 Hello, everyone, and welcome to this broadcast. Firstly, I'd just like to acknowledge the traditional owners of the land on which we meet, the Non-Aul and Nambri peoples, and pay my respects to their elders past, present, and emerging. We're delighted to have you all join us this morning, and this is our second HMI Data, AI, and Society seminar in the series. And we're absolutely delighted, of course, to welcome the wonderful Lily Hu from Harvard University and hear more about her work. She's going to discuss what is race in algorithmic discrimination on the basis of race. And just before we get into that, I'll just do a little few little announcements. Lily will speak for 30 minutes, in which time please feel free to add any of your questions to the chat here on the Zoom channel. And then we'll have a 30 minute Q&A session, and we'll have a hard stop at 10am. And after that, we can continue our conversation in Slack. So without further ado, Lily, please, if you would like to start. Thank you. Thank you so much for that introduction, Shell. First, I want to say thank you to everybody who's at HMI and making the Data, AI, and Society seminar series possible. They have been so gracious, allowing me to join you all and give this talk in spite of a rather foolish and extremely foolish error last week, failing to do my time zone calculation correctly. Fortunately, a brave soul jumped right in. So Claire, I believe, thank you for which I'm so grateful. So I hope today's talk will make up a bit for last week and maybe even exceed what I would have done last week because I'm super prepared now. Okay, well, thanks again. So this is what is raised in algorithmic discrimination on the basis of race. I'm going to start off with this court case. In August 2018, the United States Department of Housing and Urban Development, that's HUD, filed a housing discrimination complaint against Facebook, saying that the company's algorithm based advertising system quotes minds extensive user data and classifies its users based on protected characteristics, and thereby quote unlawfully discriminates by enabling advertisers to restrict which Facebook users receive housing related ads based on race, color, religion, sex, familial status, national origin and disability. Facebook responding to this complaint pointed out that its machine learning system does not even make use of a race feature. So it doesn't even have the ability to target individuals based on race. But this line was unconvincing. In March of the following year, HUD put forth an actual charge of discrimination that earlier was a complaint and I'm going to read the text in its entirety here but it's on the slide. It's an excerpt respondent that is Facebook combines the data it has about user attributes and behavior on its platform with data it obtains about user behavior on other websites and the non digital world. Respondent then uses machine learning and other prediction techniques to classify and group users so as to project each user's likely response to a given ad. In doing so, respondent inevitably recreates groupings defined by the protected class by grouping users who like similar pages unrelated to housing and presuming a shared interest or disinterest in housing related advertisements. Respondence mechanism functions just like an advertiser who intentionally targets or excludes users based on their protected class. So I think there are two questions that I want to focus on here. First, we might ask, well, why does the algorithm reconstitute these protected class groups. And I think some simple laws are sociological observation and as well as observations about how machine learning works makes that quite clear right. There are sociological facts about how certain groups of people, what kinds of pages they tend to like and also facts about clicking on certain and apparently as long as the behavioral tendencies are true in the world. The machine learning system will learn them. So I think the first question. I can kind of set aside. I think the second question though is rather interesting, which is, why does count as discrimination. So if we suppose that HUD reconstruction of the events is is in fact what is happening why would this kind of behavior constitute discrimination. This exchange between Facebook and HUD drills down on a critical point of contention in debates about algorithmic discrimination, which is, what does it take for a database system to act on or on the basis of race. And here HUD proffers a negative proposition. They say look an algorithm can discriminate on the basis of race, even when it does not have access to a race feature. So I want to ask that my kind of talk today will be asking, you know, could such a bold claim be true in what theory of discrimination what theory of race would make it so. So, I think database predictive tools like the ones at the center of the HUD versus Facebook case clarifying underappreciated problem for discrimination theory, which I'm going to call the puzzle of algorithmic discrimination. And I'm going to refer keep referring back to this as the puzzle. So this is really, we got to keep this one in our minds here. And the question is simply this when our decisions made on the basis of features that are correlated with race decisions made on the basis of race. And I think this question is not just relevant to current debates about algorithms, I think a version of it has dog philosophers legal scholars for decades under the descriptor proxy discrimination, in which an attribute that is a proxy for race or sex or so on is used in decision making. And I think predictive machine learning simply kind of cast this old problem in new light, but also allows us I think to drill down on like, you know what exactly is going on here that's making this so puzzling. So, here is the roadmap for today's talk. So I'm going to set up the puzzle, I'm going to pursue a solution to the puzzle, not by elaborating a distinctive account of discrimination, but rather a distinctive a social constructivist account of race. And I'm going to especially talk about a subset of social constructivism which I'm going to call thick constructivism. I'm going to show how that will offer a solution to the puzzle. And then I'm going to talk about some implications for discrimination theory more broadly. So let's set up the puzzle setting up the puzzle. Okay. The hub versus Facebook case, I think well illustrates a central dispute in cases of potential algorithmic discrimination with each party, taking a distinct interpretation of this notoriously vague because of race clause that we tend to see an analysis of discrimination. So consider the following I think rather uncontroversial, but non moralized analysis of discrimination, which I'm going to call our D. So X discriminates on the basis of race if X treats why differently from how X treats or would treat some Z because why supposedly has racial status R and Z supposedly has a different racial status or crime. So, because in the literature tends to be read in two ways. On one reading, we're asked because refers to