The Tyranny of the Ideal

by Gerald Gaus

  To generalize the lessons of the Owenite tale, in order for actual small-scale experiments to help the proponents of perspective Σ to better know the social worlds in domain {X} almost all the elements of Σ’s perspective would have to characterize each experimental group. They would have to concur on evaluative standards, on an understanding of the relevant features of the social world, on the trade-offs of values characterizing the (second part of the) mapping relation, on an understanding of the similarity of the underlying structures, and on the distance metric (§II.1).98 They would, presumably, omit the modeling part of the mapping function (element [i]), as they are engaging in actual experiments to discover the social realizations of these worlds. Enthusiasts of social experiments often fail to perceive that the very value of the experiment depends on not permitting a number of variables to be altered; if the basic features of the perspective can be altered, the results of the experiment will not be helpful in orienting a particular perspective’s pursuit of justice. It is noteworthy that Owenite communities such as New Lanark and Ralahine were owned by proprietors or subscribers; this sought to convert the social experiment into a management exercise, where the manager (owner) sets the criteria for success and leaves the employee teams to find the optimal solutions. Owen was prescient: these are precisely the cases where it can be shown that different teams exploring the same problem can yield real benefits.99 The great barrier to social experimentation, as the fate of the Owenite communities suggests, is to stably maintain this model when the experiment is about the pursuit of justice.

  Needless to say, the problems of small-scale social experimentation are even more daunting if it is supposed that these small-scale experiments are informative about the application of the perspective to large-scale societies. If they are so intended, it will often be doubtful that their results scale up; large-scale societies will have features that small-scale societies lack, so it will be uncertain how well their lessons apply to large-scale societies. Moreover, unless we have a rather large number of experiments, we will not be able to employ statistical techniques for judging our inferences, which will be based more on hunch. If we try to avoid the problem of inferences from small-scale experiments by conducting large-scale ones, then we know that the number of experiments will be very limited, and so drawing inferences from such small-n experiments will generally be uncertain.100

  4.2 Improving Predictions: Diversity within, and the Seeds of It between, Perspectives

  Perhaps Owen’s son had it right: “the enjoyment of a reformer is … much more in contemplation than reality.”101 Not all investigations of alternative social worlds need to engage in actual social experimentation. As one scholar has put it, “Just as the artist invents imaginary worlds, so the social theorist invents pure states of society.”102 Of course our problem is that, while in one sense the social theorist is “inventing” a social world in the model, the theorist is also trying to discover its justice so as to make sound recommendations about where utopia lies—and to do that, the theorist needs to figure out how the recommended social world will function. An ideal theorist seeks to understand far-off social worlds and then report back to the rest of us on how they function, and how just they are.

  We saw that actual social experiments more or less bracket one part of a perspective on justice—the modeling of worlds and their social realizations—which is replaced with social experimentation. This suggests a way forward for the utopian: an ideal theory might employ multiple predictive models, and see when these different models agree and when they diverge. In this sense we can think of an internally diverse perspective, one that adopts a variety of ways of modeling social worlds, and which seeks to combine them to arrive at an overall prediction of the way a world might work. One case seems especially clear: namely, when the perspective’s models almost converge on estimates in the same small range. To be sure, even here we can go wrong, but our confidence in our estimates will be much higher.103 If we suppose the approach to understanding complex systems that we considered earlier (§II.3.2), convergence of models is most expected up to the borders of our neighborhood. Take again a case drawn from meteorology—hurricane prediction. Predictions typically draw on a number on models based on very different methods and assumptions. As those of us who have lived in New Orleans and have closely followed the models in the summer months know, they usually have markedly diverging results three to five days ahead but converge on one-to-two day predictions. This is precisely what we would expect in modeling complex systems: the further we get from the observed system the more even our best models diverge.

  However, an internally diverse perspective can improve the reliability of its predictions even without such convergence. Scott E. Page has stressed what he calls the Diversity Prediction Theorem, according to which collective predictive error = average individual predictive error minus predictive diversity. The upshot of the theorem (explained in appendix B) is that “individual ability (the first term on the right-hand side) and collective diversity (the second term) contribute equally to collective predictive ability. Being different is as important as being good. Increasing diversity by a unit results in the same reduction in collective error as does increasing average ability by a unit.”104 Although an excellent predictive model can still beat a collective prediction, the theorem tells us that we can compensate for an error in our predictive model by employing a greater diversity of models, and essentially averaging the result. This is an important theorem: even if our predictive models are not very good, a perspective that draws on diverse predictive models can significantly enhance its confidence in its estimates of the justice of alternative social worlds. Predictive diversity thus can expand the neighborhood by expanding the range of sound predictions, and so mitigate the problems posed by the Neighborhood Constraint. By drawing on a variety of models, diverse information can be put together to form a more adequate composite prediction.105 Any ideal theory committed to the Social Realizations Condition and cognizant of the Neighborhood Constraint must value predictive diversity.

