Do Daily Interactions With Autonomous Tech Shape the Way We Think?

Avriel Epps Darling is a researcher, entrepreneur and artist. She is currently a PhD student at Harvard Graduate School of Education, seeking to make a meaningful impact through researching how online machine learning-driven ecologies influence the youth of colour, as they construct and affirm racialized and gendered identities.

Avriel has received numerous awards and honours, including an invitation from the US Department of Education to present her work for Congress. As well as recognition as part of the Top 10% of Undergraduate Social Scientists in The World. Today her research in partnership, with organisations such as Spotify and Snapping, focuses on the intersection of algorithmic bias in content recommendation systems, and racial identity development.

Nikita Johnson, RE•WORK Founder, chatted with Avriel Epps Darling, for our Women in AI Podcast. This is an interview transcript.

Topics explored:

The Intersection of Developmental Psychology and Machine Learning With Computer Science
Bias in Tech Products
The Reason Why Female Artists Are Chronically Underrepresented in the Music Industry
The Relationship Between Algorithmic Streams and Organic Streams
The Role of Streaming Services for Challenging Inequalities by Spotlighting Underrepresented Artists in Their Recommendations
How Different Age Groups Have Different Effects on Streaming
The Correlation Between Fluid Identity and Influence Ability
The Challenges of Gender Labeling
Does Your Background Influence Your Academic Research Later On?
Avriel's Next Step and Ideas for New Research
Advice for Those Wanting to Start a Career or Progress in the Field of AI

🎧 Listen to the podcast here.

Nikita RE•WORK [1:03]

So Avriel, you have a really interesting and varied background. And you're currently a PhD Candidate Forge Fellow and Presidential Scholar at Harvard University. And your passion for your research really comes across in your work, which is something that we had the pleasure of seeing when we heard you at our webinar a couple of months ago. I just wanted to, first of all, start by asking if you could share a bit more about your current research focus to start with.

Avriel [1:30]

Yeah of course. So my work is really at the intersection of developmental psychology and machine learning with computer science. I'm particularly interested in how our daily interactions with autonomous technologies, like recommendation engines and search engines, things like that, shape the way we think about ourselves, and the way we think about our place in the world. I think this is particularly relevant when you take into consideration all of the biases in these tech products, and particularly how they don't work well or as intended, in a lot of cases when interacting with folks of colour or with women. And then, I'm also really interested in adolescence as a period of development, because that's kind of when our identities are most fluid and when they really take shape. And I think an under-researched piece of this whole conversation is around algorithmic biases, how we don't take into account age as a form of bias and how a lot of tech products are normed on adult users. So that is kind of the overarching theme or the overarching themes in my research.

Nikita RE•WORK [2:53]

I mean there's a lot that we could go a bit deeper into. But I think one that's really interesting, so just earlier this month, actually, you published a paper on Artist Gender Representation in Music Streaming. So I want to just have a chat in a bit more detail about that, and one thing I saw from a presentation that you did on some of the findings was one of the findings, in particular, showed that about 1 in 5 streams goes to female artists. So how have music recommendations influenced this?

Avriel [3:26]

So it's no surprise that female artists are chronically underrepresented in the music industry. So this idea that female artists are not getting their fair share of streams is really a historical issue, which is the case for a lot of the biases that we see in AI and machine learning, right. And so what we end up seeing in the data when we look at it is editor programme streams. So the streams that are coming from playlists that human domain experts are curating, those actually have the highest representation of female artists. And it's the streams that people are kind of going and searching for on their own, as well as the streams that are coming from purely algorithmically curated playlists that have the lowest number of female streams, and that's in this 1-in-5, 1-in-6 range. But the interesting thing about that paper and that research is that there is a positive relationship between streaming female artists in algorithmic settings and streaming female artists organically such that as more

Unfortunately, we had technical difficulties & lost connection momentarily.

Nikita RE•WORK [5:25]

So we got most of what you said there, but do you want to just start again or do you want me to ask you the question? Or do you know exactly where you were?

Avriel [5:39]

I think so, I was talking about the positive relationship between algorithmic streams of the artists and organic streams. So what we found through the research was that there is this positive association between algorithmically recommended streams of female artists. And then these kinds of user-initiated organic streams of female artists, such that, as users are listening to more female artists that are recommended to them by the recommendation engines, they will then stream more female artists organically down the line. And so that's really promising because it provides a potential path toward equity using computational systems. I think one of the things that this means for me, and this is a complicated problem, we can't just say, okay, let's go retune these algorithms to make them more equitable in terms of the gender of the artists. But I think there is a precedent to use equity as a sort of, like optimization, or as a function of optimization. So, I think there's an interesting conversation to be had both in academia and industry around what it would look like to optimise algorithms or optimise machine learning problems for equity, what would it mean to bake in gender parity, as either constraint or something that we're looking to maximise? In addition to all of these other metrics, like stream time, or likes of songs that are recommended, and things of this nature, which are just the more traditional metrics that folks use.

