Michael Kearns: Algorithmic Fairness, Privacy & Ethics | Lex Fridman Podcast #50

The following is a conversation with Michael Kearns. He's a professor at the University of Pennsylvania and a co author of the new book Ethical Algorithm that is the focus of much of this conversation. It includes algorithmic fairness, bias, privacy, and ethics in general. But that is just one of many fields that Michael is a world class researcher in, some of which we touch on quickly, including learning theory or the theoretical foundation of machine learning, game theory, quantitative finance, computational social science, and much more. But on a personal note, when I was an undergrad, early on I worked with Michael on an algorithmic trading project and competition that he led. That's when I first fell in love with algorithmic game theory. While most of my research life has been in machine learning and human robot interaction, the systematic way that game theory reveals the beautiful structure in our competitive and cooperating world of humans has been a continued inspiration to me. So for that and other things, I'm deeply thankful to Michael and really enjoyed having this conversation again in person after so many years. This is the Artificial Intelligence podcast.

If you enjoy it, subscribe on YouTube.

Give it five stars on Apple Podcasts, support on Patreon or simply connect with me on Twitter Alex Friedman Spelled F R I D M A n this episode is supported by an amazing podcast called Pessimists Archive. Jason, the host of the show, reached out to me looking to support this podcast and so I listened to it to check it out. And by listened I mean I went through it Netflix binge style, at least five episodes in a row. It's not one of my favorite podcasts and I think it should be one of the top podcasts in the world, frankly. It's a history show about why people resist new things. Each episode looks at a moment in history when something new was introduced, something that today we think of as commonplace, like recorded music, umbrellas, bicycles, cars, chess, coffee, the elevator. And the show explores why it freaked everyone out. The latest episode on Mirrors and Vanity still stays with me as I think about vanity in the modern day of the Twitter world. That's the fascinating thing about this show is that stuff that happened long ago, especially in terms of our fear of new things, repeats itself in the modern day and so has many lessons for us to think about in terms of human psychology and the role of technology in our society. Anyway, you should subscribe and listen to Pessimist Archive. I highly recommend it. And now here's my conversation with with Michael Kearns.

You mentioned reading Fear and Loathing in Las Vegas in high school and having a more or a bit more of a literary mind. So what books, non technical, non computer science would you say had the biggest impact on your life, either intellectually or emotionally?

You've dug deep into my history, I see.

Went deep.

Yeah, I think. Well, my favorite novel is Infinite Jest by David Foster Wallace, which actually, coincidentally, much of it takes place in the halls of buildings right around us here at mit. So that certainly had a big influence on me. And as you noticed, like when I was in high school, I actually even started college as an English major, so was very influenced by sort of that genre of journalism at the time and thought I wanted to be a writer and then realized that an English major teaches you to read, but it doesn't teach you how to write. And then I became interested in math and computer computer science instead.

Well, in your new book, Ethical Algorithm, you kind of sneak up from a algorithmic perspective on these deep, profound philosophical questions of fairness, of privacy. In thinking about these topics, how often do you return to that literary mind that you had?

Yeah, I'd like to claim there was a deeper connection, but I think both Aaron and I kind of came at these topics first and foremost from a technical angle. I mean, I kind of consider myself primarily and originally a machine learning researcher. And I think as we just watched, like the rest of the society, the field technically advance and then quickly on the heels of that kind of the buzz kill of all of the antisocial behavior by algorithms, just kind of realized there was an opportunity for us to do something about it from a research perspective. You know, more to the point in your question, I mean, I do have an uncle who is literally a moral philosopher. And so in the early days of our technical work on fairness topics, I would occasionally run ideas behind him. So, I mean, I remember an early email I sent to him in which I said, like, oh, here's a specific definition of algorithmic fairness that we think is some sort of variant of Rawlsian fairness. What do you think? And I thought I was asking a yes or no question. And I got back your kind of classical philosopher's response, saying, well, it depends. If you look at it this way, then you might conclude this. And that's when I realized that there was a real kind of rift between the ways philosophers and others had thought about things like fairness from sort of a humanitarian perspective and the way that you needed to think about it as a computer scientist if you were going to kind of implement actual algorithmic solutions.

But I would say the algorithmic solutions take care of some of the low hanging fruit sort of the problem is a lot of algorithms, when they don't consider fairness, they are just terribly unfair. And when they don't consider privacy, they're terribly, they violate privacy. Sort of the algorithmic approach fixes the big problems. But there's still, you get, when you start pushing into the gray area, that's when you start getting into this philosophy of what it means to be fair, starting from Plato, what, what is justice? Kind of questions.

Yeah, I think that's right. And I, I mean, I, I, I would even not go as far as you went to say that that sort of, the algorithmic work in these areas is solving like the biggest problems. And, you know, we discuss in the book the fact that really we are, there's a sense in which we're kind of looking where the light is in that. You know, for example, if police are racist in who they decide to stop and frisk and that goes into the data, there's sort of no undoing that downstream by kind of clever algorithmic methods. And I think, especially in fairness, I mean, I think less so in privacy, where we feel like the community kind of really has settled on the right definition, which is differential privacy. If you just look at the algorithmic fairness literature already, you can see it's going to be much more of a mess. And you've got these theorems saying, here are three entirely reasonable, desirable notions of fairness, and here's a proof that you cannot simultaneously have all three of them. So I think we know that algorithmic fairness compared to algorithmic privacy is going to be kind of a harder problem, and it will have to revisit, I think, things that have been thought about by many generations of scholars before us. So it's very early days for fairness, I think.

So before we get into details of differential privacy and on the fairness side, let me linger on the philosophy a bit. Do you think most people are fundamentally good, or do most of us have both the capacity for good and evil within us?

I mean, I'm an optimist. I tend to think that most people are good and want to do right and that deviations from that are kind of usually due to circumstance, not due to people being bad at heart.

With people with power, are people at the heads of governments, people at the heads of companies, people at the heads of maybe so financial power, markets. Do you think the distribution there is also most people are good and have good intent?

Yeah, I do. I mean, my statement wasn't qualified to people not in positions of power. I mean, I think what happens in a lot of the cliche about absolute power corrupts absolutely. I mean, I think even short of that, having spent a lot of time on Wall street and also in arenas very, very different from Wall street, like academia, one of the things I think I benefited from by moving between two very different worlds is you become aware that these worlds kind of develop their own social norms and they develop their own rationales for behavior. For instance, that might look unusual to outsiders, but when you're in that world, it doesn't feel unusual at all. And I think this is true of a lot of professional cultures, for instance. So then you're maybe slippery slope is too strong of a word, but you're in some world where you're mainly around other people with the same kind of viewpoints and training and worldview as you. And I think that's more of a source of abuses of power than sort of, you know, there being good people and evil people, and that somehow the evil people are the ones that somehow rise to power.

Oh, that's really interesting. So it's the. Within the social norms constructed by that particular group of people, you're all trying to do good. But because as a group, you might drift into something that, for the broader population, it does not align with the values of society.

That's the word.

Yeah. I mean. Or not that you drift, but even that things that don't make sense to the outside world don't seem unusual to you. So it's not sort of like a good or a bad thing. But, you know, like. So, for instance, you know, in the world of finance, right, there's a lot of complicated types of activity that if you are not immersed in that world, you cannot see why the purpose of that, you know, that activity exists at all. It just seems like, you know, completely useless. And people just like, you know, pushing money around. And. And when you're in that world and you learn more, your view does become more nuanced. You realize, okay, there is actually a function to this activity. And in some cases, you would conclude that actually, if magically we could eradicate this activity tomorrow, it would come back because it actually is serving some useful purpose. It's just a useful purpose that's very difficult for outsiders to see. And so I think lots of professional work environments or cultures, as I might put it, kind of have these social norms that, you know, don't make sense to the outside world. Academia is the same, right? I mean, lots of people look at academia and say, you know, what the hell are all of you people doing? Why are you paid so Much in some cases at taxpayer expenses to do, you know, to publish papers that nobody reads, you know, but when you're in that world, you come to see the value for it and. But even though you might not be able to explain it to, you know, the person in the street. Right.

And in the case of the financial.

Sector, tools like credit might not make sense to people. It's a good example of something that does seem to pop up and be useful or just the power of markets and just in general, capitalism.

Yeah. In finance, I think the primary example I would give is leverage. Right. So being allowed to borrow to sort of use 10 times as much money as you've actually borrowed. Right. So that's an example of something that before I had any experience in financial markets, I might have looked at and said, well, what is the purpose of that? That just seems very dangerous. And it is dangerous and it has proven dangerous. But if the fact of the matter is that sort of on some particular timescale you are holding positions that are very unlikely to lose your value at risk or variance is 1 or 5%, then it kind of makes sense that you would be allowed to use a little bit more than you have because you have some confidence that you're not going to lose it all in a single day. Now, of course, when that happens, we've seen what happens not too long ago, but the idea that it serves no useful economic purpose under any circumstances is definitely not true.

We'll return to the other side of the coast, Silicon Valley, and the problems there. As we talk about privacy, as we talk about fairness at the high level, and I'll ask some sort of basic questions with a hope to get at the fundamental nature of reality, but from a very high level, what is an ethical algorithm? So I can say that an algorithm has a running time of using big O notation, N log n. I can say that a machine learning algorithm classified cat versus dog with 97% accuracy. Do you think there will one day be a way to measure sort of in the same compelling way as the big O notation of this algorithm is 97% ethical?

First of all, let me riff for a second on your specific n log n example, because early in the book, when we're just kind of trying to describe algorithms, period, we say, like, okay, what's an example of an algorithm or an algorithmic problem? First of all, like, it's sorting, right? You have a bunch of index cards with numbers on them and you want to sort them. And we describe an algorithm that sweeps all the way through, finds the smallest number puts it at the front, then sweeps through again, finds the second smallest number. So we make the point that this is an algorithm and it's also a bad algorithm in the sense that it's quadratic rather than n log n, which we know is optimal for sorting. And we make the point that sort of like. So even within the confines of a very precisely specified problem, there might be many, many different algorithms for the same problem with different properties. Some might be faster in terms of running time, some might use less memory, some might have better distributed implementations. The point is that already we're used to in computer science thinking about trade offs between different types of quantities and resources and there being better and worse algorithms. And our book is about that part of algorithmic ethics that we know how to kind of put on that same kind of quantitative footing right now. So, you know, just to say something that our book is not about, our book is not about kind of broad, fuzzy notions of fairness. It's about very specific notions of fairness. There's more than one of them. There are tensions between them. Right. But if you pick one of them, you can do something akin to saying that this algorithm is 97% ethical. You can say, for instance, the, you know, for this lending model, the false rejection rate on black people and white people is within 3%. Right. So we might call that a 97% ethical algorithm. And a 100% ethical algorithm would mean that that difference is 0%.

In that case, fairness is specified when two groups, however they're defined, are given to you.

That's right.

And then you can sort of mathematically start describing the algorithm.

But.

Nevertheless, the part where the two groups are given to you, unlike running time, you know, we don't in computer science talk about how fast an algorithm feels like when it runs. True, we measure it and ethical starts getting into feelings. So for example, an algorithm runs, you know, if it runs in the background, it doesn't disturb the performance of my system, it'll feel nice, I'll be okay with it, but if it overloads the system, it'll feel unpleasant. So in that same way ethics, there's a feeling of how socially acceptable it is. How does it represent the moral standards of our society today? So in that sense, and sorry to linger on that first high level philosophical question is, do you have a sense we'll be able to measure how ethical an algorithm is?

First of all, I certainly didn't mean to give the impression that you can kind of measure memory speed trade offs and that there's a Complete, you know, mapping from that onto kind of fairness, for instance, or ethics and accuracy, for example, in the type of fairness definitions that are largely the objects of study today and starting to be deployed, you as the user of the definitions, you need to make some hard decisions before you even get to the point of designing fair algorithms. One of them, for instance, is deciding who it is that you're worried about protecting, who you're worried about being harmed by, for instance, some notion of discrimination or unfairness. And then you need to also decide what constitutes harm. So for instance, in a lending application, maybe you decide that, you know, falsely rejecting a creditworthy individual, you know, sort of a false negative is the real harm and that false positives, that is people that are not creditworthy or are not going to repay your loan, that get a loan, you might think of them as lucky. And so that's not a harm. Although it's not clear that if you are don't have the means to repay a loan, that being given a loan is not also a harm. So, you know, you know, the literature is sort of so far quite limited in that you sort of need to say who do you want to protect and what would constitute harm to that group? And when you ask questions like will algorithms feel ethical? One way in which they won't under the definitions that I'm describing is if, you know, if you are an individual who is falsely denied a loan, incorrectly denied a loan, all of these definitions basically say like, well, you know, your compensation is the knowledge that we are also falsely denying loans to other people, you know, in other groups at the same rate that we're doing. It's to you and, and, and, you know, there, and so there is actually this interesting, even technical tension in the field right now between these sort of group notions of fairness and notions of fairness that might actually feel like real fairness to individuals, right? They, they might really feel like their particular interests are being protected or thought about by the algorithm rather than just, you know, the groups that they happen to be members of.

Is there parallels to the big O notation of worst case analysis? So is it important to looking at the worst violation of fairness for an individual, is it important to minimize that one individual? So, like worst case analysis, is that something you think about or.

I mean, I think we're not even at the point where we can sensibly think about that. So first of all, you know, we're talking here both about fairness applied at the group level, which is a relatively weak thing, but it's better than nothing. And also the more ambitious thing of trying to give some individual promises. But even that doesn't incorporate, I think, something that you're hinting at here is what a CHAI might call subjective fairness. Right. So a lot of the definitions, I mean, all of the definitions in the algorithmic fairness literature are what I would kind of call received wisdom definitions. It's sort of, you know, somebody like me sits around and thinks like, okay, you know, I think here's a technical definition of fairness that I think people should want or that they should, you know, think of as some notion of fairness. Maybe not the only one, maybe not the best one, maybe not the last one. But we really actually don't know from a subjective standpoint, like what people really think is fair. There's, you know, we just started doing a little bit of work in, in our group and at actually doing kind of human subject experiments in which we, you know, ask people about, you know, we, we ask them questions about fairness, we survey them, we, you know, we show them pairs of individuals in, let's say, a criminal recidivism prediction setting. And we ask them, do you think these two individuals should be treated the same as a matter of fairness? And to my knowledge, there's not a large literature in which ordinary people are asked about, you know, they, they have sort of notions of their subjective fairness elicited from them. It's mainly, you know, kind of scholars who think about fairness kind of making up their own definitions. And I think, I think this needs to change actually, for many social norms, not just for fairness. Right. So there's a lot of discussion these days in the AI community about interpretable AI or understandable AI. And as far as I can tell, everybody agrees that deep learning, or at least the outputs of deep learning, are not very understandable. And people might agree that sparse linear models with integer coefficients are more understandable, but nobody's really asked people. There's very little literature on showing people models and asking them do they understand what the model is doing. And I think that in all these topics, as these fields mature, we need to start doing more behavioral work.

Yeah. Which is. So one of my deep passions is psychology. And I always thought computer scientists will be the best future psychologist, in a sense, that data is, especially in this modern world, the data is a really powerful way to understand and study human behavior. And you've explored that with your game theory side of work as well.

Yeah. I'd like to think that what you say is true about computer scientists and psychology. From my own limited wandering into human subject experiments we have a great deal to learn. Not just computer science, but AI and machine learning more specifically, I kind of think of as imperialist research communities in that, you know, kind of like physicists in an earlier generation. Computer scientists kind of don't think of any scientific topic as off limits to them. They will like, freely wander into areas that others have been thinking about for decades or longer. And you know, we usually tend to embarrass ourselves in those efforts for, for some amount of time. Like, you know, I think reinforcement learning is a good example. Right. So a lot of the early work in reinforcement learning, I have complete sympathy for the con, the control theorists that looked at this and said like, okay, you are reinventing stuff that we've known since like the 40s, right. But you know, in my view, eventually this sort of, you know, computer scientists have made significant contributions to that field, even though we kind of embarrassed ourselves for the first decade. So I think if computer scientists are going to start engaging in kind of psychology, human, human subjects type of research, we should expect to be embarrassing ourselves for a good 10 years or so and then hope that it turns out, as well as some other areas that we've waded into.

So you've kind of mentioned this just to linger on the idea of an ethical algorithm of idea of groups, sort of group thinking and individual thinking. And we're struggling that one of the amazing things about algorithms and your book and just this field of study is it gets us to ask like forcing machines, converting these ideas into algorithms is forcing us to ask questions of ourselves as a human civilization. So there's a lot of people now in public discourse doing sort of group thinking, thinking like there's particular sets of groups that we don't want to discriminate against and so on. And then there is individuals sort of in the individual life stories, the struggles they went through and so on. Now, like in philosophy, it's easier to do group thinking because you don't, you know, it's very hard to think about individuals. There's so much variability. But with data you can start to actually say, you know what, group thinking is too crude. You're actually just doing more discrimination by thinking in terms of groups and individuals. Can you linger on that kind of idea of group versus individual and ethics and is it good to continue thinking in terms of groups in algorithms?

So let me start by answering a very good high level question with a slightly narrow technical response, which is these group definitions of fairness, like here's a few groups, like different racial groups, maybe gender groups, maybe age, what have you. And let's make sure that for none of these groups do we have a false negative rate, which is much higher than any other one of these groups. So these are kind of classic group aggregate notions of fairness, but at the end of the day, an individual you can think of as a combination of all of their attributes. They're a member of a racial group, they have a gender, they have an age and many other, you know, demographic properties that are not biological but that, you know, are. Are still, you know, very strong determinants of outcome and personality and the like. So one, I think, useful spectrum is to sort of think about that array between the group and the specific individual, and to realize that in some ways, asking for fairness at the individual level is to sort of ask for group fairness simultaneously for all possible combinations of groups. So in particular. So in particular, if I build a predictive model that meets some definition of fairness by race, by gender, by age, by what have you, marginally, to get slightly technical, sort of independently, I shouldn't expect that model to not discriminate against disabled Hispanic women over age 55 making less than $50,000 a year annually, even though I might have protected each one of those attributes marginally.

So the optimization, actually, that's a fascinating way to put it. So you're just optimizing the one way to achieve the optimizing fairness for individuals, just to add more and more definitions of groups that each individual belongs to.