X is kind of taking race as a reason for the differential treatment, right. So the reasons interpretation of because asked us to look into the considerations X took and acting the way that she did. Another reading because refers to X is treatment of why is just being constantly affected by race in just a non reasons way. So we're going to consider how race might have affected or played a causal role and how X treated why, aside from being a part of X is reasons perhaps even in a way unbeknownst to her. So I want to first set aside some easy cases of algorithmic discrimination. So I think that everybody would, I hope that everyone would gather that center cover intentions to recreate racial groupings and then assign them differential ad outcomes on the basis of those groupings would be discrimination on the basis of race and wrongful discrimination on basic race at that. So for the sake of argument I want to set aside those cases where there's like, you know, a band of engineers in the background doing something. I also want to set aside cases where the algorithm explicitly has a race feature because remember this is the kind of key point of Facebook's defense they didn't even have a race feature right. And this is a bit of a simplification which I'll continue I'll talk more about later on, but I don't want to have to develop a full account of what it takes for a machine to take something as a reason. So I'm just going to say look now, you know, take. In fact, that that'll be the only way that it can take races or reason for now I'll make that stipulation and then that simplification and then I'll go go on to revise it. And that's because I just want to make I want to set aside kind of more complicated reasons or sorry more complicated accounts of what is take something for a reason. Okay, so the really the tough cases are cases when the race features not in the machine learning algorithm or machine in the database and there's also not these covert intentions right. I think for these cases, the reasons versus causes framing exposes a rather deep fault line. Right, you go you go for the reasons view and it would seem that algorithms never. I mean so long as they don't have the race feature, but if you go with the cause interpretation it looks like if, at least if the machine learning system is working properly in many cases, they almost necessarily discriminate on the base of race. Right, like race just cause effects so many aspects of our lives. I want to set aside the reasons account because I want to talk a little bit about the literature the kind of more technical literature on the causes account which there's been quite a bit on so scholarship on algorithmic discrimination has overwhelmingly advanced a version of this causal interpretation. And I want to say a little bit about why I find these accounts to be lacking, but I'm happy to discuss this more in question Q&A. So I can't fully address the full breadth of the scholarship that I wanted to kind of acknowledge it for especially computer scientists who are maybe quite familiar with causal fairness or causal counterfactual accounts of fairness. So there's this there's a lot of literature on what's called like path specific accounts of fairness where really what the algorithm with the kind of framework is to think about this causal diagram of how race affects all the other attributes in your data system. And then to designate which causal pathways are bad, and then to say my algorithm computation is not going to take into account those pathways. I think the problem with this approach is that the grounds that you need to identify those suspect pathways they can't come from solely within the causal framework. And I think most work in the area concedes as much, but and they submit the task of that designating the the unfair pathways should be should be just directly drawing on normative considerations. But I think this suggestion is simply unresponsive to the puzzle of algorithmic discrimination as I framed it, because that calls for a descriptive account of what it is to act on race, right, setting aside any particular moral theory about why acting in that way would be wrongful. And I think skipping to a moralized conception of discrimination actually overlooks what makes algorithms so interesting in the first place because oftentimes we're asking like, did the algorithm even act on race right like that's kind of the thing. That's at the heart of this HUD versus Facebook case. So I think, so I think just kind of going the moralized direction kind of misses it. Furthermore, though, I think any such method that does aim to provide a descriptive account is only able to pick out those pathways that count as acting on the base of race in light of a fundamental theory of race. Right. But I think causal causality based approaches have been largely silent on these matters of metaphysics. I think causal counterfactual approaches. And if you're not familiar with these, I'm almost done here, causal counterfactual approaches also I think suffer from similar silences. You know, in virtue of what facts do we decide what the relevant counterfactuals are, or what are the counterparts of certain features in our, you know, non discriminatory possible worlds. I think there too, a theory of race is required to even get the method off the ground. Just saying that race stands in an important causal relationship with a given feature doesn't by itself make any headway in determining whether acting on that feature is acting on race, unless that observation works from an account of race in light of which that feature and causal relation are privileged over others. Okay, I'm going to really quickly mute my video because they're trained passing by. Okay, so the principal issue with these approaches is not just that facts about causation don't give us enough information to pick out, you know, the race features from the non race features, but it's that any implications that these approaches have for solving the puzzle will just depend entirely on the theory of race that they posit. Right. And the theory that I think that the approaches themselves can't deliver. So I think the disagreements across the different solutions to the puzzle stem from disagreements about what race as a social category. And I do want to say, though, approaches are in fact suggestive of an account of race, right, they highlight a picture of race as producing causal effects in the world. And they suggest in doing so they suggest that what race does is a key feature of what race is. How is it possible that there is another train. I'm sorry about this. I wholeheartedly endorse this proposal of thinking race to be a social cause but I think we need to think really deeply about what kind of a social cause is and I think they're a social constructivist account can. Okay. Well, surely there won't be another one in the next 15 minutes. All right. So, I think to make sense of a causes interpretation of discrimination on the basis of race, we need to elaborate an analysis of race as a social category that is causally efficacious. Right. But