Nikita RE•WORK [7:38]

And so what is the role, specifically of streaming services, for challenging inequalities by spotlighting underrepresented artists in their recommendations?

Avriel [7:50]

That's a really good question. So I'm definitely not a spokesperson for Spotify, I just want to be clear about that, but I think that it's a two-pronged problem. On the one hand, because of this historical under-representation of female artists before the advent of streaming, there is a pipeline problem and that's not something that we can ignore. There's a supply problem on the platform. We, in addition to, auditing the algorithms for bias, we also look to see how many female artists were just available to stream and there is severe underrepresentation just in terms of the number of artists who are available to shoot but that becomes a chicken and egg problem because as the money and the resources in the music industry goes to folks who show good metrics in terms of how much they're being streamed, then there are fewer resources for folks who aren't able to show those metrics. And if those happen to be female artists because of biases and recommendation engines, then you just kind of create this self-fulfilling prophecy of who's being allowed to rise in the ranks of popular artists in the industry. The other interesting piece of this though, is that if you look at the superstar’s categories of artists, the Beyonces’, the Ariana Grande's, there are way more female artists represented in that kind of upper echelon of the industry in comparison to the kind of middle-tier and lower-tier artists.

Avriel [9:44]

And so what we really have to do is figure out how we can support female artists who are kind of like middle-class musicians, folks who are able to make a living making music through streaming, through playing at shows, through selling merch etc. Who are likely never going to become Beyonce but need to be supported by these streaming services because Beyonce needs streams but she doesn't really need streams, she's fine regardless if the algorithm is biassed against her or not. But these middle-class artists who rely on this to put food on the table, they're the ones who need to stream. And so we have to think really critically about how to best support them and I don't think there's a clear cut answer to that just yet.

Nikita RE•WORK [10:31]

How is it impacted by those factors?

Avriel [10:34]

So that positive relationship that I was talking about earlier, between such that, the more you stream female artists algorithmically, the more likely you are to stream female artists organically, that relationship is actually stronger for male users and it’s really interesting. Despite the fact that male users on average stream fewer female artists, so they kind of start out below female users, but then as, as the recommendation engines recommend more female artists, then their relationship is like very strong. So I think that's super interesting and then with respect to age, what we see is that the. I'm sorry, let me collect my thoughts. This is like the project that I'm deep into so I want to make sure as I haven’t spent a lot of time talking about this part of it. Anyway with respect to age, we see that younger users, users ages between the age of 12 to 17, so it's kind of a younger adolescent group, they do not have as strong of an effect in terms of what we've been talking about streaming algorithmically, leading to streaming organically. But older adolescents do. So the 18 to 24-year-old group, you see a pretty strong effect size there and then as users get older, the weaker that relationship becomes. And so one of the things that previous research has told us is that people's taste in music basically solidifies, around like in their 30s. And so it's less likely that you're going to see some kind of influence on what somebody wants to stream organically based on what they are recommended as they get older.

Avriel [12:40]

But the thinking behind, or my interpretation of this increase in influence ability, if you want to call it that, which it’s not, this is not a causal study. So I don't want to misrepresent this as like, if you do this, then this will happen. It's really correlational. But the one way to interpret this phenomenon with the 18 to 25-year-old group, or 24-year-old group is that they're in this kind of fluid identity stage where they are trying to figure out who they are, what their tastes in music are, maybe they're in college, or they have left home for the first time and they're really kind of exploring all of it or experimenting with different tastes and identities. And so I think you see that reflected in their kind of influence ability from these algorithmically recommended songs. One might assume that this young girl adolescent group might work similarly or might act similarly, but I think there's something to be said about younger adolescents really taking their cues, in terms of taste, not from personally tailored algorithms but more so from what pop culture is telling them is important, or what their friends in middle school or high is telling them.

Nikita RE•WORK [14:07]

That will also be a factor, I'm sure.

