Finding Solutions: Choice, Control, and Content Policies
AMRITHA: Okay. Well, hi everyone. Thanks so much for joining us today. My name is Amritha Jayanti. I direct the Tech and Public Purpose Project, also referred to as TAPP here at Harvard's Belfer Center for Science and International Affairs. TAPP was founded in 2018 by former US Secretary of Defense Ash Carter, to work to ensure that emerging technologies are developed and managed in ways that serve the overall public good. We have a very exciting session today. The first in a new public events series titled “Finding Solutions,” in which we host practitioners across the public and private sectors, many of whom will be our very own Fellows, to dive into a myriad of potential solutions to address unintended consequences of emerging technology.
We're planning to talk honestly about problem definition, solutions, feasibility, and intentionality of impact across various solution types. So I'd like to quickly take a moment and thank Karen Ejiofor, TAPP's project manager, for all her work in ideating and planning this series. To follow the series, as well as other TAPP events and publications, we encourage you to follow us on Twitter at TAPP_Project, and sign up for a newsletter on our web page. And with that, let's kick things off for today. So we're hosting two awesome TAPP non-res. Fellows,
Karen Hao and Joaquin Quiñonero Candela, sorry, Joaquin, if I messed that up. But I'll give them a brief introduction and pass it over to them for a fireside chat style conversation, for the first 40 minutes. We'll save some time at the end for audience Q and A. So please use the Q and A feature in Zoom, not the chat box, but the Q and A feature, to submit any questions you may have for Karen and Joaquin. So with that, and I'll go ahead and introduce them. And if you all want to pop into the
screen, you can do that now. So Karen is the Senior AI Editor at MIT Tech Review, covering the fields, cutting-edge research and its impacts on society. She writes a weekly newsletter called “The Algorithm,” which was named one of the best newsletters on the internet in 2018 by the WEBBY Awards, and co-produces the podcast “In Machines We Trust,” which won a 2020 Front Page Award.
In March, 2021, she published a nine-month investigation into Facebook's responsible AI efforts, and the company's failure to prioritize studying and mitigating the way its algorithms amplify misinformation and extremism. Her findings were cited in a congressional hearing on misinformation two weeks later. In December of 2020, she also published a piece that shed light into Google's dismissal of its ethical AI co-lead, Timnit Gebru, which congressional members later cited in a missive to the company.
Prior to MIT Tech Review, she was a tech reporter and data scientist at Quartz. In a past life, she was also an application engineer at the first startup to spin out of Alphabet's X. She received her B.S. in mechanical engineering and a minor in energy studies from MIT. Okay. And Joaquin serves on the Board of Directors of The Partnership on AI, a nonprofit partnership of academic civil society, industry, and media organizations. Creating solutions, so that AI advances positive outcomes for people in society.
And is a member of the Spanish Government's Advisory Council for Artificial Intelligence. Until September, 2021, Joaquin was a distinguished technical lead for Responsible AI at Facebook, where he led the technical strategy for areas like fairness and inclusiveness, robustness, privacy, transparency, and accountability. Before this, he built and led an applied ML learning team at Facebook, driving product impact at scale, through applied research and machine learning, language, understanding, computer vision, computational photography, augmented reality, and other AI disciplines. AML also built the unified AI platform that powers all production applications of AI across the family at Facebook products. Prior to Facebook, he taught a new machine learning course at the University of Cambridge, worked at Microsoft Research, and conducted post-doctoral research at three institutions in Germany, including the Max Planck Institute for Biological Cybernetics. He received his PHD in 2004 from the Technical University of Denmark.
So our session today is going to focus on solutions involving social media recommendation systems, particularly some of the experiences that both Karen and Joaquin have, thinking about applications, as well as societal impacts. Just reading their bios, I'm sure you all know that we're in for a really amazing discussion. So thank you both, Karen and Joaquin, for joining us today. I'm really excited to see where this goes. So Karen, I'll pass it over to you, to get the conversation rolling. [00:06:36] KAREN: Awesome. Thank you so much, Amritha. Hi everyone. I am super excited to be here. Thank you to the Belfer Center for having both of us. Just for a little bit of background, Joaquin and I met when I started working on my story about Facebook. Joaquin was the former responsible AI lead there. And we spent quite a lot of
time talking together about some of the issues that we're going to talk about today. And I was very impressed, throughout my time speaking with him, about his thoughtfulness, and his really deep caring for these issues. And so I'm really pleased that we get this public forum to talk a little bit about some of the things that we've been talking about behind the scenes.
So Joaquin, obviously there's this huge ongoing debate that's happening today about social media recommender systems. And we're here to talk about it head-on today, and propose some solutions. But I first wanted to give you an opportunity to actually tell the audience a little bit about your background, and how you ended up at a place like Facebook. [00:07:44] JOAQUIN: Yeah, thank you, Karen. Hi everyone.
