Welcome to the Brains Blog’s Symposium series on the Cognitive Science of Philosophy. The aim of the series is to examine the use of diverse methods to generate philosophical insight. Each symposium is comprised of two parts. In the target post, a practitioner describes their use of the method under discussion and explains why they find it philosophically fruitful. A commentator then responds to the target post and discusses the strengths and limitations of the method.

In this symposium, Jingyi Wu (UC Irvine) discusses how she uses network models to explore what happens when the testimony of people who are socially Marginalized is devalued. Commentator Kevin Zollman (Carnegie Mellon) discusses the challenge of figuring out when a network Model applies to a particular social system.

* * * * * * * * * * * *

Modeling Injustice in Epistemic Networks

Jingyi Wu

* * * * * * * * * * *

Here is a scenario that is perhaps all-too-familiar to some of you. Imagine that you are a female scientist in a meeting with potential collaborators, brainstorming about possible research projects. People in the room start going through everyone’s skillsets and common interests, and suddenly you see the dots connecting and pitch an idea. But, as it happens, no one seems to notice that you just had an idea. The conversation flows forward as if it never happened, until 10 minutes later, a male scientist finally sees the dots connecting and says the same idea as yours. Everyone applauds. “What a great idea, Matt. Let’s do that!”

This is a hypothetical scenario, of course, but I’m sure many of you recognize the pattern (it even has a name—hepeating). It is empirically well-documented that testimony given by socially marginalized people is often devalued by people in Dominant groups. Law professor Meera Deo conducted a meticulous study about the role of race and gender in the academic context, and found that most women in the study sample, “regardless of racial/ ethnic background, have endured silencing, harassment, mansplaining, hepeating, and gender bias” (Deo 2019, 47). One study subject counted over ten times that she was hepeated in faculty meetings. Similarly, studies have also found that people of color’s testimony in the court system is taken to carry less weight (Carlin 2016).

Philosophers have introduced theories to better understand this phenomenon. For instance, Kristie Dotson (2011) coins the term epistemic silencing, which describes situations in which an audience, often from a dominant social background, fails to identify a speaker, often from a marginalized background, as a knower. Devaluing marginalized people’s testimony constitutes a paradigmatic case of epistemic injustice , because socially marginalized people are given less credit than they deserve as knowers (Fricker 2007).

Importantly, various authors point out that this devaluation of testimony happens in an asymmetric fashion. In order to survive in a hostile world, socially marginalized people simply cannot afford to devalue dominant groups’ testimony. Philosopher of race Charles Mills quotes James Baldwin’s brutally honest line to demonstrate this point, “I have spent most of my life, after all, watching white people and outwitting them, so that I might survive” (Baldwin, 1992, 217, qtd. Mills 2007, 18).

While the phenomenon of asymmetric testimonial uptake and and its (un)ethical nature is well-studied by philosophers and social scientists, its epistemic consequences are a lot less clear. We might wonder: how does asymmetric testimonial uptake impact knowledge production in a group? In other words, if marginalized people’s testimony is routinely devalued, does it make them better or worse at learning something? And how does this impact the creation of knowledge in the entire community?

These questions are difficult to answer. For one thing, while we have access to subjective reports and behavioral observations of testimonial uptake, it is much harder to measure who has better knowledge. Moreover, even if we had such a measure, reality is complex with multiple causal pathways. It would be difficult to figure out if it is indeed the asymmetric testimonial uptake that is actually responsible for the observed consequences.

In my recent paper (Wu 2022a), I study the epistemic consequences of asymmetric testimonial uptake using computer simulations of network models.^[1] In the models I build, a group of agents are connected to each other via a network (see Figure below). Every agent occupies a node in the network, and edges represent social relations. Agents are then tasked with learning to solve a decision problem. In typical versions of the model, agents share evidence they discover with their neighbors (who are connected to them via an edge), and therefore agents may learn from discoveries they personally have not explored.

We can begin to see why network models are particularly suitable to study the questions I posed. A piece of evidence in the model that is shared and then properly updated on can be thought of as being given full credit. A piece of evidence that is shared but subsequently ignored or devalued by the receiver, then, can be thought of as being treated unjustly. Furthermore, the models I use are minimal by design. Having few components in the models means that I can isolate individual causal factors and investigate their consequences. This allows me to change one aspect of the model—adding asymmetric testimonial uptake—and observe how it changes the learning dynamics.