Avriel [14:10]

Exactly. But the gender of artists is such a high-level thing like nobody, very few people, I don't want to say nobody as I'm sure there are folks who are very deliberate in streaming female artists, but very few people decide whether or not they like a song or want to stream a song purely because it's a female artist. And so I think really what's happening here is it is just the strength of the recommendations generally influencing organic streaming down the line. And so if the recommendations are just as good, regardless of if it is a male or female artist, you're gonna see people responding to that and streaming the songs organically. And so, again, I think this makes the case that if you can keep the quality of the recommendations high while also controlling for gender, or constraining for gender parity, that I think seems like a really promising way to move forward to try to address this problem.

Nikita RE•WORK [15:28]

And another challenge that you spoke about in the paper was of gender labelling. Can you share a bit more about that and what specifically you found there?

Avriel [15:41]

I think this really speaks to a lot of the issues that we have across the industry in terms of bias because most machine learning experts or folks who are in this space will say, well, the biggest problem here is that there are social biases that are reflected in the data that these models are trained on. And sometimes you can't even get good data which was definitely the case in this study, getting reliable data for gender of artists is really difficult, especially in a platform where you've got millions of artists that are being streamed. For these higher tier artists, gender data is pretty reliable because there are all these kinds of industry databases that you can draw on that carefully track the credits for songs. And beyond that, again, you could google Beyonce and you can find out what her gender is, but an artist who recently uploaded their first song on to Spotify and maybe has a handful of streams, it's going to be really difficult to find their gender data, or the gender data for those types of artists at scale. And those are the artists that make up the majority of the artists on the platform, these kinds of middle and lower-tier artists. And so short of serving every single artist on the platform, it's just nearly impossible to get that data and get accurate data. And then the other thing is that even within these larger industry databases that track artists' gender, they still exist in this kind of regime of rigid gender binaries. And so non-binary artists, even if they do identify or identity as non-binary, that's usually not reflected in these large databases, just because that's, unfortunately, not how people collect this kind of demographic data. So it raises a lot of issues, it restricts our analysis and our way of understanding this problem, such that we can't even begin to think about the representation of non-binary artists and I think that is super problematic. And then, unfortunately, our research kind of ends up reinforcing the fact that this is an unaddressed problem in the industry, despite us wanting to find a way to address it, it's just really hard to get that data at scale.

Nikita RE •WORK [18:28]

I guess on that note, what would the next steps be for research in this area, and what would you next want to focus on?

Avriel [8:38]

I mean sadly my tenure at Spotify is over. So there are all kinds of things that I would love to do to kind of follow up on this. One is setting up some kind of experimental condition on the platform and see what happens if we expose folks to more female artists kind of in a very systematic way on purpose, rather than just looking at what is happening naturally on the platform, that's one thing. And then the other thing I think I'd be really interested in looking at it more deeply, at least, because we touched on this in the research, but it's not something we looked at super in-depth, is the role of the genre. And so if we can show that there is some kind of nudging effect here that you can get people to like more female artists and nudge those numbers up a little bit. That probably doesn't work across all genres, right? It might be super jarring for somebody who only listens to metal, given that there are virtually no female artists that are metal artists and I'm not a connoisseur of that genre so I hope somebody listening to this podcast emails me and says here are some great fans. But it might be super jarring for somebody who is used to listening to that genre to suddenly have a 50/50 split. And all of a sudden they're listening to female voices. Maybe, for some people that may be welcomed, and they might enjoy it. But I think something like that might backfire in a lot of situations. Such that people just don’t want their experience on the platform to be disrupted in a way that they aren't expecting, and it's just a kind of jarring.

Nikita RE•WORK [20:37]

And I guess, stepping a bit away from that research just for a moment. So prior to your academic career, you have, as I said before, quite a varied background. And you've worked in multiple areas, including storytelling and digital media, and many others I'm sure. But how has that background influenced your academic research? And has this at all?

Avriel [21:00]

Yeah so before I started grad school I was a musician myself, and gender and gender fluidity was a pretty big theme in my music. And I too had, despite getting or despite some relative success in a project that I put out, I really struggled in navigating the music industry. It's incredibly male-dominated. And particularly, I think, female artists of colour or non-male artists of colour, especially when they're vocal and kind of furtive then they're not super welcome. Folks want non-male artists to be packaged like products and they want to be able to figure out how to sell them and there's kind of this tried and true way of doing that. And if you're not willing to be commoditized in that way then it makes it really difficult to navigate the industry. I also, so I think that's one of the things that just made me interested in the research that we have just been talking about. But I think beyond that, identity is something, specifically how our identities are shaped by experiences online and interacting with technology is, I’ve got to say as a millennial, as a young millennial, my entire life was shaped by being on social media, and the rapid changes in technology across my lifetime so far and I don't know anybody who can say otherwise, really.

