During the fall semester, Yoel Roth '11 returned to campus to deliver a lecture titled "Swiping Left: Identity, Preference, and the Politics of Online Dating." In the lecture, Roth gives an overview of the development of online dating sites and explores notions of privacy and self-expression with a particular focus on the case studies of Tinder and Grindr.

Roth earned his B.A. from Swarthmore College with honors in political science. He received his Ph.D. from the Annenberg School for Communication at the University of Pennsylvania. His research focuses on the privacy and self-expression choices of gay men using geosocial networking services like Grindr. He currently works on the Trust & Safety team at Twitter, developing policies and products to promote the privacy and safety of Twitter users worldwide.

---

Audio Transcript

Patty White:  Good evening. Hi, I'm Patty White from the Department of Film and Media Studies, and also Gender and Sexuality Studies, and I'm delighted to welcome you on behalf of both of those programs and the Tri-Code Digital Humanities, as well as the Department of History, and there's many fans of Yoel so everyone came together to pull this off.

Yoel, as you saw from the poster, is a graduate of Swarthmore College. He graduated in 2011, and we were talking about things have changed since then. I'm sure he'd be happy to talk with you about that. When he was here, he was at one point an editor of The Phoenix. He graduated with highest honors in Political Science and a minor in Film and Media Studies, and Yoel was, for me, one of my first introductions to a really robust kind of media studies research profile. He was really interested in media studies in a way that he realized in his graduate work at the Annenberg School, which is a communications school, and it's not really how we teach media studies here so I learned a lot. I remember we had an independent study and it was at least as much me learning from him. He was also an early adopter of all technologies, so everything I know about Twitter I learned from him in 140 characters.

Yoel went to Annenberg School, as I said, and his dissertation was called Gay Data and it focused on the privacy and self-expression choices of gay men using geo-social networking services. He's published several articles from that research. "Zero Feet Away: The Digital Geography of Gay Social Media," "No Overly Suggestive Photos of Any Kind: Content Management Policies and the Policing of Self in Gay Digital Communities," and forthcoming work that my students have a sneak preview of and we'll talk about tomorrow with Yoel in class, "No Fats, No Femmes, No Privacy," forthcoming in Digital Media 2: Transformations in Human Communication. Yoel has also worked with researchers at Harvard University Berkman Center for Internet and Society to develop protocols for studying dangerous speech and counter-speech online.

After completing his PhD in four years ... I'm proud. I don't have any right to be, but I'm very proud you can see. Yoel is a program manager now on the Product Trust Team in Trust and Safety at Twitter. He develops policies and product guidance for teams at Twitter focusing on issues ranging from user protection and privacy to content management and abuse prevention. Please join me in welcoming Yoel to speak on this topic.

Yoel Roth '11: Thanks again, Patty, for the intro and for welcoming me here, and thanks to all of the programs at Swarthmore and the Tri-Code Digital Humanities Group that were responsible for bringing me out here.

Despite Patty's intro, I wanted to spend just a second maybe trying to give you and overview of who I am and what I'm doing here and how I ended up here. As Patty mentioned, I graduated from Swarthmore in 2011. By the sounds of it, I missed a whole bunch of drama in the years immediately thereafter. There was this spring of y'all's discontent, and so, I was not discontented at all in my time here. I then went on and got my PhD from the Annenberg School focusing on geo-social gay networking apps, which I'll talk about a little bit tonight, and I'm now working at Twitter in trust and safety. Hey Bob.

My work at Twitter right now is focused in a couple of different areas, and I'm happy to talk at length about them by email or after the talk, You can reach out whenever, but the general overview of what I've been working on at Twitter for the last year and a half are, first, I work with product managers who are developing new products at Twitter to make sure we're doing them in a way that's responsible and safe. Whenever you're launching a feature that's going to affect hundreds of millions of people, you want to make sure that you're thinking about not just the awesome stuff that you're shipping out to the world but also the ways that people can misuse it, and that can be anything from abuse and harassment, which are huge topics, to child sexual exploitation, the spread of child pornography online. These are all things that we have to think through whenever we're shipping any new product at Twitter. That's one of my major areas of focus.

The other issue that I've been thinking about for the last year and a half is how to protect Twitter's data when other people are using it, so I'm responsible for our policies around our APIs and data ecosystem. You may have seen some news recently about police departments using Twitter data to track protesters, and in a nutshell, it's my job to make sure that doesn't happen, and so I'm fighting that fight at work every single data. I'm happy to answer questions about that at the end or talk about it in some more depth.

Then finally, my confession, I've actually never used Tinder myself, so I got this entire PhD in online dating before Tinder really happened. At my dissertation defense people were like, "What about Tinder?" and I was like, "I have absolutely no clue," and so, confession, even though a lot of my talk tonight is going to be about Tinder, even though I talk about this device of swiping left, I'm doing so based on mostly second-hand experience with it. Please correct me if I say anything that's completely and utterly wrong.

To try to set the stage a little bit for what I'm going to be talking about, I want to start with another one of the questions I got at my dissertation defense, which was, "Why online dating? Why does this even matter? Why do we care?" I want to talk a little bit about why I think this matters now in particular. The first is, there's an ongoing massive perception shift about online dating. In 2005, about 32% of individuals in a survey conducted by Pew Research agreed with the statement, "People who use online dating are desperate." Now it's down to 21% in 2013, but, I'm just curious. Raise your hand if you think people who use online dating are desperate. Don't be shy. One, so that's actually kind of surprising because the number is still holding stable in the low twenties, but like, in general I think there has been this massive perception shift. Anecdotally, my first year of graduate school, I was on a date with somebody who I met online, and over dinner he said, "Look, if this goes anywhere, we're going to have to tell people that we met through friends or at a party or something," and I was like, "What do you think I'm getting a PhD in?" Like, this is, and it didn't work out, and so I think that there is sort of this moment that's like, epochal shift in attitudes about online dating, so that's why it matters now.