I think the straightforward thought that race acts as a cause. And I think that's kind of an embrace being just an individual attribute that kind of trigger certain responses right so that the claim the police person stop Jamal because he was black seems to suggest something like at time T zero the police person perceives Jamal perceives race, and then that triggers a certain set of behavioral responses that ends in the police person stopping Jamal at time T one. And that kind of causal understanding leaps kind of completely opaque the background conditions of the interaction right so why did black have the causal effect that it did. Why did the police officer react the way did with, you know, with stop and search rather than with pass. And I think if a causes interpretation discrimination is primarily concerned with the systemic nature of racial disadvantage, then answers to these causal questions should explain why black has the causal capacities that it does, and why black has the causal capacities are configured in this way. So I think that here, what we want is I think a picture of race not as a kind of attribute that is perceived by, but as a social category that structures certain kinds of social responses. The definition of why Jamal being Jamal's being black caused the police officer to search him or stop him has to address the fact that the black carry social meaning, right, or that the concept of black is linked to in America concepts of kind of danger or violence. And I think we can't really understand how black acts constantly, you know, causes the police search, without understanding the origins of this representational content. And this is what social constructivism really highlights. Social constructivism highlights that the myriad causal powers of Jamal being raised black are connected with the social kind of thing that being raised black is and how the category is socially constructed as such. So constructively foreground this idea that racial categories are socially constructed to have the causal powers that they have right and socially constructed because of social and political forces continually shaping the boundaries of racial categories, their meanings and as a result of their causal capacities. And since races don't share any essential characteristics that naturally mark them off from each other, the divisions can only be maintained and produced a new by race making institutions. So I'm going to use the example of police and the prison system in the US, though I'm not a scholar of policing and prisons in the US and for people who, you know, want to learn more about this which we all should I think this little like explanation isn't going to do justice to the work that so many scholars have done and so just to name a few. Elizabeth Hinton, Ruthie Gilmore, Julie Kohler house and like I'm happy to kind of give names I think this work is really important. And also don't want people to think this cursory explanation of it is, you know, kind of covers the expansive scholarship. But, let me say a little bit about this which is that the expansion of the American carceral system, starting in the late 60s reinforced representations of blacks as deviance as criminals as generally dangerous and the rise in black crime which was itself a product of race laden policies further justifies certain policing practices right practices like racial profiling, such that blackness can be becomes both equated with criminality at this, you know, micro level of everyday social meeting and police interaction, but also this macro level of legitimized racial policing policies and just general policymaking rationale. And these effects have spilled over in the past 50 years, and in many other spheres that making blacks further socially economically and politically marginalized and very in highly specific ways. And the concept of race is thereby further embedded in our society figures into new and different non accidental generalizations and asserts new and different causal pressures on black lives. So I think a further way to flesh out this critical constructivist project is to think about race as essentially hierarchically constructed. And I'm going to call this kind of strand of constructivist thought thick constructivism. So, for instance, Sally Haslinger for Sally Haslinger, you know, race and gender or social positions, where in certain characteristics of phenotype and ancestry. In the case of race justifies certain kinds of differential treatment and positioning. In quote, racial oppression or racial oppression is directly connected to correlations of race. So she writes group domination and the effects of earlier injustice positions, coordinate groups socially and economically so that their members have much more in common than their group membership. And I think seeing that the race structure as a hierarchical structure means seeing continuity in how blacks are subordinated, even if the particulars of that subordination might vary by historical moment. Because the category is essentially constructed in such a way that black individuals occupy a social position that is defined by a set of norms and expectations. And as a whole oppressive, and I think this fact of racial oppression connects directly with the correlations with race issue at the heart of the puzzle of algorithmic discrimination. Because these correlations, you know, robust non-accessual correlations between outcomes of disadvantage and the black label that we often see in machine learning and in data sets generally, there are simply just manifestations of an unjust race structure at work. Because look, if races are constructed as different, then seeing these consistent non-accidental correlations of race are not just kind of unfortunate statistical regularities that, you know, algorithms have to find a way to skirt around in order to not, you know, not to discriminate. But I think they track empirical facts descriptive of what races qua thick social positions are, right, what it is to be raised in the first instance. So, for the thick constructivist the racial category black is defined in reference to its overall position of disadvantage. So for instance to be black is to be marked on the basis of one's phenotype and supposed ancestry for certain norms expectations and practices of support nation. And I think different thick constructivist accounts can be distinguished by the kinds of subordinating factors that they're emphasizing and defining the category. So for instance, do boys in dusk of dawn he has this dialogue that he's speaking with a white interlocutor and the white interlocutor asks him, what is this group and how do you differentiate it and how can you call it black when you admit it is not black. So the boy's response, he says, the black man is a person who must ride Jim Crow in Georgia. And he therefore he thereby defines black by its relationship to legal institutions of racial subordination. While Taylor expands this kind of thick constructivist like approach to the race concept more generally so for Taylor races are probabilistically defined population that result from the white supremacist determination to link appearance and ancestry with social location and life chances. Notice in these accounts that race continues to be a causal factor in individuals lives absolutely right being marked black causally subjects into black individuals to certain types of social treatment. But the constructivist account is of the categories constitutive because black is in fact defined as to be black is to define to be standing in some relation to social factors and social agents. So taking this view we would kind of you know if we have for instance an example here being black is partly defined in terms of standing one standing in a certain relation to institutions of policing. Right. If we took this view then we would expect policing data to be racially skewed. Right a positive correlation between the features black and past searches like police doesn't just expose a causal effect of what an individual is being marked black but a descriptive effect about what black is as a social category. And if we kind of continue on thinking about the racial hierarchy racialized policing institutions actually for the distinct pillar in that racial hierarchical race structure. And so they partly actually form the thick social position that is the category black. And you know if you take this view then there is really no clean distinction between something. This race are kind of itself and the complex of social factors that constitute the thick social position of race are. Okay so I want to make a few caveats and clarifications before I get to the upshot. At this point I want to first say that I think constructed this account of race doesn't need to take all non accidental correlations with the feature black to be constitutive of the racial category black. I think some I think I would posit like most causal effects of being labeled black and encountering the world as black are not defining features of the category but I think the question of which ones are essential and which ones aren't is an important conceptual. And I was just like a political question that isn't reducible to data mining right so I don't want I don't mean to suggest like. Oh constitutive you know those features that are constitutive are the ones that exhibit the strongest or most robust correlations or anything like that. I also, you know, a potential objection might be well, you know, doesn't a feature need to be kind of perfectly correlated with the racial label black in order for it to be kind of this essential. constitutive feature of the category. And I want to say no and I think this is because standing in certain relations of subordination to policing doesn't mean having to have certain kinds of policing outcomes right so not every individual experiences. The last thing I want to say, which again will sound very cursory because I think a lot of work should be dedicated to this. But I think that doesn't mean that subordinating standing a subordinate relation with to policing isn't constitutive of the category, because I think in fact all black individuals quaw black individuals do experience race based harms and injustices because of policing and I can talk more about that maybe in the q amp a. I just want to say, which again will sound very cursory because I think a lot of work should be dedicated to this is that, you know, intersectional theorists intersectionality theorists have, you know, urged us to break out of single axis thinking that there is on the one hand, you know, the race hierarchy and the gender hierarchy and you can kind of like, just add them up to think about oppression, you know, at the intersection and I want to say that that's absolutely not the case. And I think that constructivism, what must be developed further to kind of account for intersectional social categories. But I think that that's going to require a ton of other work and deserves its own kind of dedicated attention. Okay, so what is the cash value of this. So I think with these theoretical resources in place I claim we can elaborate a causal analysis of discrimination on the basis of race. The answer is the puzzle that I set forth at the beginning of this talk. But I think we need to understand how race acts as a cause in the right way. And I think the right way is understanding advanced by the thick constructivist account. So here it is. When it comes to algorithmic decision systems, discriminating on the basis of features correlated with race constitutes discriminating on the basis of race, when those features are not merely causally related to but constitutive of the racial category So let's kind of run through an example so you know that constructivists define being black in terms of a certain sort of supporting social factors so here maybe black is a thick social position whose occupants are probabilistically more likely to And I have some examples here be on the receiving end of a stop and frisk action be a victim of state violence and so on. Now, here we have social facts in this box there are social facts that constitute the racial category black quite thick social position. And I want to say that acting on the feature black in a data set and acting on these features the features in this box are kind of the dividing line between those for the thick constructivism that constructivist is not really there, right. The constructivist might see that an algorithm that includes the race race feature and algorithm that includes these features that exhibit non accidental correlations that tracking the social facts are constitutive of being raised or might both be discriminating on the basis of race. And I think this helps understand something that I think maybe many of you might have heard in discussions about algorithmic discrimination, which is this this claim that an algorithm can learn race, even if you exclude race, right. I think the thought here seems to be that in so far as race leaves its marks, leaves its mark on so many non race features, keeping those features around in the machine learning pipeline is going to contribute to certain racially biased outcomes. I think the thick constructivist interpretation is to interpret the claim quite literally it literally learns race, it learns via these correlations with race what race is and acting on the basis of correlations produced by social facts that constitute race in a social position just is acting on the basis of race, because races just are social positions that subject their members, member individuals to this matrix of privileging and supporting social relations that are just precisely what it is, you know, define what it is to be raised. Okay, some implications for discrimination theory I know I'm at the half hour mark so I'll kind of try to run through this. Rather quickly I think I only have like two slides, maybe one slide for I think one slide. I think there are a lot of implications for discrimination theory more broadly I want to discuss a few here. So I think one place where the standard machinery of discrimination theory seems to malfunction is this commonly