It's a pleasure to be here today. And thank you so much to the Belfer TAPP team for hosting us. I didn't think I would one day work at Facebook or work on the—at the intersection of machine learning and social media, if I'm honest. I think it happened a little bit like many things, by accident through connections, through good friends who had gone to work at Facebook back in 2011. But one piece of context that I think is really important, is that although I was born in Spain and raised in Spain until age three, my parents, my sister, and I moved to Morocco when I was three years old. And so I grew up in the Arab world, surrounded by
many friends of all kinds of origins, but many friends of Muslim and Arabic origin. And so when I visited Facebook in 2012, socially, with no idea that I would end up working there, the Arab Spring had been going on for about a year. And to me, it had caused a very, very profound impression, because countries like Egypt or Tunisia were seeing massive revolutions, in a way. And I had close friends living in these two countries. For example, even my sister ended up working in Tunisia for a while.
And so I was blown away by the power of tools and platforms like Twitter or Facebook, to help people communicate, mobilize, and really change society for the better. So it's in that context that I felt compelled to join Facebook. I felt the mission was simply incredible. KAREN: Yeah. I kind of want to bring people back to that particular time in social media, because in 2012 Facebook was quite young at that time. So could you like paint a little bit of a picture of
just what stage, I guess, the company was in, and when you came to join the company, what your professional background was, at that point, and what you were sort of tasked to do? [00:10:10] JOAQUIN: Yeah, absolutely. I was on a journey. As a professional background, I was on a journey from almost having been an academic in machine learning. I was a post-doc, had my constituent in [00:10:21] Germany doing pretty, pretty theoretical and abstract research. But in the intervening time, I had spent five and a half years at Microsoft Research in Cambridge in the UK, initially.
Also doing research But veering towards applications quite—quite rapidly. And, in fact, at Microsoft, together with some colleagues and now very good friends, we developed the click prediction and ranking algorithms that helped power ads on the Bing search engine. [00:10:55] So I had a bit of background, now, not only in the theory and research of machine learning, but also in its application at scale. And in fact, I did something considered a bit crazy at the time, which was, I accepted an offer to leave Microsoft Research and join a product team, and become an engineering leader. So I had a little bit of experience on both sides of the aisle, if you will. So that was the context.
And I joined Facebook, as Facebook was a pretty young company, like you said. It was growing extremely fast. I remember this, during the first months that I was at Facebook is when we crossed one billion users. And that's not even daily active users or anything like that. It was one billion users total. And a lot of things were in their infancy, you know. There was, of course,
a couple of pretty strong machine learning teams in feed ranking, but also in ads. But, when you compare the number of people working in ML, at Facebook or at Meta today, you know, to what it was when I joined in 2012, you could fit everybody in the room I'm sitting here, you know, who was working on ML. And I would know them all by name, right. [00:12:15] So I joined in. And the task that I took on was, well, let me build out the machine learning team for ads, at first. And then, very quickly,
I realized that we needed to invest extremely heavily in platforms, in tools. I used to say wizards don't scale. We need to be able to factor. We can't hire enough ML people to do all the work that needs to be done. We really need to take our tools and our platforms to the next level. And, to try and make a long story short, that led to the creation of the team that Amritha mentioned earlier, the Applied ML Team, which then had the scope to help bring ML to everybody across the company.
[00:13:03] KAREN: Yeah. So one of the things that, when we first started talking about your personal background, that I was very touched by, was the fact that the Arab Spring was quite personal for you. Because when you talk to people in social media, they often reference the Arab Spring. But it was very different for me hearing it from you,
because it was something that you were literally sort of seeing in your life, and through your own friends and your own family. And that was sort of the mission or the vision that you took on when you joined Facebook, of this is what social media could be. And obviously, to sort of skip a lot of things that happened along the way, social media didn't really quite turn out the way that it was originally envisioned, this grandiose ambition to connect people, to create these like powerful positive social movements without any of the costs. This past year, there has obviously been a lot of talk, now, in particular with Frances Haugen and the Facebook Papers, that are reexamining some of the core challenges with social media platforms, and why we might be seeing some of the adverse effects.
And Frances Haugen specifically pinpointed a number of the risks to recommender systems. So what are the risks that you see for social media recommender systems, in particular, as someone who, more than most people in the world, understands how it works, and how it was built? [00:14:35] JOAQUIN: Yeah, no, it's interesting. What you say is exactly right. It's been a really interesting journey for me. I remember the first few years at Facebook, my obsession was in scaling ML, and making sure that we can build bigger and better recommender systems, using larger models, being able to refresh those models as fast as we can, et cetera, and make them evermore powerful. And it is true that, along the way, a lot of issues did arise which I did not anticipate, and I think many people did not anticipate.