Avriel [22:58]

I often think about how I would feel about myself, and how would my racial identity have been shaped, or my views around gender have been shaped had I not been exposed to such a wide range of content online. And I also think about, what could have happened if I didn't end up in the rabbit hole of activists and folks who think Black Lives Matter is a good thing? What if I had ended up in the rabbit hole of folks calling people rioters and looters and the world kind of the white supremacists and folks who are people of colour who don't understand the movement and don't feel a part of it. You can see all of these kinds of simulated realities that you could have, or I could have experienced and I just find that so fascinating. It's kind of due to chance, it's kind of due to my own agency as a human being navigating these systems, it's kind of due to the folks who designed these systems and may or may not have taken my individual experience into account when doing it. I probably didn't take my individual experience into account when doing it. And that's just this terrifying and fascinating and wondrous world that I want to keep thinking about for, at least the next few years well into my career wherever I'm going after I graduate.

Nikita RE•WORK [24:43]

I think there’s no doubt everybody has individual things that influence us, especially through adolescence and 20s 30s, when things can change quite quickly. Especially with this current technology era, as you said, there is absolutely no way that that's not going to influence your work. And for you, in your case, you’ll be searched today I'm sure. And so quite a lot of our listeners are coming into AI or quite new to AI and looking to really start their career. What advice would you give for anybody that's currently listening today and is thinking that they'd like to get started in the field of AI, or they just recently started not really sure how to progress in their career, or what to focus on even as well?

Avriel [25:29]

So I think I'd like to tailor my advice to other people who are like me, who are kind of underrepresented in this field. And a lot of times I hear the advice that like, you can get into this field or this industry, and not necessarily be a technical person, which is true. There are all sorts of roles for non-engineers within organisations or teams that are working on AI problems or trying to solve problems with AI. But I want to push back on that a little bit. It's hard to learn the technical skills, it's hard to go back into the recesses of your mind if you haven't taken calculus, or like matrix algebra, or whatever, for a long time. But I would say just kind of push ahead and pick up the things that you need as you go along. And figure out how to actually be somebody who builds these technologies, even if you may not have originally seen yourself as the person doing the building, because we need the people who don't automatically see themselves as the builders to be the builders. The fact that those folks are not doing the work right now is one of the big reasons why all of these issues are popping up around. And then the other piece of advice I would have, or I would say is, before you build something, think about if it really needs to be built. A lot of times we build technologies because they're fun, they're interesting, which are perfectly valid reasons to want to do something. But in many cases they end up doing much more harm than good. And so, I'm less interested in building facial recognition technology that eventually will be used for mass surveillance or for criminalising the folks whose faces don't get easily recognised by the facial recognition technology, then I am in figuring out how to use AI to solve a number of the social issues, or social justice issues that exist in the world. And so I know that those are not always the projects that get funded and they're not always the projects that get fond over by investors and things like that, but I think we need folks to really start thinking deeply about what actually needs to be built. And in doing that, even those tools end up probably being much more complex and difficult.

Nikita RE•WORK [28:42]

Well, I think those both of those answers will resonate with our listeners, a lot of our audience are data scientists and engineers, so the first answer will work well for them. And then also, we get a huge amount of questions about AI for Social Impact. And it comes up on every single podcast that we do, so it has a perfect ending. So Avriel, for any of our listeners that want to connect with you or to get in touch, what would be the best way for them to do that?

Avriel [29:14]

I'm reluctantly on social media. So you can find me on Twitter and Instagram.

Nikita RE•WORK [29:20]

We can list your social media accounts in the podcast notes so that people can see it when they check out the recording. But thank you so much for taking the time to chat about your work today. Some really interesting topics and I think, as I said from the start, the recent paper that you published, we could have had a whole conversation just focused on that and your previous research. So this is just a very short snippet of some of the work that you're doing. But I'd encourage anyone listeners to get in touch, if they're interested in any of their topics that we've explored. And thanks again. It's been a pleasure speaking to you today.

Avriel [29:58]

Thank you for having me.

Nikita RE•WORK

A huge thank you to Avriel for taking the time to chat with us for this week's episode. If you're keen to learn more about Avriel's work, you can check out our research in the links below. It was really insightful to hear more about the important findings of her current research, as well as how her background has influenced her professional choices today. Check out our website for upcoming Women in AI events for 2021. Until next time, take care.

Are you interested in reading more AI content from RE•WORK and our AI experts? Read these articles below: