensive knowledge of the landscape of justice.

  1.2 Diversity of Meaningful Structures and Finding the Ideal

  Before turning to the details of Hong and Page’s analysis, we can get an intuitive grasp of how a team of sophisticated ideal theorists (who share an evaluative core of a perspective) might employ their work by considering a real-world example. Suppose that instead of searching for the most just social world among a set of possible worlds, our team is searching for the most just state within a domain or set of actual states. If they are genuinely searching, of course, they do not yet know which is the most just state. They know that there is a most just state in the domain, but not what member it is. Because team members concur on the evaluative core of a perspective (state features, evaluative standards, and mapping function), they all concur on where any given state should be placed on the y-axis. However, because they do not share the similarity ordering and/or distance metric, they may disagree on a state’s location on the x-axis. Suppose one member of the team has a straightforward economic similarity metric, which arrays states simply in terms of per capita GDP; a state is similar to another if and only if it has a similar per capita GDP.3 Figure 3-1 uses this perspective to search for the most just state among a group of twenty selected current states: Brazil, Bulgaria, China (PRC), the Czech Republic, Guyana, Haiti, Honduras, Jordan, Macedonia, Madagascar, Mexico, Moldova, Pakistan, Romania, Russia, Saudi Arabia, Senegal, Serbia, Swaziland, and Zimbabwe. Justice is understood in the classical liberal sense of the best protection of individual rights and autonomy.4 (Note: figures 3-1 to 3-3 label only the states in the group that are “local optima” on a perspective—states at the top of a gradient.)

  Figure 3-1. Searching for justice on an economic perspective

  Note that on the economic perspective in figure 3-1 there are nine local optima. Although there is certainly some correlation between high GDP and protection of individual rights, seeking justice on this perspective does not create a smooth optimization problem. Suppose a perspective searches for the ideal via a simple climb-the-gradient strategy: if it finds itself on a slope, this strategy instructs it to climb up to the peak, but it will never go down a gradient.5 Such a procedure will thus never embark on a path that leads to worse results—it “climbs” only “upward.” The downside to this strategy, however, is that it will get stuck at the first local optimum it comes to, which are rather abundant on this perspective. To be sure, if the search commences with the very richest states, the global optimum (the Czech Republic) will be quickly hit upon; but from any other starting point there is a deep valley between the global optimum and all other states. Suppose that another member of the team has a different, more libertarian, perspective that arrays states according to their economic freedom, as in figure 3-2.6

  Note that for every state, this perspective fully concurs with the first with regard to justice scoring. This perspective, however, employs a different similarity metric; as a result it eliminates local optima at Honduras, Moldova, and Madagascar, while adding optima at Mexico and Guyana. Assume the simple per capita GDP perspective is stuck at Honduras, Moldova, or Madagascar and it asks the economic liberty perspective, “Can you see any way that goes up from here?” The economic liberty perspective will, since it will still be on a gradient; in contrast, the GDP perspective will never get stuck at Mexico or Guyana, so the two perspectives can assist each other. Nevertheless, we have quite a few shared local optima; if they are both on one of these neither can see a way upward. Now add a third member of the team, employing a more political perspective. On this third perspective the underlying structure of the relevant aspects of states concerns how well their government functions, whether its elected leaders determine economic policy, whether the government is free from corruption, and so on.7 Figure 3-3 gives this perspective on the problem.

  Figure 3-2. Searching for justice on a libertarian perspective

  The only optima shared by all three perspectives are the Czech Republic, Brazil, and Serbia: if each perspective can rely on the other members of the team to get it over its own local optima, (and, we shall see that is a very big “if”), our diverse team is bound to at least get as high as 13 on the individual rights protection scale. (At Brazil or Serbia none of the three perspectives can see a way upward.) Notice that alone, no single perspective can be assured of doing that well. This is true even though the third perspective is in an obvious sense the best perspective; its understanding of the similarity of states tracks the underlying structure of the rights protection problem better than the others (it has fewer local optima). As we shall see, this is an important point: under certain conditions a diversity of perspectives does better than even the best alone.

  Figure 3-3. Searching for justice on a government functioning perspective

  1.3 The Hong-Page Theorem

  The crux of our example is how searching a rugged landscape can be improved by what we might call a handing-off-the-baton dynamic.8 A perspective takes the baton (it takes charge of the search) and runs uphill as far as it can, to a local peak. But on some other perspective this is not a peak, so the baton can be handed off to it, and then it will run uphill until it comes to one of its local optima, where it will hand off the baton to yet another perspective that is still on an upward gradient. Hong and Page have identified conditions under
which this general type of dynamic is guaranteed to find the global optimum—when the ideal is guaranteed to be discovered.9 At least in some of its core suppositions, their theorem fits well with the model of ideal theory developed in chapter II, as they (i) suppose precisely the sort of rugged landscapes that underlie ideal theory (§II.2.4.) while (ii) showing under what conditions a diverse community will locate the ideal. They thus appear to prove the conditions that guarantee locating utopia! No small feat.

  Page provides an excellent informal discussion of the conditions of the proof.10 (i) The optimization problem has to be sufficiently difficult such that no single perspective always finds the global optimum. I stressed this point throughout chapter II; if the optimization problem is smooth, then a single perspective will climb to the global optimum on its own. But then, as I have stressed, orientation via the ideal is not necessary for the search for justice. (ii) At the same time, the problem cannot be too complex. Page glosses this as “the problem solvers are smart,” and this parsing is at the heart of Hélène Landemore’s use of the Hong-Page theorem in her excellent account of democratic deliberation—indeed, as she puts it, the problem solvers cannot be “too dumb.”11 Recall our perspective on height, which sees the underlying meaningful structure in terms of the alphabetical ordering of first names (§II.2.1). This perspective created maximal ruggedness; the height measure (y-axis), is uncorrelated with the name ordering (x-axis). It thus gets constantly caught at local optima; a group of perspectives like this cannot make much progress on the problem of finding the tallest person. However, this assumption is not really only about whether perspectives are “dumb” but also about whether the problem is “easy enough.” The real question is whether perspectives understand the problem as posing a very high-dimensional landscape (a maximally rugged one) (§II.2.2). In such landscapes there really are no gradients (each point’s justice is uncorrelated with its neighbors), and so no one can really make progress in solving the optimization problem. In maximally rugged landscapes there are a great number of poor local optima at which the group can get stuck.12 Under conditions of maximal complexity, then, increased diversity does not promote optimization.13 We can understand the gloss of this condition as the problem solvers are “not too dumb” in the sense that they find the problem sufficiently tractable—they possess perspectives in which the underlying structure is correlated with the optimization problem. Nevertheless, I believe it is better to focus on the question of complexity: if we are confronting Kauffman-like complexity catastrophes (§II.2.2) we may all be “dumb.”