I don't claim to have a comprehensive overview of what the risks are of large-scale recommender systems, and in particular, social media recommender systems. But since our past conversation, and your article, a year ago, I've been thinking about this a little bit. And so I have maybe four buckets of concerns that I'd like to propose. And again, disclaimer, there may be more. And I'm super excited. I know that we have, at the Belfer Center, people who are thinking about these things very deeply as well. I don't know if Aviv Ovadya
has been able to join today. But I'm going to put him on the spot and embarrassing him. I've been reading his work with a lot of interest. And I know there's many other people as well. [00:16:05] So let me list out the four buckets. And then we go into them one by one. The first one would be mental health concerns. The second bucket would be bias and discrimination. And then buckets three and four
are similar but different. So bucket three would be the propagation of misinformation in particular. But just of harmful content in general. And then bucket four builds on this, but is much more specific on personalization, and the fact that, if you think about it, everybody has got their own unique personalized feed and newspaper. And so the risk in bucket four is polarization and divisiveness. So these are my four buckets. Should we –
[00:16:57] KAREN: Yeah. Let's go one by one. I mean the mental health one is huge, because that was probably the most explosive Facebook paper from the Wall Street Journal, was the fact that there could be possibly—that there was research done internally at Facebook, at Instagram, showing that there were some adverse effects to mental health for teenagers on the platform. And I know that you have three kids yourself. And you sort of see some of this play out. Or you have concerns as a parent, when you're watching them engage in different social media platforms. So talk a little bit about that.
[00:17:36] JOAQUIN: Yeah. Well, to make this very personal and down to earth, like you said, I have three kids. Youngest one is almost 12. Middle one is 15 ½, and our oldest is 19. And one common pattern that I see that does concern me, is that often you'll see the day pass. And then I'll sit down and say, “Well, what happened the last hour or two?” And, “Oh well, I've just been on TikTok or on YouTube or on Instagram or on SnapChat.” And I go, “Okay, well what came out of it? What can you remember?”
It's like, “Oh, I don't know.” Just like, but two hours have passed, you know. What happened there? [00:18:18] And so I am concerned about addiction to technology. And I am concerned about that time spent not being valuable. In addition to that, of course, there has been a lot of research about things like comparison. There have been some
really interesting ethical debates about things like, whether you should turn beautification filters on by default, or not, which I know are not specific to an AI algorithmic ranking system. But one point that I would like to make right now, and you're going to hear me repeat this again and again, is that when we think about these issues, we should not only think about them as an AI issue, we need to think about the end-to-end design of the system, right. And say if I'm optimizing for engagement, or time spent, of any kinds, some of it is going to be achieved through a ranking algorithm, to try to maximize some metric that relates to engagement. But some of it is also going to be achieved through
UI, user interface decisions, right, whether it's making it super easy to share stuff, or whether it is to have, like, beautification filters on by default and things like that. KAREN: Yeah. Why don't we continue going down the rest of your list. So the next one, I think you spent the most time at Facebook thinking about. This is the bias and discrimination bucket. So talk a little bit about that. [00:19:52] JOAQUIN: Yeah.
To give a little bit of a historical perspective, back in end of 2017 or 2018, I had just spent a good five years-plus at Facebook. I keep saying Facebook because I never worked at Meta. I left right before the rename. So bear with me on that. I had spent a good five-plus years really focused on scaling ML. And then, at the end of 2017, I started to really think about what I should focus on next, right. I think of myself very much as a zero-to-one person.
Once a team or a platform is working, I get itchy feet, and I sort of go , look for the next thing. [00:20:43] And I became very attuned, again, to the fairness, accountability, and transparency community. In early 2018, there was the first dedicated conference which was no longer part of the NeurIPS conference. This forum had been a workshop of NeurIPS in the past. And I thought, okay. I really want to devote my next couple of years to responsible AI in general. I found it to be such a vast area, a bit daunting in many ways. And I got very drawn, like you said, to the question of fairness.
I have a very funny story. I'll just drop it, and we don't need to spend time on it. But we can go back to it. It's a little bit like a philosopher and a computer scientist walk into a bar. So literally, that happened in New York. And I had the pleasure to meet a philosopher,
a moral philosopher, who was half affiliated with industry, half affiliated with academia. And I was so excited, because I said, “Oh, you know what? I figured this out. We're going to build this tool that's going to be de-bias data and models. And then,
you know, we're going to make it available to the entire company. And we're going to get rid of bias and discrimination in AI models within like this year.” [00:21:57] And then the person slowly—slow-motion turned their back to me, and didn't talk to me anymore the entire evening. So it was pretty interesting. And, of course, the reason is that fairness is not one of those things that you solve, right. Like there is no such thing as like a half-solved fairness.