  The ability of predictive diversity to expand our neighborhood by improving our predictions is of real importance, but it is in one respect critically limited. It remains the case that an excellent predictive model can beat the average of a mediocre collection, but how can we know which models are especially powerful and which are mediocre? “Finding out about” the terrain of justice and the social realizations of other social worlds is not so much about making a prediction about the justice landscape that a perspective can subsequently check as it is about, in a very real sense, constructing that landscape. Recall the idea with which we began this section—“the social theorist invents” social worlds. This is not to say that we are making justice up or constructing the principles of justice (though we could be), but that our only knowledge of a far-off social world is our models of it. We have no independent measurement techniques to determine when a model has gone astray, or to decide what model performs best. We cannot, at the end of the day, compare the results of the model to how the world really is—at least, not until we have actually brought it about. As far as our theory right now is concerned, we are making up the landscape while we are investigating it.

  So how should we “explore” landscapes where our exploration via a model in some sense also constructs them? Once we see that the social worlds are “being made up as we go, we can see, clearly, that there is nothing interesting to be said about how the space should be explored, except to say that it should be explored (as it is made up) in the various ways in which various enquiry teams think best. We should, in other words, devolve decision-making about enquiry to the enquiry teams and let them get on with it.”106 D’Agostino identifies this with a “liberal solution to the problem of enquiry in complex environments. Each team will construct and traverse that region of the space which they find interesting,” using the tools and models they think best.107

  This liberal solution is apt to encourage maximum discovery, as e
ach team (or ideal theorist) seeks to model the possible social worlds it is studying in a way that it deems most fruitful. The likelihood that the most appropriate tools and information ultimately will be used is thus greatly enhanced. The drawback is that each “team” judges for itself whether it has been successful; in the absence of shared standards of evaluation, a genuine insight of one investigator is not apt to be taken up by others. In the absence of shared standards of success we cannot suppose that once an inquirer has modeled a far-flung part of the landscape, and announces the resulting heavily model-dependent results, the rest take the report as veridical. Thus, the liberal approach cannot be well integrated into a single perspective. Those using diverse models tend to go their own way. Given the controversies surrounding which models are most appropriate, rather than seeing modeling diversity as occurring within a perspective on justice, it seems more a diversity among those who understand the problem in different ways. We can discern here the seeds of a dynamic that we will observe in the rest of this, and the coming, chapter: as we seek to take increasing advantage of the fruits of diversity we find that we introduce diversity not simply within a perspective, but between perspectives. The way we see the world tends to influence the models that we think are best to understand it, and so one perspective has some tendency to see some models as more sensible and reliable than do other perspectives. As Page recognizes, a perspective encourages its proponents to employ a specific set of tools for understanding its social worlds and its problems.108 To the extent this is so, maximum predictive diversity is apt to occur when many different perspectives (“crowds”) interact, and bring very different tools to bear on a predictive problem. Overall, “the logic of collective intelligence is that different individuals will apply different ‘theories,’ or more appropriately heuristics, to the guessing task, the aggregate of which results in a highly precise estimate of the variable in question. While each ‘theory’ would only be able to predict part of the variance in the observed outcome, the collection of theories brought together can explain much of the variance and lead to a highly precise result.”109 As we move from a single perspective employing diverse models, to the interaction of diverse individuals with different perspectives, we move from the importance of internally diverse perspectives to the diversity of perspectives.

  The alternative to the liberal, individualistic, and diversity-maximizing approach is what D’Agostino deems the “republican approach,” where inquirers possess common standards of assessment.110 Here we would expect agreement as to what constitutes a good model and how to interpret its results. The “republican” approach seems consistent with diversity within a perspective: when one announces to one’s Σ-perspective colleagues that one has expanded our neighborhood by identifying the working of a new social world at its edges, they are apt to see this as a real advance, as they embrace the assumptions of one’s search. Or, when one group develops a new model, others will grasp its usefulness, and so add it to the basket of models used in the perspective. But while the republican approach enhances communication of results and mutual comprehension, in our case it does so by restricting the lessons that can be drawn from different models of any given social world: only models embraced by the republican community count as informative. On the liberal approach, in a rugged landscape rugged individualist investigators can use innovative techniques to understand the justice of far-off possible social worlds; in contrast, in a republican community that commences with agreement as to what constitutes the correct range of approaches and subjects of inquiry, insightful and highly innovative approaches may be excluded. Thus an exploration of a rugged landscape that is being constructed as we explore it (in the sense that our best models of a social world are the only way to know it—until we arrive at it) confronts the critical trade-off between innovation and communication, a theme that we will explore in some depth in the coming chapter. The greater the diversity of inquiry is embraced, the more apt we are to actually uncover the best insights; but many techniques might not be accepted as reliable by others, and so these insights may not be accepted as veridical. As approaches are constrained to those endorsed by a republican community, they achieve communication of insights, though at a cost of excluding some approaches and their insights.

  4.3 Introducing Explicit Perspectival Diversity

  The costs and benefits of employing different search strategies dominate much of the literature on exploring rugged landscapes.111 However, once an inquirer seeks to evaluate social worlds beyond our neighborhood, our rationale for adopting the neighborhood constraint leads us to take these findings with more than a grain or two of salt. Ex hypothesi, the inquirer is making claims about the justice of social worlds about which we cannot be confident; the resulting model(s) is (are), in an important sense, creating the very world we are evaluating. If we are to search more widely and yet accept the reasoning behind the Neighborhood Constraint, then a theory of the ideal must explore ways to expand its ken within this constraint. Besides improving its predictive models, are there other ways it can do so?

  Recall again the five elements of a perspective: (ES) a set of evaluative standards or principles of justice; (WF) an identification of the relevant features of social worlds; (MP) a mapping relation from the evaluative standards to the features of the social worlds, yielding an overall justice score; (SO) a similarity ordering of the underlying features that provides a meaningful structure to the domain {X} of worlds to be evaluated; and (DM) a distance metric (§II.1). Thus far I have been supposing that everyone shares all these features except, perhaps, the modeling element of the mapping function. Such thorough agreement on the elements of a perspective is, however, an extreme assumption. Analytic results indicate that if we relax the assumption of a thoroughly common perspective, and consider searches among individuals who posses different perspectives, results can be greatly improved.112 Let us, then, introduce a modest degree of diversity among perspectives more formally into the analysis. Suppose that the investigators now all agree on every element of perspective Σ except the metric of distance (DM) between social worlds (§II.1.2).

  To illustrate the significance of this sort of perspectival diversity consider the idea of a distance-contracting metric. A distance-contracting metric is any metric that increases the effective size of our current neighborhood relative to some other metric. Consider for instance the most minimal and straightforward way in which two distance metrics, d1 and d2, might differ from one another, namely if d2 were to be a scalar transformation of d1. In this case if d2 = kd1 where k ∈ (0,1), then d2 would be a distance contracting metric relative to d1. Thus a perspective Σ2—identical to Σ1 except for its distance metric, d2—will view moves between certain social worlds (say, from our current socioeconomic system to property-owning democracy) as moves within our neighborhood, while Σ1, employing the d1 metric, will see these moves as beyond our ken. The result of this sort of difference is likely to be debate about the real size and scope of our current neighborhood. An upshot of this debate sometimes will be the effective expansion of our neighborhood. Should those with distance-contracting metrics like d2 convince Σ1 adherents that Σ2 is a more plausible perspective, we take a small but significant step toward mitigating the Neighborhood Constraint.113

  The point here is subtle and important: just what is within our neighborhood is partially a matter of how far away—how dissimilar—we view certain social worlds. For example, two metrics d1 and d2 might agree on the similarity of the underlying structure of social worlds, such that they both arrange them a–b–c–d–e. On d1, though, a and b are very close social worlds, both considerably distant from c, which is nearer to d and e, while on d2 c is much closer to b than to d (recall figure 2-1). Suppose we are at world a; d1 will not see knowledge of a and its neighbor b as very informative about c, while d2 will; as a result d2 may include c in the neighborhood of a while d1 will not. If Alf can convince Betty, who employs d1, that his d2 is the superior metric, then he will have somewh
at mitigated the Neighborhood Constraint by bringing new social worlds into Betty’s current neighborhood, and thus perhaps a better local optimum. This constitutes a minimal difference in perspectives: while the similarity ordering (SO) of the domain {X} is the same, the distance metrics (DM) are different.

  Once again, it is important to stress that the benefits of perspectival diversity are not merely an upshot of my formal representation of the problem. Consider again Mill’s case for what he called “socialism.” From Mill’s perspective Victorian capitalism fell far below the moral optimum; a form of society centered on worker cooperatives was far better, and perhaps even the ideal. Mill did not simply analyze this ideal, though. Instead he sought to show how a society that might appear very far from the one he inhabited could be achieved via the institutions already in place. Mill insisted that the evolution of new forms of partnerships and corporations that render capitalism more efficient would also allow competitive market processes within capitalism to test the viability of socialist experiments. By connecting the idea of worker cooperatives to a series of intermediate social worlds, he sought to bring socialism into the neighborhood of Victorian capitalism. Rather than a leap into the dark, Mill depicted socialism as a form of industrial organization within the current neighborhood.

  5 THE LIMITS OF LIKE-MINDEDNESS

  As we saw in chapter I, for a theory of ideal justice to orient the search for improvements in justice and, simultaneously, for the identification of the ideal to be critical, two conditions must be met: Social Realizations and Orientation. The first ensures that our ideal theory will help us make the choices between less-than-ideal social worlds that usually compose our option set. When the Social Realizations Condition is met our theory can provide guidance as to whether one social world secures more or less justice than another. This allows us to form judgments of comparative justice as well as sometimes being in the position to recommend reforms that increase justice. As I stressed, however, the Social Realizations Condition alone does not require reference to an ideal; Sen’s “climbing” model meets this condition, and it strenuously abjures any appeal to the best or optimal social realization of justice. The ideal is necessary to orient us not simply when we are concerned with ranking the options in terms of their justice, but when our choice confronts at least two dimensions: how just a social world is, and whether changing the features of the world moves it closer to the features of the ideal.