held distinction between direct and indirect discrimination. You know, kind of to explain that distinction rather roughly, which there's still like scholarly debate about how best to characterize it. It's, you know, the thought roughly is that indirect discrimination kind of against blacks harms blacks, but not by kind of acting on race itself but it has some negative disparate impact on the group right so for instance, you know because blacks have been more often been subject to certain policing practices, they indirectly bear the brunt of other policies that kind of use those police encounters as a decision making criterion. Right, that's an indirect kind of discrimination. But if as the constructivist holds the category black is defined in terms of those social facts that constitute the position in a hierarchical race structure, then the gap between acting on being black and acting on having been targeted by violent policing practices narrow down significantly. Right, it seems to act on black in an important sense when that challenges the idea that doing so only acts on being black kind of indirectly. I also, you know, I said, you know, here's this account it solves the puzzle and I want to give you a quick reason or two for why we should adopt the fit constructivist account. I think that the accounts of our social categories, you know, should be constructed with certain theoretical and practical goals. And I think the fit constructivist account has, you know, gives us payoff. And let me, I'm going to work with an analogy to economic class. Okay, to show that I think the fit constructivist account like gives us theoretical and practical payoff. I suppose I'm an orthodox Marxian and I take class to be defined as an orthodox objective as Marxian terms so I think your economic class 100% established by your relationship to the means of production. So under capitalism, you're a member of the proletariat or a member of the bourgeoisie. I remember that what Marx called being a proletarian is correlated with having a long working day, limited personal property, and being subject to behavioral control on and off the job. I find that my algorithm can predict with great accuracy which of these individuals will experience financial hardship in the next months, and that the people who will experience financial hardship. It's heavily skewed towards those labeled proletariat, even if I remove the class category. One interpretation that says okay well being a proletariat causally you know influences you having a long working day and so on. I think this conclusion is fine but it seems unsatisfying and maybe even a little bit wrong. It seems that when I use an algorithm to draw on the fact that you frequently experience financial hardship and behavior discipline. I act directly on the fact that you are any particular precarious and subordinated position within the economic structure. A distinctive social position that perhaps warrants attention as a class category. But this thought isn't really available to me if I'm kind of pre committed to the Marxian account. I kind of can't see that in fact these features in fact constitute a really important category of theoretical interest. It seems right to me to say that if I deny you social benefits on the basis of the fact that you have a long working day limited personal property and are subject to behavioral discipline that I'm doing so on the basis of your class directly. And I think this is a claim that we can make on a thick constructivist conception. So that example is just to show I think different conceptions of our categories bear different theoretical and practical normative fruit. The thick constructivist category would serve us well both theoretically and in many of our political projects or many of my political projects. I also think okay this is really one minute now I also think that the constructivist the thick constructivist account of race shows why acting on the basis of races, often wrongful. So in discrimination theory, there's a question of what the wrong making feature of discrimination is. I think if you take it that constructivist account. It's kind of clear to see how acting on the basis of being black is demeaning as Deborah Hellman would say is a wrong making feature of discrimination or compounds injustice or reinforces hierarchy. So I think it kind of connects the descriptive and the nondis or sorry that more lies than non more lies quite well. And the last thing I'll say is that there are really serious practical limitations of talking about certain features is only causally related to and not constitutive of race. And that cashes out in a series of really horrible in my view judicial decisions in discrimination cases in the US context and I'm happy to talk about that more in Q&A, but I'll leave that at that. Thank you for inviting me and allowing me to go a bit over. Thanks. Thank you so much, Lily. That was a really wonderful talk. I realize I didn't introduce my own self, which is also I'm Shelly Adlundson, the Chief Operations Officer of the HMI Grand Challenge. And what I'd like to do right now that was hand over to Sarita Rosenstock, the research fellow at HMI and she's going to moderate our Q&A. So Sarita, would you like to take over please? Thank you. Hi everybody. So today I'm going to start by calling on our interdisciplinary panel who's associated with the HMI group here. And then I'm going to move on to audience questions. And so you can ask the question in the Q&A function here and you can either I can either read it for you or you can ask it for yourself just to indicate in the question. So I'm going to start by calling on Atuza, who's another research fellow here with the HMI group based in the computer science department. Atuza. Hi Lily and thank you so much for the great talk. So I'm totally in board with the defense of the thick constructivist account of race. And I think maybe that's the right way to think about race if we are thinking about like ontological questions about race. But what I'm a little bit fearful is that if we like imagine we want to apply this notion, which is a correct conception of metaphysics of race to the case of Facebook, like the example that you started. And so the worry is that if we adopt this dislike very socially kind of picture, it could be that in a couple of legal systems, it wouldn't allow us or we wouldn't allow the legal system to demarcate when this notion of race or basically when there was a racial discrimination or not. Because at some point, maybe you can, you know, if you say that, okay, there's not much, we don't need to have like very strong correlations between features, particular features and the notion of race or any other social category. Then at some point, we can end up like saying that yeah, at some point, many different features are correlated with the social categories. And that seems to be a kind of a problem for, for like legal decision making in practical settings. And I was wondering if you have something to say about that kind of worry. Yeah. Um, sorry, so I'm going to try to phrase