And to an engineer like me, with an optimization mindset, and we get—I'm queuing this saying for later, we'll talk about the optimization mindset, I hope—that's very hard to grasp, right? Because as an engineer, you think about the world as having states, right. And so something is either sober or it's not sober, it's broken or it's not broken. And I got so fascinated about fairness, because it's one of these places where there are just many answers. And all of them are good. It's just that the context is going to dictate which one you should choose. And even more, it's not even clear who should choose, right. And it's very clear that I shouldn't choose. So anyway, it's such a rich area. But bringing it back to the topic, in the context of recommender systems and discovery and suggestions, you can think about concrete examples. You can think about building a tool that helps people connect to
job opportunities, for example. And then you can think about the bias that exists in real world data, right. So there are many studies that show, for example, that if you look at a certain job opening, that has a certain description and some requirements, and a certain salary, right, there is a bias by which men tend to apply, even if they have lower qualifications than women for the same job. So, if you're not careful, and you're trying to learn from the data, your AI might learn that
it should prioritize males for higher paying jobs, for example, right. Which would be pretty terrible, because it would reflect and cement, right, and reinforce biases that exist in society. [00:24:04] Other forms of bias that I guess have been very much discussed, are biases in content moderation. For example, I guess like in the US, in 2018, there was a lot of discussion around anti-conservative bias. And so that led to a lot of interesting discussions too, right, which is like, well, should you, you know, should you sort of like suppress a comparable amount of content from conservative versus liberal outlets? Or should you instead apply equal treatment, and have like procedural consistency, where you say, “No. this is the bar.”
And then, if one outlet produces more misinformation, or violating content, then there would be a bigger chunk of their content removed. And then, different people would have different opinions, right. Different people would say, “Well, I want equal outcomes.” And other people would say, “Well, I want equal treatment.” And then you have
to sort of explain, “Well you cannot have both,” right. And then the question is, who decides? [00:25:06] [00:25:06] So in computational advertising, there's been some brilliant papers on the ways in which algorithms can discriminate. Again, if you're not careful when you're showing people ads, in particular sensitive ads about employment, or credit, or education, things like that. So anyway, I could go on. Like you said, fairness is the area that I spent a ton of time thinking about. And maybe we need a dedicated session to talk about that in depth, yeah.
KAREN: I think we do. Yeah. I think we do. But we'll move along to the last two. And I think I'm going to put them together, because they are very interrelated. And these are true, for me. But the questions that I sort of [00:25:53] is like how do recommender systems end up propagating misinformation, harmful content, or polarize people? [00:26:04] JOAQUIN: Yeah, no, absolutely. Let me—they
are interrelated, I agree with you. I'll still try to break it down. But it's like two sub-bullet points, or something like that, if you will. In a very naive way, imagine that you were to build an algorithmic recommendation system for social media. Now take a moment to think about how you
would do that. The first thing that you need to do, is you need to figure out, well, I need to train my algorithm I need to ask it. I need to give it a goal. So what's that goal going to be? Well, one goal could be, I want people to be engaged. Because if they're engaged, that's a good thing. Well, what does it mean to be engaged? Well, it can mean I click on stuff.
Or I like it. Or I emit any of the reactions—different platforms allow for different type of reactions. Figure it out. Maybe you give a bigger weight to a love reaction, and a smaller weight to a sad reaction. And there's many ways to calibrate these things, right. [00:27:11] Maybe comments matter, right? Because you feel like, oh if, actually, people actually take the time to comment, then that means that they're engaged. So let's focus on comments for a second, right. If you go ahead and maximize comments, then
on the positive side, you might see more content that you care about and you react to. So you might see, I don't know, a friend who got married, and you might have missed it, right, if the ranking system hadn't caught it. And that's good, right. Or, because I'm balanced, I try to be balanced. [00:27:42] Another positive example, you know, I'm super excited about my guitar teacher, James Robinson. Hey James. I don't know if you're listening. Probably not. My guitar teacher, like many creators, builds their livelihood by trying to reach the relevant audience on social media. Not only Facebook, but many others, right. On the flip side, you can sort of see that if I am either a bad actor, or I'm trying to game the system, I'm going to call out Aviv one more time. I don't know if he's there. But
he makes this very clear statement, which is that the way we design our recommender system actually defines the rules of the game, right. If you sit down to play Risk, or Monopoly, or Settlers of Catan, or whatever game you like playing, there are some rules. And you know that those rules are going to incentivize a certain behavior, right.
[00:28:34] If I know—If I figure out that comments are rewarded, I'll be tempted to write inflammatory comments, you know, maybe something that will trigger strong reactions. And I'll get like a ton of comments, right. And then, it's almost like the—I don't know how to pronounce it—so ouroboros serpent that eats its own tail, right? Like you get this sort of a [00:28:53] saying, “Look what cheap Julien, who is an incredible ML researcher by the way, has been talking about.” She calls this the generate feedback loops, right, where people get, then, shown more of that content and less of other things, right. And then you get—People also act by imitation a little bit, right. If you see more of that, you might be tempted to create more of that conflict, right. Well,
until everybody leaves the platform, which is sort of one mode of failure. [00:29:21] So there is a danger that optimizing for engagement will, one, incentivize people to create content that is inflammatory in many ways. But also, the problem is, that you get this “winner take all” phenomenon, where some of that content can, as you were saying, go viral and dominate.