There's also been a tremendous shift in usage. About 15% of adults in the United States say thar they've used online dating sites or apps, and 22% of 18 to 24-year-olds have said that they used one of these services. Let me ask another question. Again, raise your hand if you feel comfortable, but how many of you have ever used an online dating app or service or website? Right, so that bears out this trend, right, especially among young people, there's a tremendous shift in usage.

The third piece of this is that I think that online dating services actually give us a really valuable lens, not just into the places where we talk about politics or share pictures of our dogs and cats, or you know, look at what our grandma is doing, but sort of the sites of the most personal and intimate exchange of information that we have online. The information that we share in the context of online dating is both quantitatively and qualitatively very different from what we put on Facebook and Twitter, and so I think it's important to unpack that and understand why that's the case.

Finally, I want to make the argument that we can learn a lot from online dating about the shape of the web more generally. I think that online dating indexes broader trends in how online identity construction, privacy, and safety have evolved, sometimes years before it happens at scale on the rest of the web.

Tonight, I'm going to focus on two case studies, Grindr and Tinder. Each of these apps are major dominant players in the mobile online dating space, and each of them have millions of daily active users. Each of them have become, in the few short years since their launch, major cultural touchstones. Using these two apps as points of reference, I'm going to try to sketch out kind of like a brief, very selective historiography of online data, and my goal with this is to show that there is an essential continuity here around the types of questions we're grappling with vis-a-vis online dating. How we express our identities and our bodies and our preferences, if or why we feel unsafe or safe expressing certain things, and how structures of institutional and political power play a part in these processes.

I also, in sketching out this history, want to stress the discontinuities, the big moments of transition where I think something really substantial happened, and something that's changing the dynamics of how we express ourselves and connect with other people. Here, I'm going to try to map out two points of inflection, one exemplified by Grindr, one exemplified by Tinder. The first is a move from open-ended, unstructured interactions to high-bandwidth but controlled interactions in the early 2000s, and the second is the rise of personalization, prediction, and what I'm calling online dating, algorithmic data, rather, a trend that I think has only really begun to unfold in the last two years. It's this more recent shift, the thing that's happened in the last two years, that I'm going to spend most of my time on, but my goal here is not to vilify it or try to portray it as something terrible, but to explain why I think online dating looks more and more like the rest of the web and why that might not be a good thing. In the end, I think I'm going to try to make the case for why online dating and this most recent election can teach each other, we can sort of learn from one and apply it to the other and vice versa, and maybe I'll test out an idea for what we can do and you can tell me if you think it's completely bullshit or not.

We're going to go from there, but before I do that, I want to start with a pre-history of online dating. You've probably never seen this thing, but the computer on the screen is a pre world wide web French platform called Minitel, which was popular in the mid-1980s. Among other features, the pay-per-message Minitel played host to a wide range of gay and lesbian services, including discussion boards, listings of gay-friendly businesses, and erotic personal ads known as pink messages. This was pre-internet, pre world wide web, pre-everything, and yet, these services, these dynamics, already existed. Around the same time in English-speaking countries, the Usenet group Social dot MOTSS, or Members of the Same Sex, became popular as the first iteration of what researchers have called cyber queer communities, which is sort of a dated-sounding, I mean, it's cyber Monday so it can be cyber everything, which is like a dated term but ultimately it's describing in the sort of open-ended fashion the types of interactions that happened on MOTSS and other discussion boards, bulletin boards, chat rooms, things that emerged in the mid to late eighties and thereafter.

These groups developed their own shorthand for identity construction, and the exemplar of that shorthand are the three letters A, S, L, posed as a question. A is age, S is sex, L is location, and in these early days of online interaction, the hypothesis, the practice, in fact, was that using these three letters and the answer to them, you could map out who you're talking to and why you want to talk to them and what's important about this person. They tell you at least the beginnings of everything that you need to know, but what's kind of interesting about this history is that ASL and this text-based online expression that emerged in the late eighties didn't change much for almost 15 years. Have any of you ever used this, like maybe I'm dating myself. Yeah, not as, that's, oh man. This is AOL Instant Messenger, it's running on Windows XP, this was like the rage of 2001, but interestingly, the structure of those conversations is roughly the same, if you look at it, as it was in the late 1980s. IM exchanges on ICQ, I bet very few of you know what that was, AOL Instant Messenger, MSN, looked basically the same in 2001 as they did in 1991.

In the early 2000s, things started to change. The key factor in this brief and extremely selective history of the internet was the rise in popularity of broadband, particularly in the United States and Western Europe. Fast internet connections that were capable of supporting rich media went from something that was only available at universities to something that not everyone, but many people could afford to have in their own homes. I remember the moment when I got my first cable modem. I remember that moment of thinking, "Now I have the internet as it's supposed to be." This was like, the world wide web counterpart to the PC revolution of the early nineties. The legacy dial-up internet providers, AOL, Compuserve, Prodigy, and so on, suddenly started looking like dinosaurs, and this new open media-rich web, let's call it Web 1.0, brought with it a whole host of new websites and formats for interaction. Specifically, this unfolded first in the arena of dating and sexual relationships.

The screenshot you see here is from Manhunt, this is actually a fairly recent screenshot but it's what this website looked like for years. Manhunt, along with other portals like Gaydar and Gay.com, rose to immense popularity in the early to mid 2000s in the gay world by offering users something radically new. They offered them visual profiles, heavy on images. Everybody loved the opportunity to exchange dick pics easily, and then they scrapped this sort of really labor-intensive text-heavy version of identity that was ASL and replaced it with a set of structured drop-down menus. Instead of needing to creatively author yourself anew, you could just fill out a form, and these forms are exhaustive. You can see it gets quite detailed.

Then in 2008, something else really pivotal happened. Apple launched the app store on iPhone, and a year later the first version of Grindr was released. With no further ado, this is Grindr. You can see in Grindr that the main interface in the second screenshot from the left is a grid of photos ordered by proximity. That in itself was a radical technological development. The notion that you can integrate multiple data points, including what people have shared about themselves and also automatically gathered location data from somebody's phone, was pioneered first on Grindr. These are major technological shifts that are happening first in the context of online dating. The Grindr cascade, as the service's developers call it, gives you a proximity-ordered list of profiles, image-forward, there's not too much text as you can see, and you can scroll through the several hundred guys who are nearest to you. The individual's profile is, again, media and data forward. There's not too much text. We've moved far from the authorship of a chat room, where you really need to describe yourself, and instead you get a few sentences, maybe 140 characters. You get some stats, you get data, you get somebody's height and weight. Different apps have different ways of handling this, some ask you how hairy you are, and then of course you have chats and one-on-one interactions which is where all of the good stuff happens. That's the general structure of Grindr.

I want you to try to bear these images in mind, because I'm going to talk a little bit about why I think they represent a very particular image of what gay community looks like. Think about who's missing from these illustrations as well as who's depicted. I'll sort of leave it at that, but it's worth noting that these apps have been tremendously successful. I'm focusing here on Grindr because it was my focus in my dissertation, but it's worth noting that there were a few other apps, all of which share roughly these same properties. The first is, Grindr has two million daily active users in 196 countries, and the 196 countries thing includes some perhaps unexpected locations. There's a population of Grindr users in Russia, in Iran, places that are not regarded as typically receptive to gay community, and that's both, in my view, a benefit but also a security risk and a hazard to users, and I can talk about that more later on if you're interested, but two million daily active users is a massive community by any standard.

Those users are also tremendously engaged with these apps. If you've ever seen one of your friends on Grindr you'll see that they're constantly on it under the table at [inaudible 00:17:12] and you kind of see that orange glow, like, somebody's probably on Grindr right now. It happens, and users are on these apps constantly. Nine log-ins a day and 54 minutes in-app per person per day, on average, so there are people who are literally spending all of their time in these apps. There's probably people at the other end of the spectrum, but on average, this is tremendous engagement, and people exchange a ton of content. 70 million messages and five million photos per day, and that was in 2015. Those are the most recent numbers Grindr has made available publicly, but you can imagine that that usage has only grown.

This is how Grindr describes itself. "Grindr has become a fundamental part of users' daily lives across the globe. Grindr has supplanted the gay bar as the best way for gay men to meet the right person at the right time in the right place." I think this is telling of the aspirations of Grindr. This is an app that's positioning itself as a transformative form of what gay sex and gay community and gay sociability look like. Whether or not it's actually achieved those goals, I think it's telling that this app is setting that as the scope of it's purpose, right. That's what they want to accomplish with Grindr, and that in itself is a tremendous statement.

 Signed up for Grindr, interestingly, looks a lot like the Web 1.0 portals of previous days. For all the sophistication of our iPhones, things haven't changed all that much in the last 10 years. For all of the visual polish of this application, its underlying approach to user data still requires individuals to select and disclose particular pieces of information about themselves, ad hoc, with whatever measure of honestly or dissemblance they see fit. You have the ability to upload your own photo, you can pick a name, you can write some stuff, you can share your age, maybe it's your real one, maybe everyone is always 22, and then of course you can associate yourself with different tribes.

On the point of profile photos, I usually spend most of my talk talking about content management, and why you can't put dick pics on Grindr. The really cumbersome paper title Patty brought up, "No Overly Suggestive Images," like, it should really be called, "Where Have All the Dick Pics Gone?" because they have vanished, and I think that there's some really interesting reasons for why that's the case and it's not always intuitive why that happened on Grindr when it didn't happen on these Web 1.0 sites, but in the interest of time and in the interest of focusing on Tinder, I'm actually going to move on from that, but the point is, these apps are still asking a lot of their users. They're asking them to specify a tremendous amount of information about what they're looking for, what their body is like, and what groups that they affiliate themselves with. There are even tools that are not maintained by Grindr but are used by many of their users that propose to define you automatically, so if you're confused by this array of terms here, bear, daddy, discreet, geek, [inaudible 00:20:09], what does it mean? Don't worry, there's an app for that, and so you can enter information about your body and be automatically assigned a new label of your choosing.

These websites also then offer you analytics about what types of guys you're likely to find in different cities, so you can plan your moves based on who you think you're into, and they'll also predict who you are likely to be interested in on the basis of your tribal affiliation. This kind of feature has even started rolling out to apps. Scruff, for example, which is another gay-targeted app, offers real-time analytics to its users about the type of interest that they generally express, on the order of, "You reply to 65% of very hairy 29 to 35-year-olds who message you," and if you reply to another one that percentage goes up in real time. You get live feedback about your preferences narrated back to you by an application in the moment. It's really a brave new world.

My term for this, and the term that Grindr itself uses in their app is tribalism, and whether it's through quantifiable attributes like how old you are or how tall you are or how much you weigh or how much hair you've got on your back, or more nebulous labels like bear, and geek, or twink, whatever, apps like Grindr are active participants in a process of subdividing gay sociability into ever-smaller units.

This kind of highly structured expression of preference, tribalism even, if you want to use my term, isn't new. Gay community in particular has never made a secret of its subdivisions, and they've in fact sort of had a really pivotal part in the formation of gay community, particularly in the United States. You can imagine that bars cater to different types of patrons. If you've ever been out in Philly you'll notice there is a very different crowd at Woody's than you'll see at The Bike Stop, and in the 1970s, what I have pictured on screen here was sort of this defining symbol of US gay culture, the hanky code. The hanky code is a series of colored handkerchiefs that you can use to express the intimate details of what you're looking for and from whom you're looking for it and when you want it to happen, which invariably is right now.

Of course, there's a visuals culture, sort of a style here, and I always kind of go back to Dick Hebdidge's work on post-war British punks and skinheads to think through why gay men in the US in the seventies adopted the sort of trade aesthetic of Levi's jeans and hankies in the back pocket to express how they want to get fucked. There's something more than just group formation or sort of the crass expression of sexual preference here. The hanky code doesn't just advertise that you're gay. It provides an immense amount of data about what you're interested in and what you're looking for, and I view the hanky code as a direct ancestor of those kinds of drop-down menus that we saw in Web 1.0 portals, and that we continue to see reflected in gay apps today, especially as we talk about Tinder which works totally differently. It's worth bearing in mind that this kind of highly selective, careful, considered presentation of self is something that's unique to the gay world. Straight apps work different. Bear the hanky and the profiles that came from it in mind as we move on.

Grindr does add something genuinely new. It's not just a digital hanky code. Grindr and every other similar app on the market gives users the ability to filter people out on the basis of the data they share in their profiles, rendering entire groups of people completely invisible. The stakes for Grindr's claim to supplant the gay bar, in this regard, are huge. In a gay bar, you don't have the option to just disappear people you don't want to look at, if you don't like what color handkerchief they're wearing. On Grindr you do have that option, and that option is presented to you and you are encouraged to use it to make your experience better. This is a sword that cuts both ways. For the person doing the filtering, I actually personally don't believe that it's always in your best interest to only be exposed to things that you already believe you're going to like. Maybe you have a well-established sense of self and you know your type, but in my own experience and probably in many of yours, we know that romantic types change over time. If you're only ever seeing people that corresponded with the type you expressed when you signed up for Grindr, the likelihood that you're going to have those kinds of unexpected encounters that lead to a shaping of self is substantially diminished, but I also want to argue that those filters have a profound effect on the people they filter out.

Racism, ageism, fat-phobia, and a host of other prejudices are all too common on gay apps. Many of the non-white Grindr users that I spoke with when I was doing my PhD research noted that they felt like they didn't get contacted as often as their white friends, or if they messaged other people, felt like they were being ignored. I don't think it's coincidental, going back to Grindr's own marketing images, that most of the men here are white, and all of them look like they're under 40 and athletic. This is an app saying who belongs on a service, and then gives users the tools by which to enforce that sense of belonging. Perhaps the users who don't belong could go to some other app, right, and there's a slew of them, right. Grindr, for instance, Scruff, rather, started as an app that was targeting older and hairier men. Growlr targets bears. Daddyhunt targets daddies, and so on, but that's not actually solving the tribalism problem, it's just relocating it elsewhere.

Then there's some more recent developments. Earlier this month, Grindr published a blog post about the latest addition to user profiles, an explicit disclosure of sero-status, and the last date on which a user was tested for HIV. I see this move, which for the record, is not unprecedented in the gay dating world, as sort of a mixed bag. On one hand, I think it's absolutely a good idea to prompt frank and explicit conversations about safer sex. On the other hand, I think there's a definite, and this is the technical term that we use at Twitter, ick-factor to collecting that kind of information from users. Even voluntarily, particularly when this is an app whose revenue model is based on serving highly targeted advertisements.

Will Grindr, for instance, once it has this information about its users, let advertisers target ads only to HIV-positive people? What if Grindr tried to predict what users are likely to engage in risky sexual behavior? Academic studies have already shown this is totally possible, and then targeted those people with a PSA? What if employers or insurers or random people tried to use this information to identify somebody on Grindr and connect them with their presence offline, whether or not they've actually publicly shared their sero-status in other contexts? None of these are far-fetched possibilities. None of this is fear-mongering. Every single one of these things has already happened with other pieces of Grindr data, and every single one of them could happen with HIV status. Especially now that we're in a world in which some of the major protections introduced under the Affordable Care Act for HIV-positive individuals are looking like they're more and more in question, we need to think about what the consequences are for this kind of disclosure.

Of course, Grindr is also asking its users whether they'd be interested in filtering out HIV-positive individuals. This practice is called sero-sorting, and it's something that Grindr is now, they've done it yet, but they ran a survey, this has all happened in the last couple months. They ran a survey asking users whether they'd want that to be an affordance of the application itself. The point here, and I want to avoid getting into the politics of sero-sorting, is that apps like Grindr are offering these users, in a really easy-to-engage-with package, a level of sophistication for evaluating other people that is unprecedented in the history of networked interaction. These changes roll out so quickly and feel so seamless and easy that we only rarely stop to ask whether we should make them. Whether we should think of people and sex and relationships this way, and by the time we do ask, the next version of the app is already out in the app store and we're already using it, and there's a whole new set of changes to grapple with.

Ye, even though the ink on my dissertation only dried in 2015, even though I think Grindr and Scruff really are the state of the art in gay social networking apps, I want to try to look ahead and think about what's next, and I think that's happening on Tinder. Unlike Grindr, or OkCupid, or Match, or most dating services that had existed, Tinder doesn't give you a menu of choice. Tinder looks nothing like the Grindr cascade. Instead, Tinder shows you one profile at a time, allowing the user to move sequentially through a list of nearby profiles. If both users swipe right on each other, congrats, you matched. You're able to kick off a conversation. If one or the other swipes left, end of the road, there's no match, no option to chat. That's the basic mechanic, and yet, it's been explosively popular. Since Tinder first launched in 2012, it's taken off tremendously quickly. By 2014, the service first crossed the one billion swipes per day mark, and while the service's owners, Interactive Corp, who also own Match.com and OkCupid, have generally resisted sharing granular data about how many people actually use Tinder, by most measures it's been tremendously successful. Some estimates peg the app as having 50 million active users. Those users have 1.4 billion swipes per day, it's a ton of swipes, resulting in 26 million matches with each other, or 10 billion total matches since the app's inception.

