Jon Krohn: 00:03
This is Five-Minute Friday on the Equality Machine.
00:19
Professor Orly Lobel, welcome to the Super Data Science Podcast. It’s awesome to have you on the show. You are a distinguished professor of law at the University of San Diego. You are the director of the University of San Diego Center for Employment and Labor Policy. You have two previous books in the employment and IP spaces. Your third book, the Equality Machine: Harnessing Digital Technology for a Brighter More Inclusive Future came out in October. We’re going to talk about your book in a second, but for our listeners, this law professor has a very interesting data analyst background. So you’re a data analyst in the Israeli Military Intelligence. That is super interesting. I don’t know if you want to… You probably can’t share very much about that.
Orly Lobel: 01:06
Well, first of all, thanks for having me, Jon, and I’m excited to be on your show. It’s terrific. You’re right. I can’t share that much about the military intelligence experience, just to say that I do actually write about it in the book that it was really my first encounter with the power of technology and digital data mining to do good and bad, to have this equalizing force. So it was the one place in the military that was very gender equal because we were using our minds and having this digital paper trail of our contributions. And it gave me a lot of sense of how I wanted to think about tech policy and research what has been going on. And since then there’s been so many developments. I’m not going to claim expertise like some of your listeners of real data science these days, but I’m very interested in the developments and we’ll talk about that.
Jon Krohn: 02:12
Yeah. Well, you certainly do have expertise in using digital technologies for equality. You mentioned equality there already. So in your book, the Equality Machine, you provide countless examples of how digital technologies including AI can address issues as diverse as poverty and justice, climate issues, health safety and many more. That kind of thing sounds amazing to a techno optimist like me. This is what I want to hear. I’m like, “This kind of interview lets me really get my confirmation biases going.” If you have a few pertinent examples of how you think specifically data and AI digitization could make a big positive impact, I’d love to hear them.
Orly Lobel: 02:57
Yeah, the whole purpose of the book is really having a more constructive and cautiously optimistic vision and blueprint of where we’re going with technology. Because as you know, we are in this moment of a techlash and a mindset not only that is critical of big tech, but really of the technologies themselves, which I find really problematic. I see a lot of not very informed critiques of AI and algorithms. The book takes us on this wild ride of the celebrations and great stories and success stories and trajectories, real changes that have been made even when there were fails of using AI and automation in certain contexts, showing how we can improve systems. So one example that’s near and dear to my research that I talk about in chapter three is how data science has been changing and tackling this very stagnant pay gap and salary inequities.
04:16
For years we’ve had laws that say equal pay for equal work, but because there’s been such knowledge asymmetries, what is your worth? How is talent valued in the labor market? It’s been closely held just by companies, by employers. We’ve seen real stagnation in the gender and race pay gaps. Now things are really shifting very fast with third party platforms where there’s crowdsourcing of your current salary, your past salary, looking at policies actually that demand to introduce pay scales and governments that are collecting data about compensation and analyzing that. And also just software that leaders in this space, private industry leaders are integrating into their HR systems to constantly audit the ways that unintentionally a lot of times pay gaps just grow and stay.
05:40
I’ll just mention there’s so many… that’s just one example in the labor market, but I look at, as you mentioned, correcting health disparities and how we’ve traditionally counted only certain demographies when we did clinical trials and personalized medicine just in the design of a lot of systems like speech to text, text to speech, personal digital assistance, what I show is that we need more data. One of the things that as a law professor, I get very nervous about when I look at the federal government and bills before Congress right now, and also at the EU, there’s this idea of data minimization, and I consult governments all the time and I’m trying to shift that mindset of more data is harmful. Again, I think your listeners are sophisticated to understand and that that’s just not true, that we need data to really detect disparities and know what’s wrong and correct things. But yeah, we can really take this to any field and lots of examples.
Jon Krohn: 07:02
Nice. The main idea here is that whether we’re talking about labor markets or healthcare, safety climate issues, it’s by having data on sociodemographic characteristics, for example, that we can have a record of how different groups are being treated and so we can actually reduce historical disparities that have occurred?