And one thought that I wanted to plant here, I know that I might be jumping ahead a little bit, because I know we want to talk about solutions. But you can think that you can police these things after the fact. You can think, well, this is not a problem with the recommended system itself. It's just like bad behavior. And then we'll add, you know, some integrity or trust or health. Different companies call this differently.
But it's policing—afterwards. I think, of course, you should do that. But I think the more you can frontload incentivizing the right behaviors, the better. [00:30:24] So now the second sublet point, which is polarization. Well, one of the key challenges here is that
the recommender systems are highly personalized, right. And that, again, can be a good thing. That helps James, my guitar teacher, reach the right audience. So that's awesome when it works. But it also means that you get phenomena like what Guillaume Cheslot has studied, which is how, if you create a blank YouTube profile, and you start clicking on the next suggested video, and you start to sort of go a little bit more towards either liberal or conservative content, it kind of tends to tend a little bit to the extremes, and going away from the center. So one of the things that is also heavily being studied, is this idea of the disappearing common ground, right. And this is interesting, because people have been talking about this before social media. People have talked about this already with the emergence of cable TV, where, you know, once
you have hundreds of channels, it's easier to sort of live in your own little silo and only see your own media. And then, if you sit down with other people, you might not have ever watched any piece of news in common, right. So it's difficult to have any kind of civic engagement in that context. [00:32:01] So, there is a risk, as well, that people get polarized, or even radicalized. People are—There are some articles that study preference amplification, which is kind of interesting, which is this idea that there is actually an interaction between the recommender system and a human being, right. So I might come in with a certain set of beliefs, but then through prolonged exposure, even if—and this is very important—Even if no individual piece of content violates any community centers or policies, right, if I'm only exposed to a certain type of content, right, over time, I can become radicalized. And I can become polarized.
And I'd like to mention—I'd like to cite some work that is not from the AI community, from architecture. There is a brilliant architect called Laura Kurgan at Columbia, who has been writing beautiful articles on the analogy between urban planning and the design of social media recommender systems. So where is the connection? The connection is this. The connection is, whether neighborhoods are heterogeneous or homogeneous in the cost and size of housing and so on. And what are the factors that increase or decrease homophily?
[00:33:29] And so homophily is a mechanism by which you will feel even more connected to people who are similar to yourself by any—by specific demographics. This could be, obviously, income, education But it can also be perceived race or others, right. And so in her work, she shows that—Well, she references a lot of work, a lot of studies over the years, that in connection to civil rights, for example, that showed that more mixed neighborhoods ended up resulting in people being less racist, right. And so I find her work very, very interesting, because it points to this idea, right, like this question of how do we inject diversity into recommender systems? Is that even possible, right? I think it's very hard, and we should probably talk about that in a bit. [00:34:31] KAREN: Yeah. I think I really love that example, by the way, the urban planning. And someone asked in the
chat who that was. And it's Laura Kurgan. So I'll just type that in the chat. But, so we sort of talked about these four different buckets of risks that are coming out of recommender systems as you see them. And I want to start jumping to the solutions, seeing the time that we have right now. We obviously—Like you've touched both on the benefits and promise of recommender systems and the risks. So obviously, like we can't really just throw the recommender systems
out. That's not necessarily the solution that I think we should be spending time talking about. So like what are the possible solutions that we should be thinking about? Let's start with what could companies be doing differently? What could people within companies actually be studying? Or what could different teams be outputting, to actually facilitate better public understanding or actually change the way that recommender systems work as they do today? [00:35:44] JOAQUIN: Yeah. I think I—Again, another disclaimer. I don't have the solutions But maybe there's four—and this is a coincidence, the fact that it's four—four thoughts here, in no particular order. Because many of these phenomena are so new, I think voluntary transparency and accountability
as well—But let's just begin with transparency. I think it's extremely important, right? Again, what could this look like in practice? This could look like voluntary daily reports, or public dashboards, that show you what content is going viral where across the world, right. And, if you think about it, in most platforms, whether it's, again, like YouTube, TikTok, Facebook, Instagram, and so on, stuff-- stuff that goes viral is, by definitely, public. It's not a private message from me to you, right. Like that'd be a problem if that went viral, right. And so I think the privacy concerns can be addressed
in that context. And I think the value to society would be tremendous, right. It's almost like you, at a certain level of distribution and size and impact on the world, you're almost like a public utility, right. And you almost owe it to society to report back how things are going, right. [00:37:10] And so one would be to report on what's going viral where. But also, and it would be difficult to implement, but I think it's worth trying,
also segment the population in different ways, by age, socioeconomic status, language, region, other attributes, and try to also report, you know, what kind of content are different groups exposed to? And how heterogeneous versus homogeneous that content is, right? We won't have time to talk about something I've talked about in public in the past a lot, which was the 2019 India elections, and some fairness concerns there. But, you know, there, for example, region and language in India are two very clear indicators that correlate with other things, like religion and caste. And you would want to know, what are people seeing, especially if there was harmful content going about? So that's the first one, right. I think transparency. [00:38:05] There are some good examples. Facebook/Meta publishes these community standards enforcement reports, which are public. If you Google that,