Using these models, I find some (perhaps surprising) results. When marginalized agents’ testimony is ignored or devalued by the dominant group but not vice versa,  the marginalized group ends up developing better beliefs about the learning problem as compared to the dominant group.^[2] Dominant groups miss insights from the marginalized community, and consequently explore unpromising theories for longer. The marginalized group, learning from a wide range of perspectives, benefits from this epistemic diversity and eventually learns better.

More counterintuitively, when their testimony is fully ignored by the dominant group, the marginalized group develops even more accurate beliefs than a community without testimonial injustice—where everyone fully updates on their neighbors’ evidence. Why is this the case? In the presence of initially misleading data, a community without testimonial injustice may prematurely settle onto a worse belief too quickly without ever exploring other options. The dominant group in my model, however, functions as an isolated community, so it takes longer to learn. Different options thus circulate in the community for longer, and the marginalized group, having access to the dominant group’s data, may eventually self-correct too.

These results together support a key claim of standpoint epistemology called the inversion thesis, which states that socially marginalized people can sometimes develop better beliefs. Ever since the inversion thesis was proposed in the 1980s, its interpretation and justifications have been highly contested. Why do socially marginalized groups have better beliefs? The thesis seems ever more mysterious considering that in many cases, the very manifestation of marginalization is the exclusion of certain groups of people from epistemic communities.

My models provide a possible mechanism through which standpoint advantages for marginalized people can emerge, by casting it as a consequence of testimonial injustice or elite-group ignorance.^[3] One might regard the claim that marginalized groups’ testimony is ignored or devalued as far less controversial than the inversion thesis. Insofar as this is right, my models also have the virtue of supporting a controversial thesis by showing that it follows from something more widely accepted. Moreover, we can use the models to formulate conditions under which standpoint advantages can develop. Marginalized agents in my models still have access to dominant agents’ data, but when they are completely excluded from participating in certain epistemic communities, they may not develop epistemic advantages at all.

Ultimately, the findings suggest that there are epistemic benefits from having a diversity of approaches in social learning. The marginalized agents entertain a wider range of perspectives and approaches (ironically as a result of dominant agents’ narrowmindedness), and that allows them to explore sufficiently before developing stable beliefs. This is related to the concept of transient diversity in science (Zollman 2010). A community learns well when there is a transient period when a diversity of options are being tested. Otherwise, it risks locking in on a worse theory too early.

Interestingly, the models I develop can be reinterpreted to capture an entirely different type of injustice in social relations. Instead of an agent X sharing a piece of evidence with agent Y that is subsequently being ignored, we can interpret this asymmetry as agent X refusing to share their evidence at all. This interpretation captures a structural asymmetry between industrial and academic scientists: while industrial scientists often withhold evidence from the outside, academic scientists often adhere to what is known as the communist norm and share their work widely. 

This phenomenon is related to the concept of informational injustice in the ethics of computing literature. Here, the industrial scientists commit informational injustice by depriving the wider society of their proprietary data, while free riding off of academic scientists’ findings. Even though the contexts are vastly different, the industrial scientists in this interpretation and the marginalized social group in the previous interpretation interestingly occupy exactly the same structural position.

In a working paper (Wu 2022b), I explore this interpretation of the asymmetric relation. I argue that scientists can have epistemic incentives to unilaterally withhold their evidence from the outside, even without taking into consideration financial or instrumental incentives. The withholding scientists develop better beliefs more frequently and faster than the rest of the community. I further argue that the sharing dynamics constitutes an epistemic version of the prisoner’s dilemma: scientists learn worst when others withhold evidence, best when they unilaterally withhold, and in between when everyone shares.

Now you might wonder, reality is complex, but the models I present are simple and idealized. How do we know if the results do not just reflect idiosyncrasies of the models? How do minimal models gain traction in the real world? There is not an easy answer to these questions, but two thoughts might help. First, in my work I have tested the same structural asymmetry in different modeling paradigms of social learning, and obtained largely similar results. This should give us confidence that the findings are robust across various modeling assumptions. Second, idealized models are not the end of the inquiry, but rather, they point to concrete places where empirical testing may be efficacious. We can start by investigating whether the core modeling assumptions approximately hold in epistemic communities we are interested in.