So at the end of the day, we could think of all of ourselves as groups of size one, because eventually there's some attribute that separates you from me and everybody from everybody else in the world. Okay? And so it is possible to put, you know, these incredibly coarse ways of thinking about fairness and these very, very individualistic, specific ways on a common scale. And, you know, one of the things we've worked on from a research perspective is you. So we sort of know how to, in relative terms, we know how to provide fairness guarantees at the coarsest end of the scale. We don't know how to provide kind of sensible, tractable, realistic fairness guarantees at the individual level. But maybe we could start creeping towards that by dealing with more refined subgroups. I mean, we gave a name to this phenomenon where you protect, you enforce some definition of fairness for a bunch of marginal attributes or features, but then you find yourself discriminating against a combination of them. We call that fairness gerrymandering, because, like political gerrymandering, you're giving some guarantee at the aggregate level, but that when you kind of look in a more granular way at what's going on. You realize that you're achieving that aggregate guarantee by sort of favoring some groups and discriminating against other ones. And so there are, you know, it's early days, but there are algorithmic approaches that let you start creep, creeping towards that, you know, individual end of the spectrum.

Does there need to be human input in the form of weighing the value of the importance of each kind of group? So, for example, is it. Is it like, so gender, say, crudely speaking, male and female and then different races? Are we as humans supposed to put value on saying gender is 0.6 and race is 0.4 in terms of. In the big optimization of achieving fairness, Is that kind of what humans are supposed to do?

I mean, of course, you know, I don't need to tell you that, of course, technically one could incorporate such weights if you wanted to, into a definition of fairness. Fairness is an interesting topic in that, having worked in the book, being about both fairness, privacy, and many other social norms, fairness, of course, is a much, much more loaded topic. So privacy, I mean, people want privacy. People don't like violations of privacy. Violations of privacy cause damage, angst, and bad publicity for the companies that are victims of them. But sort of everybody agrees more data privacy would be better than less data privacy. And you don't have these. Somehow the discussions of fairness don't become politicized along other dimensions like race and about gender and whether we should. And it becomes. You quickly find yourselves kind of revisiting topics that have been kind of unresolved forever. Like affirmative action. Right? Sort of, you know, like, why are you protecting. Some people will say, why are you protecting this particular racial group? And, and others will say, well, we need to do that as a matter of, of retribution. Other people will say, it's a matter of economic opportunity. And I don't know which of whether any of these are the right answers. But fairness is sort of special in that as soon as you start talking about it, you inevitably have to participate in debates about fair to whom, at what expense, to who else? I mean, even in criminal justice, where people talk about fairness in criminal sentencing or, you know, predicting failures to appear or making parole decisions or the like, they will, you know, they'll point out that, well, these definitions of fairness are all about fairness for the criminals and what about fairness for the victims, right? So when I, when I basically say something like, well, the, the false incarceration rate for black people and white people needs to be roughly the same, you know, there's no mention of potential victims of criminals in such a fairness definition.

And that's the realm of public discourse I should actually recommend. I just listened to people listening. Intelligence Squares debates. US Edition just had a debate. They have this structure where you have old Oxford style or whatever they're called debates. It was two versus two. And they talked about affirmative action. And it was. Is incredibly interesting that it's still. There's really good points on every side of this issue, which is fascinating to listen to.

Yeah, yeah, I agree. And so it's interesting to be a researcher trying to do, for the most part, technical algorithmic work. But Aaron and I both quickly learned you cannot do that and then go out and talk about it and expect people to take it seriously. If you're unwilling to engage in these broader debates that are entirely extra algorithmic. Right. They're not about algorithms and making algorithms better. They're sort of, as you said, sort of like, what should society be protecting in the first place?

When you discuss the fairness, an algorithm that achieves fairness, whether in the constraints and the objective function, there's an immediate kind of analysis you can perform which is saying if you care about fairness in gender, this is the amount that you have to pay for in terms of the performance of the system. Like, do you. Is there a role for statements like that in a table, in a paper? Or do you want to really not touch that?

Like, no, we, we want to touch that and we do touch it. So I mean, just, just again, to make sure I'm not promising your, your viewers more than we know how to provide. But if you pick a definition, I'm worried about gender discrimination, and you pick a notion of harm, like false rejection for a loan, for example, and you give me a model, I can definitely, first of all, go audit that model. It's easy for me to go from data to kind of say like, okay, your false rejection rate on women is this much higher than it is on men. But once you also put the fairness into your objective function, I mean, I think the table that you're talking about is what we would call the Pareto curve. You can literally trace out and we give examples of such plots on real data sets in the book, you have two axes on the X axis is your error on the Y axis is unfairness by whatever. If it's like the disparity between false rejection rates between two groups. And your algorithm now has a knob that basically says, how strongly do I want to enforce fairness? And the less unfair if the two axes are error and unfairness, we'd like to be at 0, 0. We'd like 0 error and 0 unfairness simultaneously. Anybody who works in machine learning knows that you're generally not going to get to zero error period without any fairness constraint whatsoever. So that's not going to happen. But in general you'll get some kind of convex curve that specifies the numerical trade off you face. If I want to go from 17% error down to 16% error, what will be the increase in unfairness that I experience as a result of that? This curve specifies the undominated models. Models that are off that curve are, you know, can be strictly improved in one or both dimensions. You can, you know, either make the error better or the unfairness better or both. And I think our view is that not only are, are these objects, these Pareto curves, you know, efficient frontiers as you might call them, not only are they valuable scientific objects, I actually think that they in the near term might need to be the interface between researchers working in the field and stakeholders in given problems. So you know, you could really imagine telling a criminal jurisdiction, look, if you're concerned about racial fairness, but you're also concerned about accuracy, you want to, you know, you want to release on parole people that are not going to recommit a violent crime and you don't want to release the ones who are. So you know, that's accuracy. But if you also care about those, you know, the mistakes you make not being disproportionately on one racial group or another, you can show this curve. I'm hoping that in the near future it'll be possible to explain these curves to non technical people that have, that are the ones that have to make the decision where do we want to be on this curve? Like what are the relative merits or value of having lower error versus lower unfairness? You know, that's not something computer scientists should be deciding for society, right. That you know, the, the people in the field, so to speak, the policymakers, the regulators, that's who should be making these decisions. But I think and hope that they can be made to understand that these trade offs generally exist and that you need to pick a point and like, and ignoring the trade off, you know, you're implicitly picking a point anyway. Right, right. You just don't know it and you're.

Not admitting it just to linger on the point of trade offs. I think that's a really important thing to sort of think about. So you think when we start to optimize for fairness, there's almost always in most system Going to be trade offs. Can you, like, what's the trade off between. Just to clarify, there have been some sort of technical terms thrown around, but a sort of a perfectly fair world. Why is that? Why will somebody be upset about that?

The specific trade off I talked about, just in order to make things very concrete, was between numerical error and some numerical measure of unfairness.

What is numerical error in the case.

Of just like, say, predictive error, like, you know, the probability or frequency with which you release somebody on parole who then goes on to recommit a violent crime or keep incarcerated somebody who would not have recommitted a violent crime.

So in the case of awarding somebody parole or giving somebody parole or letting them out on parole, you don't want them to recommit a crime. So it's your system failed in prediction if they happen to do a crime. Okay, so that's the performance. That's one axis.

Right.

And what's the fairness axis?

So then the fairness axis might be the difference between racial groups in the kind of false positive predictions, namely people that I kept incarcerated predicting that they would recommit a violent crime, when in fact they wouldn't have.

And the unfairness of that, just to linger it and allow me to ineloquently to try to sort of describe why that's unfair, why unfairness is there, the unfairness you want to get rid of is that in the judge's mind, the bias of having being brought up to society, the slight racial bias, the racism that exists in the society, you want to remove that from the system. Another way that's been debated is sort of equality of opportunity versus equality of outcome. And there's a weird dance there that's really difficult to get right. And we don't. The affirmative action is exploring that space.

Right. And then this also quickly bleeds into questions like, well, maybe if one group really does recommit crimes at a higher rate, the reason for that is that at some earlier point in the pipeline or earlier in their lives, they didn't receive the same resources that the other group did. And so there's always in kind of fairness discussions, the possibility that the real injustice came earlier. Right. Earlier in this individual's life, earlier in this group's history, et cetera, et cetera.

And so a lot of the fairness discussion is almost the goal is for it to be a corrective mechanism to account for the injustice earlier in life.

By some definitions of fairness or some theories of fairness. Yeah. Others would say, like, look, it's, it's, you know, it's not to correct that injustice. It's just to kind of level the playing field right now and not incarcerate, falsely incarcerate more people of one group than another group. But I mean, do you think just it might be helpful just to demystify a little bit about the many ways in which bias or unfairness can come into algorithms, especially in the machine learning era? And I think many of your viewers have probably heard these examples before, but let's say I'm building a face recognition system. And so I'm kind of gathering lots of images of faces and trying to train the system to recognize new faces of those individuals from training on a training set of those faces of individuals. And it shouldn't surprise anybody, or certainly not anybody in the field of machine learning if my training data set was primarily white males, and I'm training the model to maximize the overall accuracy on my training data set that, you know, the model can reduce its error most by getting things right on the white males that constitute the majority of the data set, even if that means that on other groups they will be less accurate. Okay, now there's a bunch of ways you could think about addressing this. One is to deliberately put into the objective of the algorithm not to, not to optimize the error at the expense of this discrimination. And then you're kind of back in the land of these kind of two dimensional numerical trade offs. A valid counter argument is to say like, well, no, you don't have to. There's no, you know, the notion of the tension between error and accuracy here is a false one. You could instead just go out and get much more data on these other groups that are in the minority and equalize your data set. Or you could train a separate model on those subgroups and have multiple models. The point I think we tried to make in the book is that those things have cost too, right? Going out and gathering more data on groups that are relatively rare compared to your plurality or majority group. That it may not cost you in the accuracy of the model, but it's going to cost, it's going to cost the company developing this model more money to develop that. And it also costs more money to build separate predictive models and to implement and deploy them. So even if you can find a way to avoid the tension between error and accuracy in training a model, you might push the cost somewhere else. Like money like development time, research time and the like.

They're fundamentally difficult philosophical questions. In fairness. And we live in a very divisive political climate, outrage culture. There Is alt right folks on 4chan trolls, there is social justice warriors on Twitter. There is very divisive, outraged folks on all sides of every kind of system. How do you. How do we, as engineers build ethical algorithms in such divisive culture? Do you think they could be disjoint? The human has to inject your values and then you can optimize over those values. But in our times, when you start actually applying these systems, things get a little bit challenging for the public discourse. How do you think we can proceed?

Yeah, I mean, for the most part in the book, a point that we try to take some pains to make is that we don't view ourselves or people like us as being in the position of deciding for society what the right social norms are, what the right definitions of fairness are. Our main point is to just show that if society or the relevant stakeholders in a particular domain can come to agreement on those sorts of things, there's a way of encoding that into algorithms. In many cases, not in all cases. One other misconception that hopefully we definitely dispel is sometimes people read the title of the book and I think not unnaturally fear that what we're suggesting is that the algorithms themselves should decide what those social norms are and develop their own notions of fairness and privacy or ethics. And we're definitely not suggesting that the.

Title of the book is Ethical Algorithm, by the way, and I didn't think.

Of that interpretation of the title. That's interesting.

Yeah, yeah. I mean, especially these days where people are, you know, concerned about the robots becoming our overlords. The idea that the robots would also, like, sort of develop their own social norms is, you know, just one step away from that. But I do think, you know, obviously, despite disclaimer that people like us shouldn't be making those decisions for society, we are kind of living in a world where in many ways, computer scientists have made some decisions that have fundamentally changed the nature of our society and democracy and sort of civil discourse and deliberation in ways that I think most people generally feel are bad these days. Right. So.

But they had to make. So if we look at people, at the heads of companies and so on, they had to make those decisions. Right? There has to be decisions. So there's two options. Either you kind of put your head in the sand and don't think about these things and just let the algorithm do what it does, or you make decisions about what you value of injecting moral values into the algorithm.

Look, I never mean to be an apologist for the tech industry, But I think it's a little bit too far to sort of say that explicit decisions were made about these things. So let's, for instance, take social media platforms, right? So like many inventions in technology and computer science, a lot of these platforms that we now use regularly kind of started as curiosities, right? I remember when things like Facebook came out in its predecessors, like Friendster, which nobody even remembers. Now people. People really wonder, like, what. Why would anybody want to spend time doing that? You know what I mean? Even. Even the web, when it first came out, when it wasn't populated with much content and it was largely kind of hobbyists building their own kind of ramshackle websites, a lot of people looked at this. This is like, what is the purpose of this thing? Why is this interesting? Who would want to do this? And so even things like Facebook and Twitter, yes, technical decisions were made by engineers, by scientists, by executives in the design of those platforms. But, you know, I don't. I don't think 10 years ago anyone anticipated that those platforms, for instance, might kind of acquire undue influence on political discourse or on the outcomes of elections. And I think the scrutiny that these companies are getting now is entirely appropriate. But I think it's a little too harsh to kind of look at history and sort of say, like, oh, you should have been able to anticipate that this would happen with your platform. And in the sort of gaming chapter of the book, one of the points we're making is that these platforms, they don't operate in isolation. So unlike the other topics we're discussing, like fairness and privacy, those are really cases where algorithms can operate on your data and make decisions about you and you're not even aware of it. Okay, things like Facebook and Twitter, these are systems, right? These are social systems and their evolution, even their technical evolution, because machine learning is involved, is driven in no small part by the behavior of the users themselves and how the users decide to adopt them and how to use them. And so I'm kind of like, who really knew that until we saw it happen? Who knew that these things might be able to influence the outcome of elections? Who knew that, you know, they might polarize political discourse because of the ability to, you know, decide who you interact with on the platform and also with the platform naturally using machine learning to optimize for your own interests, that they would further isolate us from each other and, you know, like, feed us all, basically just the stuff that we already agreed with. And so I think, you know, we've come to that outcome. I Think largely, but I think it's something that we all learned together, including the companies, as these things happen. Now you asked like, well, are there algorithmic remedies to these kinds of things? And again, these are big problems that are not going to be solved with, you know, somebody going in and changing a few lines of code somewhere in a social media platform. But I do think in many ways there are definitely ways of making things better. I mean, like, an obvious recommendation that we make at some point in the book is like, look, you know, to the extent that we think that machine learning applied for personalization purposes in things like newsfeed, you know, or other platforms has led to polarization and intolerance of opposing viewpoints, as, you know, right. These. These algorithms have models, right? And they kind of place people in some kind of metric space, and. And they place content in that space. And they sort of know the extent to which I have an affinity for a particular type of content. And by the same token, they also probably have that same model. Probably gives you a good idea of the stuff I'm likely to violently disagree with or be offended by. Okay, so, you know, in this case, there really is some knob you could tune that says, like, instead of showing people only what they like and what they want, let's show them some stuff that we think that they don't like or that's a little bit further away. And you could even imagine users being able to control this. You know, just like everybody gets a slider and that slider says, like, you know, how much stuff do you want to see? That's kind of, you know, you might disagree with or is at least further from your interest. Like, it's almost like an exploration button.

So just get your intuition. Do you think. Think engagement. So, like, you staying on the platform, you be staying engaged, do you think fairness, ideas of fairness won't emerge? Like, how bad is it to just optimize for engagement? Do you think we'll run into big trouble if we're just optimizing for how much you love the platform?

Well, I mean, optimizing for engagement kind of got us where we are.

So do you. One, have faith that it's possible to do better? And two, if it is, how do we do better?

I mean, it's definitely possible to do different. Right. And again, you know, it's not as if I think that doing something different than optimizing for engagement won't cost these companies in real ways, including revenue and profitability, potentially.

In the short term, at least.

Yeah, in the short Term. Right. And again, you know, if I worked at these companies, I'm sure that it would have seemed like the most natural thing in the world also to want to optimize engagement. Right. And that's good for users. In some sense. You want them to be, you know, vested in the platform and enjoying it and finding it useful, interesting and, or productive. But you know, my point is, is that the idea that there's, that it's sort of out of their hands, as you said, or that there's nothing to do about it, Never say never. But that strikes me as implausible as a machine learning person. Right? I mean, these companies are driven by machine learning and this optimization of engagement is essentially driven by machine learning. Right? It's driven by not just machine learning, but, you know, very, very large scale a B experimentation where you kind of tweak some element of the user interface or tweak some component of an algorithm, or tweak some component or feature of your click through prediction model. And my point is that anytime you know how to optimize for something, almost by definition that solution tells you how not to optimize for it or to do something different.

Engagement can be measured. So sort of optimizing for sort of minimizing divisiveness or maximizing intellectual growth over the lifetime of a human being are very difficult to measure.

That's right. So I'm not claiming that doing something different will immediately make it apparent that this is a good thing for society in particular. I think one way of thinking about where we are on some of these social media platforms is that it kind of feels a bit like we're in a bad equilibrium. Right. That these systems are helping us all kind of optimize something myopically and selfishly for ourselves. And of course, from an individual standpoint at any given moment, like why would I want to see things in my newsfeed that I found irrelevant, offensive, or you know, or the like. Okay. But, you know, maybe by all of us, you know, having these platforms myopically optimized in our interests, we have reached a collective outcome as a society that we're unhappy with in different ways. Let's say with respect to things like, you know, political discourse and tolerance of opposing viewpoints.

And if Mark Zuckerberg gave you a call and said, I'm thinking of taking a sabbatical, could you run Facebook for me for, for six months? What would you.

How I think no thanks would be my first response. But there are many aspects of being the head of the entire company. That are kind of entirely exogenous to many of the things that we're discussing here.

Yes.

And so I don't really think I would need to be CEO of Facebook to kind of implement the, you know, more limited set of solutions that I might imagine. But I think one, one concrete thing they could do is they could experiment with letting people who chose to, to see more stuff in their news feed that is not entirely kind of chosen to optimize for their particular interests, beliefs, etc.

So the, the kind of thing, so I could speak to YouTube, but I think Facebook probably does something similar is they're quite effective at automatically finding what sorts of groups you belong to, not based on race or gender or so on, but based on the kind of stuff you enjoy watching. In the case of YouTube, sort of, it's, it's a difficult thing for Facebook or YouTube to then say, well, you know what, we're going to show you something from a very different cluster, even though we believe algorithmically you're unlikely to enjoy that thing. Sort of. That's a weird jump to make. There has to be a human like at the very top of that system that says, well, that will be long term healthy for you. That's more than an algorithmic decision.

Or, or that same person could say, that'll be long term healthy for the platform or for the platform's influence on society outside of the platform. Right. And you know, it's easy for me to sit here and say these things. Yes. But conceptually I do not think that these are kind of totally or should they shouldn't be kind of completely alien ideas. Right. You know, you could try things like this and it wouldn't be, you know, we wouldn't have to invent entirely new science to do it. Because if we're all already embedded in some metric space and there's a notion of distance between you and me and every other, every piece of content, then, you know, we know exactly. You know, the same model that tells, that dictates how to make me really happy also tells how to make me as unhappy as possible as well.

Right. The focus in your book and algorithmic fairness research today in general is on machine learning, like we said, is data. And just even the entire AI field right now is captivated with machine learning, with deep learning. Do you think ideas in symbolic AI or totally other kinds of approaches are interesting, useful in the space, have some promising ideas in terms of fairness?

I haven't thought about that question specifically in the context of fairness. I definitely would agree with that. Statement in the large. I am one of many machine learning researchers who do believe that the great successes that have been shown in machine learning recently are great successes, but they're on a pretty narrow set of tasks. I mean, I don't think we're kind of notably closer to general artificial intelligence now than we were when I started my career. I mean, there's been progress and I do think that we are kind of, as a community, maybe looking a bit where the light is, but the light is shining pretty bright there right now and we're finding a lot of stuff. So I don't want to like argue with the progress that's been made in areas like deep learning, for example.

This touches another sort of related thing that you mentioned and that people might misinterpret from the title of your book, Ethical Algorithm. Is it possible for the algorithm to automate some of those decisions? Sort of higher level decisions of what.

Kind of like what should be fair?

What should be fair?

The more you know about a field, the more aware you are of its limitations. And so I'm pretty leery of sort of trying. You know, there's so much we already don't know in fairness. Even when we're the ones picking the fairness definitions and comparing alternatives and thinking about the tensions between different definitions that the idea of kind of letting the algorithm start exploring as well, I definitely think, you know, this is a much narrower statement. I definitely think that kind of algorithmic auditing for different types of unfairness, right? So like in this gerrymandering example, where I might want to prevent not just discrimination against very broad categories, but against combinations of broad categories, you quickly get to a point where there's a lot of categories, there's a lot of combinations of N features, and you can use algorithmic techniques to try to find the subgroups on which you're discriminating the most and try to fix that. That's actually kind of the form of one of the algorithms we developed for this fairness gerrymandering problem. But partly because of our technology and our sort of our scientific ignorance on these topics right now, and also partly just because these topics are so loaded emotionally for people that I just don't see the value. I mean, again, never say never, but I just don't think we're at a moment where it's a great time for computer scientists to be rolling out the idea like, hey, not only have we kind of figured fairness out, but we think the algorithms should start deciding what's fair or giving input on that decision. Just don't like the cost benefit analysis to the field of kind of going there right now just doesn't seem worth it to me.

That said, I should say that I think computer scientists should be more philosophically should enrich their thinking about these kinds of things. I think it's been too often used as an excuse for roboticists working on autonomous vehicles, for example, to not think about the human factor or psychology or safety in the same way.

Like computer science design algorithms they've been.

Sort of using as an excuse. And I think it's time for basically everybody to become computer scientists.

I was about to agree with everything you said, except that last point. I think that the other way of looking at it is that I think computer scientists, and many of us are. But we need to wade out into the world more. Right. I mean, just the influence that computer science and therefore computer scientists have had on society at large just like has exponentially magnified in the last 10 or 20 years or so. And you know, before when we were just tinkering around amongst ourselves and it didn't matter that much, there was no need for sort of computer scientists to be citizens of the world more broadly. And I think those days need to be over very, very fast. And I'm not saying everybody needs to do it, but to me, like the right way of doing it is to not to sort of think that everybody else is going to become a computer scientist. But, you know, I think, you know, people are becoming more sophisticated about computer science, even lay people. Yeah. You know, I think one of the reasons we decided to write this book is we thought 10 years ago I wouldn't have tried this just because I, I just didn't think that sort of people's awareness of algorithms and machine learning, you know, the general population would have been high and I mean, you would have had to first, you know, write one of the many books kind of just explicating that topic to a lay audience first. Now, I think we're at the point where like lots of people without any technical training at all know enough about algorithms, machine learning that you can start getting to these nuances of things, like ethical algorithms. I think we agree that there needs to be much more mixing, but I think, I think a lot of the onus of that mixing needs to be on the computer science community.

Yeah. So just to linger on the disagreement, because I do disagree with you on the point that I think if you're a biologist, if you're a chemist, if you're an MBA, business person, all of those things you can, like, if you learn to program, and not only program, if you learn to do machine learning, if you learn to do data science, you immediately become much more powerful in the kinds of things you can do and therefore literature, like library sciences, like, so you were speaking. I think def. I think it holds true what you're saying for the next few years. But long term, if you're interested. To me, if you're interested in philosophy, you should learn to program because then you can scrape data and study what people are thinking about on Twitter and then start making philosophical conclusions about the meaning of life.

Right.

I just, I just feel like the access to data, the digitization of whatever problem you're trying to solve, it fundamentally changes what it means to be a computer scientist. To me, a computer scientist In 20, 30 years will go back to being Donald Knuth style theoretical computer science. And everybody would be doing basically the kind of exploring the kinds of ideas that you're exploring in your book. It won't be a computer science.

Yeah, yeah. I mean, I, I don't, I think, I disagree, but I think that that trend of more and more people and more and more disciplines adopting ideas from computer science, learning how to code, I think that that trend seems firmly underway. I mean, an interesting digressive question along these lines is maybe in 50 years there won't be computer science departments anymore because the field will just sort of be ambient in all of the different disciplines and people will look back and having a computer science department will look like having an electricity department or something. It's like everybody uses this. It's just out there. I mean, I do think there will always be that kind of Knuth style core to it, but it's not an implausible path that we kind of get to the point where the academic discipline of computer science becomes somewhat marginalized because of its very success in kind of infiltrating all of science and society and the humanities, et cetera.

What is differential privacy or more broadly, algorithmic privacy?

Algorithmic privacy, more broadly is just the study or the notion of privacy definitions or norms being encoded inside of algorithms. And so, you know, I think we count among this body of work just the literature and practice of things like data anonymization, which we kind of at the beginning of our discussion of privacy, say, like, okay, this is sort of a notion of algorithmic privacy. It kind of tells you something to go do with data. But our view is that it's, and I think this is now quite widespread, that it's. Despite the fact that those Notions of anonymization, kind of redacting and coarsening are the most widely adopted technical solutions for data privacy. They are deeply, fundamentally flawed. And so to your first question, what is differential privacy? Differential privacy seems to be a much, much better notion of privacy that kind of avoids a lot of the weaknesses of anonymization notions while still letting us do useful stuff with data.

What's anonymization of data?

So by anonymization, I'm kind of referring to techniques like, I have a database. The rows of that database are, let's say, individual people's medical records, okay? And I want to let people use that data. Maybe I want to let researchers access that data to build predictive models for some disease. But I'm worried that that will leak sensitive information about specific people's medical records. So anonymization broadly refers to the set of techniques where I say, like, okay, I'm first going to, like, I'm going to delete the column with people's names. I'm going to not put. So that would be like a redaction, right? I'm just redacting that information. I am going to take ages, and I'm not going to say your exact age. I'm going to say whether you're 0 to 10, 10 to 20, 20 to 30, I might put the first three digits of your zip code, but not the last two, et cetera, et cetera. And so the idea is that through some series of operations like this on the data, I anonymize it. Another term of art that's used is removing personally identifiable information. And this is basically the most common way of providing data privacy. But that's in a way that still lets people access some variant form of the data.

So at a slightly broader picture, as you talk about, what does anonymization mean when you have multiple databases, like with a Netflix prize, when you can start combining stuff together.

So this is exactly the problem with these notions, right, is that notions of anonymization, removing personally identifiable information, the kind of fundamental conceptual flaw is that these definitions kind of pretend as if the data set in question is the only data set that exists in the world or that ever will exist in the future. And of course, things like the Netflix prize and many, many other examples since the Netflix Prize, I think that was one of the earliest ones, though, you know, you can re identify people that were, you know, that were anonymized in the data set by taking that anonymized data set and combining it with other allegedly anonymized data sets and maybe publicly available information about you.

For people who don't know, the Netflix prize was what was being publicly released as data. So the names from those rows were removed. But what was released is the preference or the ratings of what movies you.

Like and you don't like.

And from that combined with other things, I think forum posts and so on, you can start to figure out in.

That case, it was specifically the Internet Movie Database, where, where lots of Netflix users publicly rate their, you know, their movie preferences. And so the anonymized data in Netflix, when it's just this phenomenon, I think that we've all come to realize in the last decade or so, is that just knowing a few apparently irrelevant, innocuous things about you can often act as a fingerprint. Like if I know, you know, what, what rating you gave to these 10 movies and the date on which you entered these movies. This is almost like a fingerprint for, for you is in the sea of all Netflix users. There was just another paper on this in Science or Nature about a month ago that, you know, kind of 18 attributes. I mean, my favorite example of this was actually a paper from several years ago now where it was shown that just from your likes on Facebook, just from the time, you know, the things on which you clicked on the thumbs up button on the platform, not using any information, demographic information, nothing about who your friends are, just knowing the content that you had liked was enough to, you know, in the aggregate, accurately predict things like sexual orientation, drug and alcohol use, whether you were the child of divorced parents. So we live in this era where, you know, even the apparently irrelevant data that we offer about ourselves on public platforms and forums, often unbeknownst to us, more or less acts as signature or, you know, fingerprint. And that if you can kind of, you know, do a join between that kind of data and allegedly anonymized data, you have real trouble.

So is there hope for any kind of privacy in a world where a few likes can. Can identify you?

So there is differential privacy, right?

What is differential?

So differential privacy basically is an kind of alternate, much stronger notion of privacy than these anonymization ideas. And it's a technical definition, but the spirit of it is we compare two alternate worlds. So let's suppose I'm a researcher and I want to do, there's a database of medical records and one of them is yours. And I want to use that database of medical records to build a predictive model for some disease. So based on people's symptoms and test results and the like, I want to, you know, build a probability model predicting the probability that people have disease. So, you know, this is the type of scientific research that we would like to be allowed to continue. And in differential privacy, you act ask a very particular counterfactual question. We basically compare two alternatives. One is, when I do this, I build this model on the database of medical records, including your medical record. And the other one is where I do the same exercise with the same database with just your medical record removed. So basically, you know, it's two databases, one with n records in it and one with n minus one records in it. The n minus one records are the same and the only one that's missing in the second case is your medical record. So differential privacy basically says that any harms that might come to you from the analysis in which your data was included are essentially nearly identical to the harms that would have come to you if the same analysis had been done without your medical record included. So in other words, this doesn't say that bad things cannot happen to you as a result of data analysis. It just says that these bad things were going to happen to you already, even if your data wasn't included. And to give a very concrete example. Right. You know, like, we discuss at some length the study that, you know, in the 50s that was done that created the. That established the link between smoking and lung cancer. And we make the point that, like, well, if your data was used in that analysis and the world kind of knew that you were a smoker because there was no stigma associated with smoking before those findings, real harm might have come to you as a result of that study that your data was included in. In particular, your insurer now might have a higher posterior belief that you might have lung cancer and raise your premium. So you've suffered economic damage. But the point is that if the same analysis has been done with all the other n minus one medical records and just yours missing, the outcome would have been the same. Your data wasn't idiosyncratically crucial to establishing the link between smoking and lung cancer, because the link between smoking and lung cancer is like a fact about the world that can be discovered with any sufficiently large database of medical records.

But that's a very low value of harm. Yeah, so that's showing that very little harm is done. Great.

But how.

What is the mechanism of differential privacy? So that's the kind of beautiful statement of it, but what's the mechanism by which privacy is preserved?

Yeah, so it's basically by adding noise to computations right. So the basic idea is that every differentially private algorithm, first of all, or every good differentially private algorithm, every useful one, is a probabilistic algorithm. So it doesn't on a given input, if you gave the algorithm the same input multiple times, it would give different outputs each time from some distribution. And the way you achieve differential privacy algorithmically is by kind of carefully and tastefully adding noise to a computation in the right places. And you know, to give a very concrete example, if I want to compute the average of a set of numbers, right, the non private way of doing that is to take those numbers and average them and release, like a numerically precise value for the average.