you will sort of see the latest one. I think it's a good step. It sort of shows, hey, what harmful content, you know, is going on? It's not a realtime dashboard. It doesn't get into the specifics. You don't get to see, you know, what exactly is going on. I think we need a lot more of that. So that's one. KAREN: I want to briefly follow up on that. I mean, one of the things that Facebook has
been criticized for is the fact that it sort of games its metrics on those reports. And the realtime dashboards that we do have available, which is through the CrowdTangle tool, is something that Facebook has deprioritized and underfunded and started shutting down. And there have been other examples of, you know, data that used to be available to researchers that was meant for transparency, that it's now been revealed that Facebook is either denying access to these researchers, or it's giving them incomplete data. So how do we actually—I agree with you, that that's a great idea. How do we make Facebook do it properly? Like who are the people involved that should be holding Facebook accountable to transparency standards? [00:39:25] JOAQUIN: Yeah. Facebook, and
everyone else I would say. I'd love to see TikTok do this. I'd love to see YouTube, everybody. My intuition is that there would have to be regulatory pressure for this, which is maybe the second bucket here, right, is accountability mechanisms or accountability infrastructure. I am not an expert in freedom of speech. It's a fascinating topics, and especially in the US, and especially coming from Europe, where the perspective is a bit different, especially now that I have close friends who have gone back to China. And we talk about the cultural differences. And again, it's like fairness. There's no right answer, right, to freedom of speech. [00:40:19] But, when you look at Section 230 of the Communications Decency Act, I think it's reasonable to ask oneself, well, in what circumstances does it make sense to uphold that? And in what other circumstances should we be actually asking for more accountability, and for an obligation to actually report on what's going on? So that would be the second—yeah, the second thing.
KAREN: Yeah. You've also talked to me about this idea of having participatory governance, this idea that it is rather concerning to ask platforms themselves to be deciding some of the things that we sometimes impose responsibility or burden on them to decide. So could you talk a little bit about what participatory governance is, and how you could see that actually functioning in practice, given that Facebook already has an external oversight board, but it's not necessarily working the way that it was originally envisioned? JOAQUIN: What I—I don't know whether the external oversight board is working or not the way it was originally envisioned. For me, I view it as a very good proof of concept that may not scale, that may not satisfy the—it may not go as far as we need it to be. But let me just give you an example. So in May, 2020, back then President Trump wrote, both on Twitter and on Facebook, and probably in other places too, a tweet along the lines that contained his statement, “When the looting starts, the shooting starts,” right.
And not only Facebook, but most media had intense debates on like, okay, what do we do? Do we keep this up, or do we take it down? [00:42:18] And at the time, the external oversight board was not operational. And I remember that it was extremely painful. Because I was dying for it to be there, and to sort of say, “Hey, here you go. Here is an interesting and extremely difficult example where Facebook should not be making the decision on whether that content should be up on this site or not.” And I say the site, it's an app. It's a platform.
Another example would be, right after the Capitol riots, again, the same, I guess—I don't know whether he was formally former President or not. But sort of said things like, you know, “We love you. You're very special. Great patriots,” and so on, you know, talking to the rioters. And then, that was like the straw that broke the camel's back, right, in a way. Then Facebook indefinitely suspended Trump at the time. [00:43:16] And then, I think the fascinating thing is that then, Facebook still did, what I think is the right thing, and sort of went to the external oversight board, once it was stood up, right, in 2021, and said, “Hey. This is what we decided on. Here is the data. Did we do the right thing?” And the way the external oversight board came back, was really interesting. They said, “Well, yes and no.
Yes, that content was unacceptable and should be taken down. But no, you cannot indefinitely suspend anyone, because you haven't defined what that means, and in what conditions you would do that, right. So you've got to clean up your rules, right.” So I think what was really interesting, is that the external oversight board provided feedback in two different ways. One was very specific, right,
on a piece of content and a behavior. But the other one was feedback, even about the governance itself, right, saying, like, “Hey, improve your rules.” And I think that is us, right. [00:44:10] And so I think we need to see a lot of that. But what I don't know, and another sort of plug here, is for Gillian
Hatfield, who is an amazing researcher in Toronto. She has been working a lot on regulatory marketplaces, and this idea that we need to—we need to find other ways to create regulation, because the old ways are too slow. And the reason I'm interested, the angle that fascinates me, is one of like, how do we bring democracy into the process of deciding the rules that we big tech companies need to create, in order to operate? KAREN: Yeah. JOAQUIN: That was three. So I have one more. [laughter]
KAREN: I'm going to interject right before you get to that last one, and just remind people that, if you have questions for the Q and A portion, you can put it in through the Q and A feature on Zoom. And I see that there's already one question. And so I'll get to that shortly. But if other people want to pop that in while Joaquin talks about the third one, I can [simultaneous conversation] [00:45:18] JOAQUIN: The fourth, the fourth, yeah, yeah, yeah. So I wanted to have a prop, and I didn't get organized. My prop is a book. And the book is
called System Error: Where Big Tech Went Wrong and How to Reboot, by three phenomenal authors. They teach the CS-182 course at Stanford called “Ethics, Public Policy, and Technological Change, I think. I might mis-remember. But the important thing is that Rob Reich is actually a philosopher. Mehran Sahami is a computer scientist. And Jeremy Weinstein is a political scientist. [00:45:58] And the reason I'm bringing this as an example, is that book talks about the dangers of the optimization mindset, which is, “Hey. The mindset I grew up with. I'm an engineer, right. I like to optimize things.”
And the problem there is that you might incur failures of imagination. And one of the things that has been very painful, for me, in my last years at Facebook, was people would tell me, “Why are you working at Facebook? It's an evil company.” And I would just not understand, because that's not the reality I was living on the ground. I'm like, “Listen. I can give you countless examples where we've done
risk assessments and have not shipped something because it wasn't ready.” And zooming out, when you think about what might be happening, is less about anyone being evil, and it's more about systemic issues, where, almost like the culture on the approach of things, it's just not diverse enough. Like you can't only have engineers making the most consequential decisions. You need the Rob Reich philosophers. And you need the Jeremy Weinstein political scientists. And you need diversity across not only, you know, your education and background, you need to really sort of embrace diversity and inclusion, and make it count. You need to give it teeth.
[00:47:21] KAREN: This is something that we've talked about a lot, that this system optimization mindset, and how—Because I was also trained as an engineer, and how I sort of went through a similar journey, as starting to realize the gaps in that. And one of my favorite quotes is like, “Engineering is all about the how. And humanities is all about the why.” And that's part of the reason why you need interdisciplinary teams to talk about these issues, because you need to figure out whether or not the question you're even asking, or the problem you've scoped, is even the right problem. We have a bunch of questions that are starting to come in. So I'm going to start taking some of
them. And then I have a couple other questions as well, that I'm going to try to weave in. But the first question from Hamid is, recent events in Ottawa—so this is referring to the trucker convoys—and elsewhere have shown that extremism has found its way into the virtual world. Given the recent revelations about Cambridge Analytica, in what way could Democratic societies protect citizens and their Democratic systems? Will it be through more public oversight and regulations of the algorithms, and/or metrics, or better enforcement of corporate tax systems, et cetera? [00:48:34] JOAQUIN: I think it's all of the above. I don't think that I have one single answer to this. Maybe I'll bring up—So I think a lot of what we've discussed already addresses the question, like the four buckets of solution that I mentioned are the ones that come to my mind. There's one more that I just thought about. In the same way as sometimes we have a public health education campaigns that
say, “Wash your hands. Because if you do, then you'll kill germs.” And I think—Oh, my memory is-- There is another Belfer Fellow—oh, who used to be Chief Technology Advisor to Obama. Damn it. You know who he is. It'll come back to me. Patel, yeah. He used to talk a lot about the example of the impact of people starting to wash hands in hospitals and so on, or like in the medical profession.
Believe it or not, there was a time where doctors and nurses wouldn't wash hands. [00:49:44] And so I think that we need to invest a lot in the public understanding of social media and educating people. Just to bring in a bit of optimism here, in the middle of this very stern conversation, I am actually encouraged sometimes. The flip side of my kids spending so much time on social media, is that we have these amazing breakfast conversations, and also it's amazing to see them educate their grandparents. Like, “Hey, grandma,” you know, in German to their German grandma, or Abuela to the Spanish one—or Abuelo, you know, the grandpas—“Like, by default, you shouldn't believe what you see people share something on WhatsApp with you. By default, don't believe it. It's like spam.” And they're like, “Why would people do that?” And my kids kind of like explain it to them.
[00:50:30] So I think—I think—I'm hopeful that, at least in some of the younger generations, people are developing some sort of an immune system of sorts. But again, like I'm an optimist. But yeah. I think investing very heavily in public education, in addition to all of the regulatory steps, from transparency, accountability, and parts through governance, I think, is necessary Will it be sufficient? I don't know. KAREN: There's another question here. It's a really hard one. So it says, it's from Derek.
It says, you pointed out the Facebook oversight board model of accountability doesn't scale well. So are there accountability models that you think do have the potential to operate at scale? [00:51:13] JOAQUIN: I don't know. I came across papers on, I think it was called fluid democracy, or something like that, which to my computer scientist brain, translating to a tree. And trees can be efficient constructions, right, where you could imagine—Well, going back to social structures that, through the centuries have ensured that social norms sort of prevailed, right, like you have maybe the elders in the village. And then the elders in the county. And then the elders in the region, and whatnot. And so the problems you need to solve are not just participatory governance in general.
It has to be localized and in context, right. I was saying earlier that freedom of speech means something very different in China, in the US, and in Germany, and in Spain, right. And there's no right or wrong. But you need to localize it, right. So the question is, it's again, like, do we need to sort of reinvent democracy in a certain way, make it very fluid, and figure out how do people elect their representatives that will actually make those decisions? KAREN: I'm going to ask you a question before I move on to more audience questions.