the worry and you can correct me if I'm getting it wrong. So the concern is if we don't have this kind of clear operationalization of the constructivist account that kind of maps on to correlations, then it might be in fact difficult. We might kind of have to like, not have any knowledge of which features are problematic to act on is that the concern. Yeah. Yeah. Yeah, I mean, I think that first I think that there are a lot of first there are a lot of like social scientists who are working on issues of race, both empirical social scientists and also social theorists who've studied all sorts of hierarchy and I think that, you know, our conceptions of race have to kind of take into account historical as well as empirical phenomena that might not be born out directly in things like correlations right sometimes they will be like, for instance, policing really you know we can see the stark divisions in policing, but you know other other kinds of cases of oppression are not so easily identifiable from the data. And I don't know I don't want this to be hand wavy question but I just think that it's something that we're going to have to do we're going to have to come come to like what in what specific ways certain groups are oppressed and marginalized and keep note of that kind of cluster of social facts. Whenever we try to adjudicate questions about group inequality or group discrimination. And I think another thing to say that about that is that we have to do this. We have to do this because this is part of, I can kind of squeeze in a response, a little bit about my practical limitations, because I write a lot about how race reasoning about race and discrimination cashes out and the US legal system, and how it cashes out is that nothing is constitutive of race right. And so all you need to do to show that there's no discrimination, the statistical the econometrician comes in waves their statistical wand and says look if I condition on neighborhood poverty rates neighborhood crime rates socioeconomic status, you know, by condition on all these extremely racialized factors, there's no race difference at all. And it's true you'll see this across in policing cases, you know, for the people from the US Floyd versus the city of New York, NYPD. That's a case where we had to ask whether stop and frisk was racially biased. And both sides said look, if we take this thin account of race, you can condition on anything and ask what the marginal effective race is so I think, just as a practical matter you do have to decide what are the things that you cannot condition on as a matter of statistical practice, because it is constitutive of race. So I think, you know, it might be hard to get out that question but like otherwise we're never going to be able to prove discrimination practically. Sorry, that didn't answer. All right, so next on the queue we have Colin Klein from here in the philosophy department. Hey, thanks, Lily. That was kind of a few here. All right. Let's hope. All right, good. No thanks. It was really interesting, like, I think much more subtle than a lot of these usually are. So I'm going to show this is partly my contrastivist roots thinking about causation but when you were thinking through this it seemed like you were focusing mainly on this question of is something discriminating on the base rather than not. But it seems like in many cases people care about some other contrast. So just put to one is on the basis of race versus income. I put it in terms of like does income completely screen off the effects of race. You might also care about, I mean, so you're quite focused on the US context and particularly African American experience. So I care does it discriminate on the basis of race. So sometimes you say look it doesn't because we treat Jews and Asians fine. So I'm not discriminating on the basis of race. We're really what you mean is African American versus everyone else or what you mean is white versus everyone else. And I'm partly curious in the last one because this is a kind of generalization question because in Australia we don't have a lot of African Americans they mostly work for the State Department. And the indigenous say, you know, they don't face heaps of discrimination unless they're mistaken for indigenous people who are horrifically discriminated against but in a different pattern than the US. I'm just trying to tease out how much the social constructivist account is just focused on that first question of race rather than not or whether you get a sense of how it could be extended to cover the other kinds of contrast as well. Yeah, that's a great question. Thanks. Well, the first thing which is raised rather than something else I think a lot of and I didn't talk about this too much in this paper in this talk but a lot of my work is interested in like well what do you mean race and some non race thing. Right like race rather than let's say like socioeconomic status right. I think that for people who are familiar with these kind of perennial debates in in the left about whether something is in fact a racial disparity or a class disparity. Oftentimes the way that these kinds of conversations cash out statistically is like, well when I just for class I the racial disparity shrinks by this much. And the thing I want to say there is is, you know, why would you think that in a society in which black and white people had the same wealth distribution that we would even care like that you know how this phenomenon let's say police brutality would affect blacks would even be the same right. Like why would you think that in a in a world in which black people had the same incomes that blackness in that world reflects what blackness in our world is today so I think. So I want to problematize the this like clean distinction. A little bit. The second thing about Remind me sorry the second thing was generalizing okay generalizing or so black so African American versus everything else African American versus white economically successful minorities versus actually in America you know like you can carve this up quite a number of different ways, especially in the American context. Yeah, yeah, totally. Um, so for people who might be familiar there's this, there was this case at Harvard discrimination against Asians and admissions to Harvard, right. And I don't want to give this more airtime than it deserves, but I've been thinking about this for a while and part of the issue is like, we have a racial minority in this case Asians, but their character. They have different kinds of attributes like academic attributes. And there's a question of whether acting on the basis of certain positive attributes or, you know, when we're acting against Asians right. I think the, you know, one of the things that I think is interesting about racial discrimination is it's is the way I parse it is like, how do we treat similar people similarly, given that they have one particular difference which is that they're different racially. Right. And I think the only way to fill in that question of given their different racially is to kind of think about