Finally, since previous work of network modeling in philosophy all involve reciprocal social relations, introducing asymmetry to how agents interact with each other is a new and exciting terrain. Besides testimonial and informational injustice, we can use asymmetric network relations to model other kinds of injustice in our society. For instance, in my current projects (with fantastic co-authors like Liam Kofi Bright, Sina Fazelpour, and Hannah Rubin), I use network models to think about situations where (a) mainstream academics largely cite work from their own circle, but marginalized academics have to engage with mainstream work to survive; and (b) marginalized social groups are pressured to conform or assimilate to dominant groups, but not vice versa. Often simple models can offer surprising insights into our target phenomena, and network models, with asymmetric social relations, can be fruitfully applied to provide new understandings of race, gender, and injustice.

[1] This modeling paradigm was developed by economists Bala and Goyal (1998) and introduced to philosophy of science by Zollman (2007, 2010).

[2] In the simplest version of the model, there are two groups: one marginalized and one dominant. The only feature differentiating the two groups is that the dominant group ignores or devalues evidence coming from the marginalized group, but not vice versa.

[3] Note that while this is a sufficient condition, it is not a necessary one. Notably, the structural relation between testimonial injustice and standpoint advantage that I am concerned with is not the mechanism many standpoint theorists have in mind (e.g. Hartsock 1983).

* * * * * * * * * * * *

Commentary

Kevin Zollman

* * * * * * * * * * *

In this blog post, Jingyi Wu provides a wonderfully lucid introduction to results from her exceptional paper about modeling epistemic networks of dominant and marginalized communities. It is an exciting application of an existing modeling framework to a novel problem that exhibits Wu’s characteristic creativity and clarity. It is a great paper, and I would encourage everyone to study it.

The paper uncovers a troubling example of “The Independence Thesis”—that individually sub-optimal behavior may in fact serve the group interest. Many of the existing examples are ones where a person does something that hampers their ability to learn in order to benefit their group’s ability to do so. Wu’s example adds an extra layer: she shows how a group might be epistemically benefited by a behavior which is epistemically bad individually and also ethically quite troublesome. In this case, the marginalized group becomes better epistemically by being the victim of testimonial injustice.

Of course, neither Wu nor I would suggest that the epistemic benefits justify the injustice. As Wu notes in her paper, there are other ways to achieve the same benefit which do not require marginalization. Should we be designing ideal epistemic communities, we ought to seek out one of those instead.

As Wu argues, however, it gives novel evidence for a claim that has often been difficult to underwrite: that marginalized groups might be epistemically superior in certain respects (the “inversion thesis”). Wu’s paper provides a clear proof-of-possibility of this claim. Like any mathematical or simulation model, it makes idealizations that are better or worse approximations of different parts of real social systems. In order to fully understand the applicability of the results to a specific social system, one might want to dig into these. Wu’s proposed mechanism applies to some, but not all, cases of injustice.

Wu employs the bandit problem model of learning, where an individual must balance the desire to explore the options with a desire to make use of their current information in pursuing the option that currently seems best. The risk is that a group may “lock in” on an option that seems superior but is not. If they incorrectly think that a good option is bad, they may never discover their error.

Critical to these models is that experimentation is a public good. You experimenting with apparently inferior options benefits me, since you are giving me information, but it only costs you (since you are the one who might choose a poor option). This undergirds Wu’s results. When the dominant group ignores the marginalized group, they harm themselves by giving up on that cost-free information.

Critical to Wu’s model is that this is a one-sided break in the epistemic network. Marginalized groups can receive evidence from the dominant group. So they have access to a public good that the dominant group does not. But since the dominant group ignores the marginalized, they forgo the public good.

I believe that there are many cases of marginalization that fit Wu’s picture. Indeed, she gives many illustrative examples. But not all cases of marginalization fit this picture. For example, if the marginalized group is systematically excluded from educational opportunities, then the marginalized group may not have access to the evidence generated by the dominant group. In that case, Wu’s model does not apply.

Of course, this doesn’t show that the inversion thesis is false in these contexts. My observation merely removes Wu’s novel justification for it. The inversion thesis is sufficiently general that it might come about through many mechanisms, including this quite novel one that Wu provides.

Nor should I be read to suggest Wu’s model is wrong. I agree with her that some cases of marginalization fit this picture well. Rather, like all mathematical and simulation models, different social systems fit the picture better and worse.

Overall, I think Wu’s paper is exciting both for the results it reports and because it shows how flexible and interesting these models can be. I look forward to more work of this high caliber from her and collaborators on this topic!

* * * * * * * * * * *

* * * * * * * * * *