Its recent success aside, I actually want to think a little bit about the history of Tinder, because I think it's an interesting one. Online dating is a really crowded field, but few apps or services ever make the kind of impact that we've seen from Grindr or Tinder, and so I think it's worth thinking about why these apps succeeded where so many others failed. In the case of Tinder, I'd actually argue that it's mostly the responsibility of one person, who receives next to none of the credit that she deserves. Whitney Wolfe was one of the co-founders of Tinder, along with six men. I would argue that she was substantially responsible in her role as Tinder's VP of Marketing for the app becoming the massive success it is today, and yet, after a series of extremely public accusations of sexual harassment, she was forced out of the company and in typical Silicon Valley style, the history of the app has largely been re-written to tell a heroic story of its male founders. I want to talk about two things that Wolfe did in her time at Tinder. She recognized early on in the app's history that its success or failure hinged on two related factors. The first is network size, the second is safety.

The first problem, network size, is expressed in a pretty straightforward idea called Metcalfe's Law that for whatever reason is always illustrated with either rotary phones or fax machines. The basic hypothesis of Metcalfe's Law is that the value of a network is proportional to the square of its number of participants, or put more straightforwardly, if I have a fax machine and you have a fax machine, it's not really useful if all we do is fax each other back and forth. That's fun for a while but it doesn't get anything done, but if everyone has a fax machine, then it still sucks, because they're fax machines, but, now lots of people have them and we can exchange useful ... How many of you have ever sent a fax? That is way more than I would've expected. It's horrible. I had to learn when I was in a fight with my insurance company. It's the worst, but the reason why fax machines have any value at all is because they have overcome the Metcalfe's Law problem, right. They have a sufficient network size that if you need to fax something to your insurance company, there's a fax machine on the other end of it. At this point they're like the only ones who still have fax machines.

The same thing is absolutely true of social networks. If no one you know or no one you want to know is using a given app, there's no reason for you to waste your time doing that either. It's part of why it's so difficult to challenge entrenched industry players like Facebook. When you have 1.5 billion users like Facebook does, you're leveraging Metcalfe's Law at a scale that's hard for anyone else to replicate or compete with. Like, raise your hand if you remember Ello. Exactly. They tried and they failed because they couldn't overcome the problem of scale. In the case of Tinder, Wolfe's answer to Metcalfe's Law was incredibly elegant. Shortly after the app launched in 2012, Wolfe went on the road. I'm quoting here from a Bloomberg story about her role at the app. She would go to chapters of her sorority, do her presentation, and have all the girls at the meetings install the app. Then she'd go to the corresponding brother fraternity. They'd open the app and see all these cute girls they knew. Before her trip, the app had 5000 users. After, it has 15,000 and the rest was history.

One of the things that made Wolfe's pitch more palatable is how easy it is to sign up for Tinder. Unlike virtually every other dating app or service, Tinder decided not to require or even allow users to construct their profiles with arbitrary information. Instead, Tinder uses Facebook data to automatically populate profiles with photos, interests, and so on. We see "Log-in with Facebook" buttons every day all over the web, but the underlying logic is always going to be the same. Instead of giving users the ability to construct new identities tailor-made to a given website, Facebook offers access via an API, or application programming interface, to the underlying pieces of a user's profile. Those pieces can be tweaked or reconstituted into whatever form of identity a given website needs. This makes signing up dead simple. Unlike Grindr's endless array of boxes to fill out, Tinder gives you essentially one-click sign-up, and when you're dealing with Metcalfe's Law as a new app, this is an unimaginable benefit.

I think Facebook authentication has a second, and more fundamental purpose. This New Yorker cartoon from 1993 is a total cliché amongst academics who study the internet, but like most clichés, I think it reflects a basic truth about the thing that it describes, namely, that in the early days of the internet, people saw online identity as something completely unmoored from its offline counterpart. Everyone was six foot two and had a six-pack in the internet in 1993, and dating services were particularly vulnerable to this critique. There are dozens of academic articles discussing the perils of fake profile photos, and we even now have the term cat-fishing to describe this practice. Tinder took, in my view, sort of two ingenious approaches to solving this problem. The first was Wolfe's college tour, and specifically the fact that it was a woman who was pitching this app to other women with whom she had a pre-existing connection of being in the same sorority. She made prospective users feel like they could trust this new app that they were trying out. Wolfe was there to walk them through setting it up, and when they logged on, because she had already gone next door to a fraternity, the people they saw in this app weren't anonymous creeps lying about themselves. They were the creeps that they knew from the fraternity next door. These were already familiar people from the same university.

Second, and this is where Facebook comes into it, using "Log-in with Facebook" keeps Tinder profiles firmly grounded in information that's already available on another, more elaborate, more grounded service. Tinder profiles are able to sort of leverage the barriers to entry associated with Facebook. It's kind of hard to create a fake Facebook profile unless you're really trying to do so, and so when you sign up for Tinder you're required to use that information that you've already shared with Facebook, potentially over years, and it credentials what you're doing on Tinder with the weight of your Facebook presence.

The interesting assumption here is that our Facebook profiles are somehow more representative of our true selves than the identities we construct on particular apps or websites. I think there's good reasons to be skeptical of that assumption, or any claim about one website's ability to be the arbiter of true identity online. We filter and reconstruct and manipulate ourselves online all the time. Alice Marwick and Danah Boyd have talked about the imagined audience that we have in our heads as we tweet or post on Facebook, but cross-platform identity as it's enacted in Facebook's log-in API suggests that whether it's for convenience or some other reason, many of us do have a canonical self online, or at least a self that we treat as canonical, and we use that self all over the web. Whether or not it's true to who we are face-to-face, it's our presence online, and the Facebook API gives any application developer access to that.

How do those properties of Tinder impact how we actually connect with each other? To go back to my brief history, the key technological developments of the early 2000s, broadband, the rise of social networking, and the popularization of smartphones and mobile app stores gave rise to apps like Grindr, that live firmly in a paradigm of browsing, searching, and filtering, but I think we're in another moment of transition now and I think Tinder is among the first apps to take advantage of it. I would argue that we are entering an era of personalization and prediction in online dating. Let's start with a pretty straightforward example, a feature Tinder calls Smart Photos, which they claim will automatically pick the most swipe-worthy you. It's worth noting that OkCupid has actually had this feature for eight years ago, and they called it My Best Face, which is a terrible name, but it's still interesting to think through what this means.

The pitch is simple. You already have all of your photos in Tinder because they came there from Facebook, and by toggling one switch, Tinder will tell you immediately, "This is the best you. This is the photo to feature and show to everyone if you want to maximize the number of people who swipe right on you." The key to understanding features like Smart Photos and how they work, and content personalization more generally, is to recognize that every action you take on a website or app can be translated into an array of data points. I'm going to try to keep things fairly accessible in the course of this talk, so this is a little bit of a simplification, but it's not too far from what's actually happening on Tinder.

This is a left swipe turned into its constituent parts, or at least a simplified version of its constituent parts. You need to identify, in a left swipe, who you are, what you're looking at, and what you did, and so here you have your user ID, you have your target's user ID, the person you're looking at, and what you did. You said, like, "Nope, not interested." That's it. That's a left swipe in data, but oftentimes, more than what you just did is stored along with that swipe.

That's what's called meta-data, so maybe for a left swipe it's important to know when it happened, right, so have a timestamp. This happened today at 9:45 am, or maybe it matters how long you viewed that person's profile, or how many of their photos you looked at. Was this an impulsive left swipe, or did you really think about it before you rejected this person? Maybe it matters who your mutual friends are. Some friends maybe are stronger predictors of whether you're likely to connect with somebody, or maybe it matters in what order you viewed the profiles, right. Maybe you were on a hot streak, maybe you were on a not-so-hot streak, or maybe it's not just who you viewed before this given person but also how you voted on them. Maybe you liked a bunch of people in a row and so your standards are really high, or maybe you're in a bad mood and everybody is getting a left swipe. Data is cheap, and processing it is cheap, and there is absolutely no disincentive for software engineers to not store this data, because they can always use it later. They'll find a use after the fact, and so the goal with every one of these interactions with an app, and this is true of every app you use every day, is to collect as much information about your activity as possible and see what you can do with it later on.

In the case of Smart Photos, Tinder is in the background automatically using all of the meta-data it collects with every swipe to carry out an individual level AB test. This is the same structure of experiment that's used to launch every major feature on any social website. The idea, without getting too into the weeds of experimentation, is basically, "I serve Patty one experience, I serve Bob another, and I see how they behave." When you replicate that times a few million people, you can actually get statistically significant results about whether your feature, that change that you're testing out, led to different outcomes of behavior. Here, Tinder is trying out different photos as the primary image of your profile, and by aggregating enough responses, enough left or right swipes, it can correlate whether somebody swiped left or right on that photo and figure out which one has the most.

Simple counting, but given all the meta-data that Tinder has available, everything from the last slide and so much more, you could imagine a more sophisticated version. Eight years ago, OkCupid's My Best Face already offered these kinds of breakdowns. You could see how different pictures performed with different ages or political beliefs, or levels of education. As far as we know, Smart Photos just picks the one best photo for everyone, but Tinder already has all of the data it needs to carry out more granular analysis, and maybe for those people who are paying 15 bucks a month for Tinder Plus, it will. All of this points towards a world in which we're giving apps and websites increasing power to construct our identities for us. Grindr takes a bunch of raw information, puts it on the screen, and lets other users manipulate it. Tinder takes data you've already shared elsewhere, does the processing for you, and will even pick your best self to showcase to other users automatically.

What about how we connect with other people? Given a large pool of potential candidates, how can Tinder select which one to show you right now? I promise there's a science to it, and it's very different from Grindr's approach. Grindr emphasizes neutrality on the part of the app. The service shows you who's near you, it lets you pick the filters, and it just is there to connect you with other people. Tinder, on the other hand, is explicitly non-neutral. The sequence of profiles you see, and in fact whether or not you see somebody at all, is determined by the app in a way that is totally obfuscated from its users. Tinder accomplishes this task using a personalization algorithm, and without getting too far in the weeds, I'm going to try to parse out what that algorithm might look like.

The core of every algorithm is an action or outcome that you're interested in. In Tinder's case, that outcome might be you swiping right on a profile that you see, and so the basic assumption expressed here in the bit that says p-swipe, is that people are generally rational and consistent and that using available data, you can come up with a pretty good prediction of how somebody is likely to behave. This is not an uncontroversial assumption, right. People are irrational all the time, but it's a pervasive one, especially amongst tech companies, and so whether or not this assumption is true, Tinder is built on the belief that they can predict pretty reliably whether you're going to swipe left or right on a given profile before you even do it. A personalization algorithm is at its core just an attempt to maximize the predicted likelihood that you'll take a given action on a given piece of content. For Tinder, the objective is the find the profile in the pool of nearby candidates who you are most likely to swipe right on.

It's simple enough, but how do you come up with that probability? Tinder can't get in its users' heads, and it's too time-consuming or difficult to ask them a million questions to try to figure it out, which is what OkCupid and Match.com and Grindr do. Instead, their prediction algorithm has to work with the data they already have available, and so for you, the newest member of Tinder, the app starts with existing predictions it's already made about other people and how they're likely to behave when they're viewing the same profile. What does similar even mean, and how does Tinder figure that out? This is the easiest part of all. Thanks to the Facebook log-in API, Tinder already knows everything about you that's on your Facebook, and it knows everything about your friends who use the app, and the other people who are engaging with the service, and so it has all of that information to base those sorts of predictions on. By looking for people who share your similar beliefs or interests or education or age, Tinder can make reasonable assumptions about what other users are likely to be similar to you in how you behave. All of those users' swipes become data that Tinder can use to predict what you're going to do.