how these two groups in not just the black white case but let's say Asians and Latinx people, how they are compared relative certain like for me like certain material outcomes. And I think those kinds of axes really affect what it means to treat those groups similarly or differently. And again to think about these groups as social positions that are constituted by certain advantages and privileges and disadvantages and injustices. But you know I do think it is comparative, it has to be a comparative question, you know, the default is not just like white right the default isn't just like how are you compared to white but they have to be comparative across racial categories. I'm being a good interlocutor and letting other people ask questions. I have plenty more to say if no one else wants to say. So next I want to call on Katie Steele from philosophy department here. Hi Lily, thanks. So yeah I like your constructivist account of race and I took your point that we no longer on this account get such a clean distinction between direct and indirect discrimination. But I mean this kind of ties in a little bit with what Atusa was getting at earlier. There's still going to be a problem of a spectrum of directness of discrimination. It'll be sort of a continuum if you like. You know there's a sort of causal like maybe past criminal record is pretty closely causally related with what it means to be black. But then there's like stuff that's a bit further causally downstream from that maybe credit history or you know whatever. So yeah there's still going to be a question of like where you draw the line perhaps. Yeah yeah totally and I guess what I would say is I think that the direct I don't want to say that my account totally like eviscerates the direct and indirect distinction. I think it you know I leave space because a lot of times the direct indirect discrimination is cashed out is just disparate impact. So I'm going to leave room for the fact that you can act on correlations with race that are not constitutive that nevertheless produce disparate impact and say I'm happy to take on board that that's going to be indirect. I also think there's a danger of subsuming too much we want to we want this concept discrimination to do too much for us. And I think the critiques of indirect discrimination are right when they think when they say something like oftentimes we're concerned about distributive justice. And you know things can be kind of wrong for many reasons and it doesn't need to be because it's discrimination on the basis of race it could be wrong or unjust as a matter of distributive justice just as a matter of what people get. So I'm kind of willing to say this might shrink the cases. You know I don't feel a need to kind of make the concept discrimination do a lot. If things are more costly distal but they nevertheless have disparate impact. I'm happy to call them indirect discrimination. I'm also happy to charge that they're unjust on other grounds. I see. All right. Next is will from the law department here. Thanks Rita and thanks Lily that was such a fantastic paper. Good on you really excellent work and I particularly liked it because I'm someone who's who left philosophy about a decade ago and then went into law and your paper was really engaged with the nuts and bolts and the concrete aspects of law in a really admirable way which is often lost in a lot of the engagements with this topic outside outside law faculties. Of course law faculties have our own problems because we don't think about any of the conceptual stuff in a sensible or sophisticated way but so thanks. Thanks so much for scratching that itch for me. I've got a comment and then a question. The comment is on the direct and indirect discrimination division as you put it. I want to suggest that indirect discrimination or adverse impact. And I know that there's overlaps between those ideas but they come from a single source. As a matter of history, I want to suggest that that is formulated as a legal category to attempt to target direct discrimination in an absence of meaningful evidence of the mental state at the discriminator. And I'm not sure exactly what impact that has on your paper but it seems that there's a weight attached to some kind of meaningful categorical distinction between the types of discrimination which are targeted by direct and indirect discrimination. And if you look at just just as a matter of the kind of the raw intellectual history of the law here, you see it really clearly in the Griggs case originated that originated adverse impact. That kind of consequential approach to attaching liability is the result of consequential consequences of actions which is discriminatory is can be understood and I think should be understood as an attempt to say look. We're actually targeting the mind meant the specific mental process of the perpetrator but we don't have sufficient evidence of that mental state so we have to adopt a proxy and the proxy is group impact. So that's the comment. The question is, and it's probably a bit of a difficult question. How would you drive the discrimination statute which expressed a constructivist conception of protected attributes. Oh boy. Oh my gosh. That's a hard question because I'm asking you to be a legislator on the spot, but it's just, it's a provocation to think about the positive. So, okay, so in the US content. Okay, I don't want to get into the nuts and bolts. I know we don't have that much time but I'm going to say something that perhaps seems like even more radical but I don't. I think as a result of my willing to direct and indirect discrimination is I'm kind of willing to give up on this like prioritization of mental states this like racial animus mechanism that direct discrimination like intentional discrimination tries to get out. I actually think that most many cases of disparate impact peer disparate impact. I'm kind of happy at you know if discrimination theory could be a lot of that you know I would be happy for that to kind of take primacy over direct discrimination. And that's not directly responsive but I do want to say something that your earlier comment made me think of so people on here might be familiar with the boss doc case the US Supreme Court had just decided, which says that discrimination on the basis of sexual orientation, gender non conforming status and trans status is discrimination on the basis of sex. I think that is quite happy, you know, aligned with something I want to say here, which is there I think there's a blurring of the lines of what it means to act on the basis of sex, where it would seem that acting on, you know, how you dress would only be kind of causally related to rather than constitutive of, but it seems like then the acting on gender non conforming status. Yeah, can be seen as saying look the way you dress is so you know how you present how you dress is so tied in with the category of sex that we're going to say that acting on the base of that is acting on the base of sex. And I think that's right. Sorry, well that that was me just shoehorning in the thing