Orly Lobel: 07:29
Yeah. That’s one way to think about the use of data, but we can also better understand root causes of ongoing disparities. We can know where to invest more resources. We can actually lower costs to increase access. So one example, for example, in the health field that I talk about is when we’re automating things like radiologists screening for breast cancer or any type of radiology screening. A lot of times I see this fallacy of comparing, the media will say, “Well, right now maybe the technology is at a point where two human radiologists that are both looking at one x-ray are as good, or maybe slightly outperforming an AI that’s doing the same work.” Using a lot of my behavioral research and understandings about cause what I argue is that that’s just not the right comparison, because most people around the world will not have access to highly trained radiologists. So if we can really spread digital literacy and digital access to these new technologies that use data in these ways, we really are creating much more inclusion and equality.
Jon Krohn: 09:09
Very cool. Yeah, that is another really great example. What do you think about how our models could incorporate historical biases? Do you discuss that in your book, how we could be resolving these kinds of… if there was an injustice in historical data and then we train a machine learning model on it, that could mean… that’s often the issue that is brought up to me as one of the big problems with data and AI. Do you have resolutions for that kind of problem?
Orly Lobel: 09:39
Yeah, absolutely. I actually started the book with that. So the equality machine is never set to deny that there have been these kinds of, what other very critical techlash mindset have called automating inequality or algorithm bias, weapons of math destruction. So there’s a lot of that sensibility that we’re just replicating and amplifying past wrongs because we’re training the algorithm not only on a skewed set of data, but also on our historical exclusions and inequities. I go through a lot of the developments, for example, in hiring automation system. It’s certainly true that if all you’re doing with trying to automate resume parsing is to train an algorithm of like, “Here’s our past successful engineers, and you just need to find people like that. We’re going to just replicate past wrongs and we’re really doing not much.”
11:03
But what I see again, and I show this that those same stories of fails just get retold. And that’s not really the gold standard these days that a lot of companies are understanding that they need to do. I go through a lot of different alternatives where it’s not about exploiting past data, but you really telling the algorithm to explore other paths of finding highly talented people, whether it’s through games or things like using facial recognition, which is very highly controversial with emotional recognition.
11:46
What I show in every turn in the book, whether we’re talking about hiring or even automating an adjudicatory system like bail releases or sentencing, is that, first of all, we don’t need to expect an algorithm to be perfect and perfectly equal in terms of its outcomes. What we really need to do is look at the comparative advantages compared to the status quo. And you actually Jon already mentioned, you said confirmation biases or you alluded to all these biases that we humans are very prone to. As somebody who’s studied, and I’ve served as an expert witness in many cases on discrimination, and I teach it, and we’re just human black box algorithms that have so many conscious, unconscious fallibilities and biases. And so we really should not be talking about just flaws in this absolute way about algorithms, but we need to see what is doing better, what can improve in a general trajectory of having more equal systems and outcomes.
Jon Krohn: 13:13
Yeah. That all makes perfect sense to me. And as an interesting example of how humans of course have that kind of black box as you described us, where this unknown black box in terms of how we make decisions. One place that we can cause a lot of damage with our machine learning systems is the way that we label data. So whether we’re using, in your example of the recruitment technology, something near and dear to my heart because I have a recruitment machine learning company myself, and so we have to be very thoughtful about what data we use and how we use it. And so one place that you can get into trouble is if you’re labeling data and saying, “Okay, this person is a good fit for the job, this person isn’t,” and you’re using those labels to train the model. If you have a human doing those labels, there’s going to be situations where they have biases.
14:08
Well they’re always going to have biases, not there’s going to be situations, they are always going to have biases. We’re not going to know what those are. Some of them are going to be subtle, some of them are going to be unconscious, some of those are going to be really bad. But there are kinds of technological solutions for getting around these issues. In episode number 635 with Shayan Mohanty, we talked about ways of labeling data automatically using rules where when we use these heuristic rules to label the data, we have a record of exactly what heuristics we chose, and then we can go back through those as opposed to just having humans making these decisions saying, “Yep, this person is a good fit for the job, this person isn’t a good fit,” and we have no way of going back and saying, “Why did they make that decision?”
14:57
Part of my techno-optimism, to use that term that I used earlier in the show is that we are increasingly aware, I think in recent years of the risks that data and AI can have in society. And so now we’re devising technological solutions like Shayan Mohanty in episode number 635 for overcoming some of those issues. And I think that it’ll continue.
Orly Lobel: 15:23
Yeah. That’s absolutely right. And I should interview you also, Jon, about the equipment tools that you’re developing, but that’s exactly the exercise and research and showcasing that I do in the equality machine I’m going through lots of different fields. So I do the same with… I have a chapter that’s called Algorithms of Desire, and I show how we can get dating apps to do better matching and helping nudging people overcome some of our historical and ongoing biases and in-group preferences, not in a way that’s too paternalistic, but that really helps us line up with the norms and values that we care about. And you can do that with everything, but you’re right, there are solutions, there are technical solutions and there are policy solutions that we need to get in place to do it better.
16:28
I also do… I talk about it in a different chapter, I do consulting on content moderation. And again, they’re like human moderators, are notorious and mislabeling and labeling according to a lot of cultural beliefs that don’t necessarily line up with what our goals are, democratic goals. But there are ways to shift beyond that and really use machine learning to highlight those very biases that we’re even doing in the labeling process. And importantly, we’re getting better at it and we can check the outcome. One of the points that I really argue for in the book is that we need to actually collect what the EU and the GDPR and a lot of laws here call sensitive information that should not be collected. We actually should collect sensitive information, demographic information in the input, so we can actually have a check with the outputs of whether we’re doing something that has a disparate impact and whether we can toss that and do something better. Because when we’re completely clean, slating it or anonymizing in that way, we don’t have that learning curve that we can develop.
Jon Krohn: 18:03
Yeah, perfectly stated. The title of your book as well as a term that you’ve mentioned a number of times in this episode is Equality Machine. Is that a discrete machine that you can define or is it more of a broad concept?
Orly Lobel: 18:23
Yeah. I think it’s actually both. It’s my response to all these titles and terms that we have that I already mentioned, algorithmic bias and automating inequality and surveillance, capitalism and weapons of math destruction. There’s a lot of that doomsday dystopian ideas of what machines can do for us. And so the equality machine is both a mindset of let’s try to do better, but it’s also a set of principles. It’s a blueprint and a vision of not having a double standard for humans and machines. So looking at the comparative advantages, looking at the trajectory of actually being able to improve something and keeping a digital trail of having these checks and using data where it’s most needed, figuring out what do we care about and count what matters.
19:28
So actually understanding that data collection has always been political, even in the most low tech settings, just like census data or whatever we’ve always done in the clinical trials and government data collection and industry, there’s always been a tilt to it. And now with our really vast leaps and computational abilities, we can put our energies to having more representative, more complete data sets and use them in the ways that we want and lining up with our democratic values.
Jon Krohn: 20:15
Nice. I love that. And I’m sure many of our listeners do too. So check out Orly’s new book, the Equality Machine. It’s currently at the time of recording, I don’t know if you know this Orly, it’s the number one new release on Amazon in the category of computer and internet law. It’s obviously doing really well as a new release. Check it out listeners, will be sure to have a link to the book in the show notes. Orly, clearly you have a lot of expert knowledge in this domain, and it’s directly relevant to a lot of our listeners, especially the ones that want to make a big social impact, a big positive social impact with the data science models that they build. How can they, after this episode follow you?
Orly Lobel: 20:58
They should check out my platforms everywhere. I’m Orly Lobel on Twitter, on LinkedIn, on Instagram, on Facebook. And I would love to connect to all readers and especially learn more examples that you are working on, whether you’re… you’re right on how you framed it that it’s relevant to anybody who wants to do good, but really AI for good is not just the purview of activists or governments, it’s really in the private industry. It is happening. And I want to really showcase those people who have skin in the game, so not just telling the bad stories, but really understanding that so many of us care about these issues and we can do good and do well. And so yes, please connect with me, Orly Lobel. And thank you for having me.
Jon Krohn: 21:55
Yeah. You’re most welcome, Professor Lobel. Thank you so much for being on the show today. All right, that’s it for this Five-Minute Friday episode. Until next time, keep on rocking it out there folks. And I’m looking forward to enjoying another round of the Super Data Science Podcast with you very soon.