They can also look at other factors. For example, there's pretty well-substantiated rumors that Tinder maintains an internal attractiveness score for every one of its users, and so your algorithm could look at the difference between your attractiveness score and the person you're looking at to figure out whether you might behave a certain way on the basis of that information. There are other factors that you could use as well. Your distance from somebody, how often you log in, how many friends you have in common, and so on, but the key thing to remember here, especially with something like the attractiveness score, is that this is all based on Tinder's parsing of readily available data. From Facebook via the API, from your activity in the app, from how you've swiped before, from what your friends are doing, and from the activity of the people that you interact with. Tinder is not using facial recognition technology to carry out an automated game of Hot or Not with people. That's not happening, the technology is not there, it's not possible, but they can still make those kinds of predictions using things like a swipe billions of times a day. That's a ton of information that they already have available.

Of course, there's no requirement that swiping is the only thing Tinder is trying to optimize for. They could, and I would guess do, rank connections by how likely you are to reply to somebody, or maybe how likely you are to carry on a lengthy conversation with them that's not just, "Hey, what's up, I'm good, okay, bye, thanks." They might want to predict whether it's likely to be a substantial connection or a fleeting one, and maybe your matching function is a combination of all of these things and a bunch of other factors. Even the sophisticated multi-variate algorithm is substantially cheaper to do than the kind of really labor-intensive work of crafting a profile and predicting, like Match.com does, whether you're going to be an 80% match with somebody, and so why not?

Maybe your business model figures into this as well, and you offer users the ability to put a thumb on the scale and artificially inflate their position in the queue. Tinder launched this feature in September, and they call it Boost. The premise of Boost is pretty simple. You pick Tinder $2.99, or you subscribe to their Tinder Plus service and Boost will increase how often you show up for people near you for 30 minutes, and afterwards, Boost will even say, "You showed up 9% more often, you got X number of swipes that you otherwise wouldn't have received," but the interesting thing about Boost is that it's completely invisible to the people you have been boosted to. They have no idea that the match they were just presented with is because you paid $2.99 rather than because you are likely to be the love of this person's life, and so we see here how business models and algorithms inevitably intersect. Tinder has total control over the sequence of profiles you see or don't see, and it has no obligation to be transparent to you, the end-user, about why you saw what you did. They've already indicated that's what they're doing with boosted matches.

It's not just a question of whether Tinder sometimes shows you an artificially inflated profile, right, so that's somebody who's score has been increased. What if Tinder automatically served you only pretty mediocre matches, all day, every day, unless you subscribe to Tinder Plus? That's entirely their prerogative, they can make the service as good or as bad for whoever they want, whenever they want, with zero visibility to users. They hold all the cards. The good news is, I don't think that's what they're doing, and I'm going to explain why. I think there's good reasons to believe that that's not the case.

Given Tinder is not going to be releasing their personalization algorithm to the public anytime soon, nor is Facebook, and Twitter's not doing it either, the best we can do is try to deduce their information from the available public information. One of the things we do know a lot about is the financial and business structure of tech start-ups. Particularly in the case of social platforms, we know that the key incentive, the over-riding imperative here is to increase user retention. Put it another way, Tinder wants to keep you on Tinder. They need to. Their success as a business depends on their ability to do that. If you imagine the business model of a yenta, I would imagine it goes something like, you have successfully arranged a marriage and you get some kind of commission of a percentage of the dowry because weddings are like buying houses, or something. Tinder has no such motivation. It doesn't receive a part of your wedding dowry. There is nothing incentivizing Tinder to efficiently pair you up with somebody, because then you'll leave Tinder forever and they'll have nobody to serve ads to or sell $12.99 monthly subscriptions to.

On the other hand, they can't make you too unhappy, because then you're going to feel like this entire app is like a colossal waste of your time. Why bother at all, if all you get are garbage matches all the time? Tinder's project is to find a happy equilibrium, where you're happy enough to stay on Tinder but not so immediately happy that the first time you open this app, you fall in love and leave forever. This is the basic logic of how every social business works, and it's pervasive enough that I'm quite confident it's happening on Tinder as well. This is absolutely the same dynamic that incentivizes every other major social platform, including Facebook and including my own employer. We want to keep you using our service. Driving usage is the key factor behind every major decision that we're making, but we're not evil. This desire to get people to use services more is what created a push towards personalization in the first place. The thinking here goes, "There's some ideal selection and sequence of content that will keep you happy and using service for as long as possible. It'll keep you retained as a user," and this thinking is not made up. It's demonstrated experimentally over and over again. People absolutely love personalization.

In my time at Twitter, we did something really crazy. It was controversial. We launched a version of the home timeline, that list of tweets you're looking at, that's not just ordered by reverse chronology, so it doesn't just show you the thing that happened most recently at the top of your timeline. People absolutely lost their shit over this. There was the hashtag RIPTwitter, it was a giant crisis. People thought, like, we're going away, we're all losing our jobs, turn off the lights, we're done. What happened when we actually launched this feature, which has the pretty simple premise of putting a few tweets that are the best for you at the top of your timeline, is that people loved it. Tiny fractions of a percent of people actually turned the product off. We give you a way to opt out, that was a giant fight I had, but people didn't. It turns out, everybody loves seeing personalized content, because it means that as soon as you open the Twitter app, you see something really cool, and so people spent more time in the app, they tweeted more, they liked more, they retweeted more, and this happens with every single instance in which somebody introduces personalization. It's not a coincidence that your Facebook newsfeed is not just a reverse chronological stream of stuff the way it used to be. Facebook tested it out and they found the same thing we did. Personalization works.

Grindr has a serious problem in this regard. For all of the subtlety of the features it offers its users, those users still complain that they can't find what they're actually looking for. The authors of a 2014 study titled, quite evocatively, "Departing Glaces," interviewed a bunch of Grindr users who said they left the app, and asked them why they left it. Many of the respondents said that they felt the app had become a huge waste of time. They invested immense effort in crafting this profile and filling out all of these interminable fields, and they still didn't find what they were looking for, and so maybe they went to another app or maybe they gave up on online dating altogether. Maybe personalization would change that. Maybe it would solve the problem. Tinder's spectacular success in this industry suggests that it would.

