The Limits of Spontaneous Moral Reconciliation
Preliminary Insights From an Agent-Based Model
Very short summary: In this essay, I discuss Gerald Gaus’s argument about the possibility of moral reconciliation in diverse societies. Like Gaus, I use an agent-based model to explore the conditions under which convergence toward a single social rule happens. The most striking result is that the presence, above some threshold, of a small number of “extremist” agents who are not willing to reconcile impedes convergence. This suggests that moral reconciliation is vulnerable even to a low number of persons who don’t share the “liberal ethos.” Note of caution: results are very preliminary!
Regular readers know that I’m a big fan of the work of the late Gerald Gaus, who was a political philosopher at the University of Arizona. Though I barely had the chance to interact with him (besides a couple of emails a few months before his death), I’ve read almost all his academic output since the 1990s. Gaus’s post-Rawlsian liberalism pioneered a new perspective on the diversity and pluralism that characterize contemporary liberal societies. At the methodological level, he showed how philosophical reasoning can be advantageously combined with the tools, methods, and theories of social sciences. In this sense, Gaus was the ideal type of the PPE scholar, and this essay is a very modest contribution to an intellectual tradition he has largely contributed to create.
Gaus’s latest writings, especially his book The Open Society and its Complexities, focus on what can be called the problem of moral reconciliation.[1] As Rawls reckoned,[2] liberal societies endogenously produce diversity and pluralism. Their open nature means that they not only refrain from slowing, but even encourage all forms of experimentation and innovation, in every domain (economic, social, political, cultural). This openness manifests itself in liberal institutions, both economic and political. It also transpires in the more difficult to characterize “liberal ethos” within which individuals are brought up, and that they continuously renew in their daily practices. Part of this ethos is constituted by norms of tolerance. Such norms mean that, at the micro level of daily interactions, if I see someone in public transport dressed in a way that I regard as “improper,” I will nonetheless refrain from making any comment, even less physically interfering with the person. Norms of tolerance ask us to let people live their lives as they see fit, even if we disapprove of their ways of life. The liberal ethos also operates in more formal domains of social life, including (and especially) in politics, where we are asked to respect people we disagree with —a demand that seems to be increasingly difficult to meet in our populist times.
The openness of the liberal society creates opportunities that are out of the reach of societies that, one way or another, impose a greater social control on behavior and practices. Authoritarian and totalitarian states enforce this control through various formal and (especially for totalitarian states) informal means through which they monitor people’s behavior and sanction attempts to deviate from the imposed social rules. Though lacking a powerful administrative control system, traditional societies also tightly monitor individuals’ ways of life through the more diluted pressure of customs and traditions. In both cases, experimentation and innovation are virtually prohibited, or at least considerably limited.[3] The result is that “closed” societies maintain, over the long run, a higher degree of social homogeneity, at the cost of missing opportunities for new sources of wealth, well-being, and ultimately, human flourishing.
There are also downsides to openness. Liberal authors like Karl Popper or Friedrich Hayek were not shy in acknowledging that the liberal society is inimical to “tribal” ways of life and, more generally, may clash with the evolved human psychology.[4] Also, there is no guarantee that the uncontrolled explosion of diversity and plurality will not, over the long run, destroy the moral and social foundations on which liberal institutions are thriving —a theme explored many times, from conservative-minded liberals to contemporary post-liberals. This latter risk is related to, but distinct from, the possibility that, as diversity is increasing, the scope of social rules over which individuals agree progressively shrinks. Rawls’s bet was that the kind of diversity encouraged by liberal institutions would not undermine social stability, building on a consensus over principles of justice. However, post-Rawlsians like Gaus rightly point out that the Rawlsian faith in the possibility of such a consensus doesn’t rest on solid grounds.
Gaus was the precursor of what is now called the convergence approach in public reason liberalism.[5] Contrary to Rawls and his followers, convergence theorists don’t assume that a consensus over justice principles should be expected. However, they contend that if public justification of rules is possible, it will be thanks to a process of convergence where diverse individuals agree on a set of rules, even for different reasons. Let’s call that the convergence conjecture. The bet is therefore of a different nature. People are likely to entertain very different reasons for accepting social rules and the authority that comes with them. This diversity makes any consensus unlikely, but is not an obstacle to public justification, provided that individuals agree it’s better to have a specific set of rules, rather than no rule at all.
What I’ve called the problem of moral reconciliation is essentially about assessing the truth-value of the convergence conjecture. There are several complementary ways of doing this, and in this latest work, Gaus explores a formal approach known as agent-based modeling. An agent-based model (ABM) studies the aggregate effects of the interactions of autonomous agents whose behavior depends on rules that are directly specified in a code. ABMs are computational rather than mathematical in the sense that their formal structure consists of commands expressed in a programming language, rather than explicit mathematical operations. ABMs have been used for a while in “complexity sciences” and have achieved some popularity among social scientists —though economists remain notoriously skeptical about the relevance of the whole modeling approach.[6]
A chapter of The Opening Society and its Complexities discusses an ABM that explores the convergence conjecture by assessing the disposition of a diverse population to converge toward the adoption of a single rule. I will leave most of the details aside, but here is a short list of features that may help to understand what’s happening in the model.[7] Agents can choose between two rules on which to coordinate. They each have an intrinsic preference regarding which rule they would like to adopt, which is captured by a cardinal utility function. They also have a preference for conformity, which consists in the fact that, everything else equal, they prefer to conform to the rule that is dominant in the population. Agents’ overall utility for following a given rule is the product of their intrinsic utility times a weight that depends on the proportion of the population following this rule.[8]
Diversity is introduced at two levels. First, agents’ intrinsic preferences over rules are heterogeneous. Second, agents are of different “reconciliation types” regarding how they weigh the fact of following the same rule as others in the population. More specifically, they can be “quasi-Kantian,” moderately or highly conditional cooperators, or “linear” agents. Quasi-Kantian agents highly value reconciliation (they give a high weight to the rule that is majoritarian), while conditional cooperators give significant weights only to rules that are adopted by a large majority. The figure below represents the different reconciliation types in Gaus’s model:[9]
Now, the question is, under which conditions does the population converge toward a single rule? Gaus runs many iterations of the model with diverse distributions of reconciliation types. The general lesson is the confirmation of the convergence conjecture, including when agents are sorted into groups, some of them “parochial” (i.e., their population of reference stops at their group and they disregard what others are doing). This is an interesting and not necessarily intuitive result that tends to support the claim that diversity is not an obstacle to moral reconciliation.
A well-known issue with ABMs is that they tend to function as “black boxes” whose output may be difficult to explain and interpret. The evidence they produce is always conditional on the range of modeling details, many of them buried deep in the code. Therefore, the rule is to run many iterations of the same model with all possible parameter configurations, but also to explore similar but different models that make slightly different structural assumptions. An obvious concern, for instance, with Gaus’s model is that it only runs simulations with two rules, while we may rightly wonder if the convergence result remains with three or more rules available. We may also wish to test different reconciliation types configurations, and especially to explore what happens when there are a few “extremist” agents that stick to a rule, whatever others are doing.
I’ve started to explore such possibilities using NetLogo. My knowledge of NetLogo is minimal and dates back more than a decade, but in this AI age, it is now easier to do basic coding without much programming skills and knowledge. So, working with a bunch of AIs, I’ve developed in a few hours a simple ABM inspired by Gaus’s, though with different structural assumptions. I also didn’t have the time to run many simulations, so the results are definitely to be taken with caution, but I think they are still worth considering. For those who are interested, I provide more technical details at the end of the essay. The NetLogo code is available on demand.
The model’s general structure is similar to Gaus’s. Take a population of agents (here 200, against 101 in Gaus’s model) who have to choose which rule to follow. However, here, three rules are available. Like Gaus, I assume that agents’ overall utility is the product of the intrinsic utility of the rule adopted (normalized between 0 and 10) times a weight (normalized between 0 and 1) that reflects the importance that an agent gives to the fact of following the same rule as other agents. I identified three reconciliation types: kantians (similar to Gaus’s highly conditional cooperators), humeans (similar to Gaus’s conditional cooperators and quasi-kantians),[10] and “random” (similar to Gaus’s linear agents). I add a fourth type that I call the “extremists.” These agents have the peculiarity of ascribing an intrinsic utility of 0 to two of the three available rules, meaning that they will always stick to the same rule, whatever others are doing (hence, reconciliation weights are irrelevant for them). As in Gaus’s model, at each period, each agent settles for the rule that maximizes their overall utility, given the aggregate behavior of others in the previous period.
The rationale for introducing extremists in the population is to check how much public justification by convergence (and thus, moral reconciliation) is robust in the face of the presence of a few individuals who are unwilling to make any compromise in a context of radical pluralism, if not outright polarization.[11] Intuitively, the presence of extremists makes convergence less likely. Obviously, a population of extremists will never converge. However, it is interesting to check which kind of social dynamics emerges when we vary the proportion of extremists at the margin. The hypothesis is that above some threshold, even a small proportion of extremists can compromise convergence.
So let’s explore what happens with different distributions of reconciliation types. First, to check if the model aligns with Gaus’s main result, we can look at what happens when no extremists are in the population with a uniform distribution of the three other types. Here is one representative outcome:
As you can see, convergence (here, on rule 3) is fairly quick. Interestingly, with a population of two-thirds Humean and one-third Kantian, the model displays highly variable dynamics, some of them converging, others not. Here is an example:
This suggests that random-type agents play an important role in the convergence process, but I’ve not explored this point further.[12] In my model, as few as 15% of “random” agents with no extremists are enough to make convergence almost systematic when humeans are majoritarian. When kantians dominate, convergence is less systematic.
However, the main point is to see what happens when we introduce extremists. At a fairly low rate, this introduction doesn’t seem to affect the overall pattern. Here is a typical result with 5% of extremists and a balanced distribution of other types:
As we increase the share of extremists, the results start to be more random. But until 10% of extremists (with an otherwise uniform distribution of types), a typical pattern is as follows:
However, the dynamic almost never converges past the 10% threshold:
Therefore, there seems to be some nonlinearity built into the model, which confirms the initial intuition. The introduction of extremists beyond some threshold impedes the convergence, not only because these extremists will never give up their rule (by definition), but also because this influences other agents who value the fact of following the dominant rule in the population. It seems that reducing the number of kantian agents increases the threshold above which extremists block convergence. For instance, with only 5% of kantian agents, convergence (ignoring the extremists who will never converge) is almost systematic with 40% humean agents and not more than 23% of extremists:
However, substituting some extremists for some random agents considerably increases the rate of non-convergence (approximately one-third of the time).
All this is arguably sketchy and calls for a more systematic exploration. The kantian agents seem to have an important influence on the sensibility of the dynamics to the presence of extremists. Presumably, that’s because I’ve modeled them as agents who have departed from their most preferred rule only if an alternative one is likely to be “universalized.” That makes them extremists of some sort, though not as much as the genuine extremists.
In any case, I think the model supports the intuition that spontaneous reconciliation is unlikely in a society where a small but significant fraction of the population rejects what I’ve called the liberal ethos. Even if there are few, these persons contribute to creating a context where even liberal-minded persons are less ready to make concessions and more likely to stand firm on their moral convictions. The ensuing social dynamic is conducive to the kind of Schmittian power politics that I’ve been discussing recently (see especially here and here) and its self-fulfilling character. Gaus was well aware of this risk, as witnessed by what he wrote in his very last paper:[13]
“For much of my career I have developed an account of how people who deeply disagree about the basis of normativity and have serious disagreements about what is right and wrong can nevertheless converge on common social-moral rules for cooperative living. The core idea is ‘convergent normativity’: while we disagree on many of the grand issues of morality we can, in the interests of achieving a cooperative order based on relations of mutual moral accountability, reconcile on common rules that each of us, for her own reasons, endorses. This tale draws on empirical literature concerning moral psychology, norms, social cooperation, punishment and practice of accountability. So, I trust it is not a ‘just so’ story. But its lesson, that each can simultaneously affirm her own moral perspective while reconciling with others on common rules, abstracts from numerous issues. It assumes, first and foremost, that most people do not simply dismiss the normative perspectives of others as befuddled crooked thinking, but acknowledge them as intelligible ways to understand social morality. In the current environment, many people apparently prefer the joys of anger and aggressive self-righteousness to reconciliation with others. Why reconcile when the enemy can be crushed? I am pessimistic about the future of social orders with this ethos, but if that is the ethos the itch my story seeks to relieve—how can a truly diverse society share a moral framework?—is not felt. These views are attracted to stories about life among truly straight-thinking people, and damning those crooked thinkers who stand in the way of true justice.”
Unfortunately, since Gaus’s death, political events seem to only confirm that “many people apparently prefer the joys of anger and aggressive self-righteousness to reconciliation with others.” What remains of moral reconciliation and of the idea of public reason in a world where more and more individuals see politics as a war?
Information about the ABM discussed in this essay
In what follows, I give more specific and formal elements regarding the ABM that I’ve discussed. The NetLogo code is available on request (if you want to work on it, I’m open to a collaboration).
· The model has a population of 200 agents who, at each period, choose among 3 rules.
· At each period t, each agent i chooses the rule Rx that maximizes their utility function U_i(Rx)=u_i(Rx).w(p(Rx)) where u_i is the intrinsic utility of Rx for i and w(p(Rx)) is a function assigning a weight of following Rx when a proportion p(Rx) follows Rx. The u_i functions are normalized between 0 and 10, and the weights w are normalized between 0 and 1.
· Intrinsic utilities are assigned randomly. A small fraction of agents (less than 5%) assign an intrinsic utility of 0 to one of the three rules.
· The weighting functions w can be of three different types, depending on the agent’s reconciliation type.
o Kantian agents: when p < 0.7, w = 0.03, and then w increases linearly above, up to w = 0.275.
o Humean agents: if p < 0.5, w = 0.05+(0.01p); if 0.5 <= p < 0.9, w = 0.1 + ((p - 0.5)/0.4)*0.5; if p>= 0.9, w = 0.9 + (p - 0.9).
o Random agents: w = p.
· Extremist agents have u_i(Rx) = 0 for two rules and u_i(Ry) > 0 for the remaining rule y.
[1] Gerald Gaus, The Open Society and Its Complexities (Oxford University Press, 2021).
[2] John Rawls, Political Liberalism (New York: Columbia University Press, 1993).
[3] An intuition that would need to be explored more thoroughly is that closed societies mostly rely on what Gaus and Nichols call “prohibitive rules,” while open societies function with “permissibility rules.” Under the latter system of rules, behavior and practices are assumed to be permitted, unless they are explicitly prohibited by social rules. Prohibitive rules work the other way round. See Gerald Gaus and Shaun Nichols, “Moral Learning in the Open Society: The Theory and Practice of Natural Liberty,” Social Philosophy and Policy 34, no. 1 (ed 2017): 79–101.
[4] Karl Popper, The Open Society and Its Enemies, 7th edition (Routledge, 1944 [2012]). Friedrich A. Hayek, Law, Legislation and Liberty, Volume 1: Rules and Order (The University of Chicago Press, 1973 [2011]).
[5] Kevin Vallier and Ryan Muldoon, “In Public Reason, Diversity Trumps Coherence,” Journal of Political Philosophy 29, no. 2 (2021): 211–30.
[6] See especially the research conducted at the Santa Fe Institute. Though I’ve read it a long time ago, a good place to start to learn about ABM, especially when using NetLogo, is Nigel Gilbert, Agent-Based Models (Los Angeles: SAGE Publications, Inc, 2009).
[7] The same model is also discussed in Gerald Gaus, “Self-Organizing Moral Systems: Beyond Social Contract Theory,” Politics, Philosophy & Economics 17, no. 2 (May 1, 2018): 119–47.
[8] Formally, the utility of agent i following rule Rx is U_i(Rx)=u_i(Rx).w(p(Rx)) where u_i is the intrinsic utility of Rx for i and w(p(Rx)) is the weight of following Rx when a proportion p(Rx) follows Rx. In Gaus’s model, u_i is normalized between 0 and 10 and the weight between 0 and 1.
[9] See Gaus, “Self-Organizing Moral Systems.”
[10] The choice of the labels is probably not the best. Contrary to Gaus, I interpret kantians as agents who value cooperation and are nonetheless predisposed to reject rules that they view as weakly justified, even if followed by others.
[11] Therefore, in the context of Gaus’s theory of public reason, I am assuming that the “socially eligible set” is empty, i.e., there is no available social rule that everybody prefers to the absence of rule. Gaus argues that such a scenario is unlikely, but this is precisely how we can interpret the return of “power politics” that I’ve discussed in detail in recent essays here.
[12] This is confirmed by the fact that convergence rarely occurs in the inverse distribution (2/3 kantians, 1/3 humeans).
[13] Quoted in Kevin Vallier, “The Social Philosophy of Gerald Gaus: Moral Relations Amid Control, Contestation, and Complexity,” Journal of the American Philosophical Association 9, no. 3 (September 2023): 510–32.