I wanted to say in response to your difficult practical question, but I'll own up to it. We can we can go bilateral on the difficult practical question. Yeah. Okay, so next I'm going to call on Tiberio who is coming from the gradient Institute in Sydney. Hello, ladies. Thank you so much for such a lovely talk. It's a subject that I'm very interested to learn more about. I think my question is around if you within this account, right, you know, how could you tell if you're making progress towards less racial discrimination. Would, you know, one way to think about it, if you, if you could conceive of a test that would determine whether racial discrimination had been completely eradicated. Which, how would you think about that? Which test would would you be would actually within this account be able to answer that question. Or even incremental progress, right, because from the point of view of, for example, the Facebook case you described, and for any pragmatic purposes. One is restricted to make decisions about make deliberations based on limited amount of information, right. And we need a specific criteria to make decisions about, you know, whether we should do A or B. And that hinges on notions of what a baseline is. Where are we today in terms of racial discrimination? Where we're going to be tomorrow in terms of racial discrimination. For instance, is there a simple test to within this account to answer the question of whether today we have more racial discrimination than we had last year. Even though today there's much more awareness than there was last year or five years ago. So how would you respond to that? Thank you. I think this question of how much racial discrimination there is in the world, or in our in a given society is a really important one. And I think that current accounts. So, okay, let me start with this. In the paper version of this talk, I kind of end on this observation of the stark disparities in how, for instance, racial minorities in the US perceive the prevalence of racial discrimination and how white people perceive the prevalence of racial discrimination. The Pew, you can like kind of every year that Pew Center does these surveys and they're they're astonishing. We're talking about three times four times the amount of expect of racial discrimination that people of color say exists versus white people. So I think what this draws out to me is the idea of like that there is an objective number amount of racial discrimination out there that an account can kind of like get a hold of is it's really challenging. And I think part of the reason that many of the conclusions kind of my paper or my talk here and how it might differ from standard accounts of racial discrimination is it makes, at least I think on my account there is a lot more racial discrimination. And I think a lot more cases look like count as cases of racial discrimination. That I think other accounts miss and I think part of the problem to kind of put the discipline a little bit on blast to say, well, it's not surprising right if the thick construct of this kind of race is true. Then it wouldn't be surprising to see that there are actually material kind of differences in the perception of race and what kinds of cases are canonical cases that my theory has to get right and what kind you know how my methodology continues right so That's to say I think like this this really important empirical question like how much racial discrimination is out there. I don't even know how to really respond to that because these stark disparities and in the perceptions of ordinary people show there to be just. I think not even overlapping, not just like empirical measurements are different like people are exposed to more or less racial discrimination, but I think it shows that there might not even be convergence on a concept right So philosophers and legal scars might have one concept of racial discrimination, but I think other people might have people people of color might have a different concept and I think that challenges the question of how much racial discrimination is out there because whose concept are we trying to measure. Again, I don't mean to be evasive. So we have one last question from Jenny in sociology before we move this conversation over into the slack. Sorry. Hi, that was fantastic talk. Really interesting to really appreciate the things that you're thinking about in the way that you're integrating the sort of social critical lens with the technical issues of how we manage it. And so I'll say that this question is also written in slack. So if people are interested in talking more about it since we've got I think a minute left. I'll come over and do so. But my question is this so, so given that race profoundly shapes life chances which it does and which I think of course underlies the thick reading of race. What if we put this idea of race as a reason back on the table, right so it's already clearly a cause might highlighting race rather than trying to diminish this importance help us sort of think about unambiguously how race matters. And potentially in both technical and social ways, mitigate racial disparities or at least sort of weigh them plain, make them clear. Yeah, thank you so much. So in I tried to cut out a lot of content but in the paper I kind of say like oh it looks like I did a sleight of hand like I started talking about the causal analysis and all of a sudden I'm saying like, no what it is to take, you know act on race is to act on these other features and I'm doing the reasons analysis like what is this trickery. So like I tried to acknowledge that a little bit. But I think the, in the end, I think that I do want to advocate for. Because like, I think it's very hard our current conceptions of what it is to take something as a reason. They don't, at least the ones that like at least how that is tends to be taken in conversations of racial discrimination are rather unsatisfying because you talk about race as a reason and all of a sudden you're in intentionality world you're in like, you know, what was your mental state, you know, were you acting with the racial animosity mechanism on turned on or not like were you in fact in, you know, implicit bias land or were you just acting, I was acting on race I was acting on like their attire or something. I think that that picture is is really unhelpful so like I don't if I if tying myself to the reasons view means to commit myself to this kind of intentionality standard racial animus racial as a prejudice view I don't want to do that. But I want to rather I want to say like we need to have a better a fuller account of what it is to act on the base of race period, even as a reason and and I don't yet have that. But I think that would require us to do a lot more thinking about how to revise what it is to take social categories as reasons generally when they're markers of really hierarchical or social meaning. So a lot more in depth revision of reasons generally, which I didn't want to have to dig through in this paper but I'm, I'm absolutely interested and advocate for that project.