The last lesson I have is maybe the most obvious. The word algorithm has become incredibly commonplace in the last few years. We've heard that everything we say and do and see on the internet has been decided for us by unseen algorithms created by unseen engineers in Silicon Valley and Mark Zuckerberg is like, a puppet master holding the strings, and I want to try and dispel or at least somewhat clarify that idea. Algorithms are just the mathematical expression of human behavior, or rather, they're an attempt to use past behavior to predict what people are going to do or like in the future. I tried to map some of that out with regards to the Tinder personalization algorithm, and obviously the way Tinder makes these decisions in practice is opaque to their users, but there's no malice or great mystery behind the underlying process. All Tinder can do is look at what you and other users have done in the past and create predictions about likely outcomes in the future. Even Tinder's desirability or attractiveness score is just a product of how other users have engaged with your profile, and everyone does this. There are more sophisticated algorithms that begin to feel truly uncanny.

A friend who worked at LinkedIn told me that their most closely guarded piece of data is a predictor of whether you're likely to leave your current job. This is obviously sensitive, but it's incredibly lucrative to make those kinds of predictions, and yet, all LinkedIn has to go on is the same kind of data that Tinder uses to personalize your matches. How often do you log on? How long do you spend on the app? Do you click on job descriptions? Which ones? How often? Are you doing it more than you used to? This is all really straightforward, simple data, but when you process it, you can come up with these really powerful, sometimes even creepy, predictions.

If these are problems you're interested in, there are, in my opinion, two truly great books that have been written about algorithms. Alex Halavais' "Search Engine Society," which focuses on the rise of Google, and Eli Pariser's "The Filter Bubble." They're both good places to start. I'm going to pause here to say that in the interest of time, I haven't given most of my citations in this talk. I know my graduate advisor is like, cringing as I don't cite any of my work, but if you want to learn more about anything that I've talked about here, please feel free to reach out to me on Twitter, by email, I'll have contact info at the end, and I'm happy to share reading lists, suggestions, any of that, if any of this is information that you want to dive into further.

I'm going to close by taking a minute to talk about the election. In the aftermath of what happened earlier this month, I think there's been a lot of public discussion about filter bubbles and fake news, and whether sites like Facebook and Twitter are somehow responsible for exposing us to the wrong information, or not enough information, in a way that hampers our ability to be informed participants in a democracy. These questions are completely reasonable, but in my opinion, they're the wrong ones to ask. They treat Facebook and Twitter like they're exercising some kind of human agency in choosing or not choosing to serve this content. Instead, I think we should focus on what the incentives are here, and why they've tended to result in certain outcomes.

The image on screen is a map of links between political blogs. It's a little confusing, but what it shows is that generally speaking, left-leaning blogs in blue link to each other, and right-leaning blogs, in red, link to each other, and only rarely do those two groups cross-pollinate. The study that this image is from is from 2005, looking at blog posts from the 2004 election, which puts it pretty firmly in what people might call Web 1.0, but the finding is an interesting one. It suggests that even when people make active human choices about what to link to on the web, they overwhelmingly tend to link to content that they think their readers are likely to agree with. In essence, they create a bubble, and it turns out that personalization algorithms are leading to basically the same result, but for a different reason. Human bloggers have a qualitative notion of whether somebody is likely to agree or disagree with something, or whether it fits into the narrative of the post they're writing. Facebook's newsfeed algorithm, on the other hand, is just trying to optimize to keep you on Facebook even longer, but the underlying cognitive process there is the same. We like seeing things that we agree with, and so the outcome of that is the same too. There's nothing sinister about a personalization algorithm. It's just a reflection of how we see and makes sense of the world.

Here's an interesting point from a study of exposure to political messages from 2006. It turns out that when you expose people to conflicting viewpoints, they're more likely to appreciate and tolerate their views. That's awesome, tolerance, great, but the downside is they also become less engaged. They're less likely to vote, or volunteer, or get in fights online, and so this is the dark side of the personalization algorithm problem. Facebook and Twitter and Tinder are all trying to keep you engaged as much as possible, and there's tons of evidence that suggests that the best way to do that is to show you things that you're already likely to enjoy. I don't think it's that anybody at Facebook is blithely unaware that there is a benefit to society by exposing people to cross-pollination, by giving them information they might not agree with, but it's not altogether clear how we should reconcile that with the problem that we do want people to remain engaged, and that too much exposure to conflicting information is likely to cause people to tune out rather than process it synthetically. Maybe the only solution is rewriting human nature, but that's not really in Facebook's purview.

I'll end with one piece of advice, and it might be one that my employer won't like. We should take active steps to confuse the personalization algorithms that surround us. Swipe right on somebody you're not immediately sure you'll be into, or that you think is the total opposite of your type. Follow somebody on Twitter that you completely disagree with. This sounds trite or contrarian or like it's going to be super annoying when you see, KellyAnne Conway for me, in my Twitter feed, and I'm just like, "Blah, really," but I'm of the belief that however nice it is to only be exposed to the things that you already agree with, it's not in our best interest to allow that to be the case. We need personalization algorithms to become a little less good at predicting what we're likely to be into, and I think that's even more so the case with our intimate relationships. I think the more we allow apps like Tinder to structure how we connect with each other, who we see, who we date, who we don't see, the more the stakes for these algorithms become embedded in every relationship we have in our lives.

Algorithms aren't going anywhere, and I don't think that the incentives of tech companies like Twitter and Facebook and Tinder are going to change either, but I would argue that by understanding where they came from and how they impact the relationships and the interactions that we have every day, we can think through what some strategies for coping might look like.

I want to again thank Patty and the Film and Media Studies Department for inviting me, as well as Tri-Code Digital Humanities and all of you for attending. I know that the Monday after Thanksgiving is like, "Oh, crap, I have six papers I haven't written and everything is awful and I have no time," so I really appreciate you coming out tonight and listening, and please reach out by email or Twitter or like, anything but LinkedIn because I don't understand how it works, and we can chat about work or grad school or life after Swarthmore, or anything. Thanks.