So one of the things that you invested a lot of energy in, at the end of your time at Facebook, was diversity and inclusion. And that's sort of part of, I think, it folds very much into this conversation of finding solutions, and how companies need to shore things up internally, to actually tackle some of these issues head-on. So why was that important to you? And how do you see that fitting into this conversation that we're having? [00:53:00] JOAQUIN: NeurIPS 2019. So, for those of you who don't know, NeurIPS is one of the main machine
learning and AI conferences. I've attended, I think, all of them since the year 2000. YoYo Ma, the famous cello player, was invited to a workshop. He came. We chatted. And then, I was lucky, we bumped into each other. I had a few minutes
to talk with him, which was incredible, one-on-one. And I told him I was working—I was starting to work on responsibility. I was starting to understand how do you help people trust AI? So I asked him that question I said, “Well, what would make you trust AI, YoYo?” And he said, “Well, the most important thing is I need to understand who is behind the AI. Who built it? What are their motivations, their concerns? Who are they? What are their lived experiences?” And I thought, okay. It's not looking super good, because it's a bunch of people like me. [laughter] [00:54:04] And, of course, every company has been talking a lot about diversity, equity, and inclusion, DNI, inclusion and diversity, you know, the order and the acronyms change and so on. And I think I was frustrated by the fact that it was a lot of talk. But I wasn't seeing the
results. And so, I decided to dive in and trying to understand, well systemically, what is going on? And I realized that the problem is that I didn't think leaders were being held accountable for creating an inclusive culture, or for evolving recruiting, to really create equal opportunity. Although I need to emphasize, the most important piece in diversity and inclusion, in my opinion, is actually the inclusion part, so actually what happens with the people who are on the team now. You don't address diversity and inclusion by hiring a black female into your team, or by hiring an Asian transgender person into your team, or a veteran, or whatever it is. That's great if that means that you have a consistent process that gives equal opportunity to everybody, and these people were great. Super. That's really good. [00:55:10] But, if you don't really have an environment where everybody can contribute equally, right, and where decision processes aren't dominated by a small minority, then you have achieved nothing. And so we—I like to work
on hard things. You know, we teamed up with HR, with legal, and said, “Well, how do we change our performance review system, so that, how much we pay people, whether we promote them or not, and so on, actually depends on very clear and concrete expectations?” So again, we did another session for this, to go into the details. But we did this. We shipped it. And I feel very happy about that step. That doesn't solve the problem. But I think creating hard accountability, right, and making what people get paid, depend on clear diversity and inclusion expectations, is essential. [00:56:01] And then people sometimes say, “Well, but how does this help the business?” Look. It's very simple. I think you make better business decisions by having a much more diverse and inclusive taskforce. And you can anticipate problems that you didn't see, right. And going back
to some of the things that we discussed earlier, and maybe the Ottawa question. I think that technology companies need to have philosophers, more philosophers [00:56:32], political scientists in the leadership team, at the very top. And that they need to be put in roles that have teeth on an even level with engineering. And obviously, those are functions and disciplines. I think that diversity and inclusion needs to sort of extend beyond there. KAREN: Yeah. I mean I think one of the things that we've talked about before is that the side effect of bringing more diverse lived experience is that you also, you have more diverse expertise experience, or whatever you want to call the expertise. Because a lot of the people who
are from more marginalized backgrounds that don't typically appear in these types of roles, are actually in the other disciplines. And they're like—it's because they're, for whatever reason, they're pushed away from tech into perhaps a social science, or from AI into AI ethics. And so there's like a really complementary effect that happens when you open up the table, I guess, to all of these different perspectives. So I'm actually really curious, because we've never talked about this, is like, what actually happened once you implemented this at Facebook? Like have you seen noticeable differences yet? Or do you think they're still to come? [00:58:08] JOAQUIN: I think they're still to come. These things take a long time. What I can say, is that the level of awareness increased dramatically,
right. And people would actually discuss, systematically, diversity, equity, and inclusion during performance reviews, which is something that happens at least twice a year. And I say at least, because you have the big ones, and the small ones, and the informal ones. And
if you're a manager, you spend your time doing performance reviews in one shape or the other. Making something be top of mind, I think, is a crucial first step. And that did happen. [00:58:48] There were a lot of pretty awesome improvements to recruiting processes, which were already pretty good. And I also saw a lot of—I saw a change towards really acknowledging and rewarding people who were investing in building community. And you and I have talked about this, in some of our conversations, this triple whammy of sorts, right, where I would be in meetings, working on diversity and inclusion and incorporating that into performance review. And I would look, and I'd realize that I was the only white male in the meeting. And so everyone else
was donating some of their time, right, to work on something that might not even be rewarded, right. So first of all, working on something that is an opportunity cost[?]. You could be working on something you know is rewarded, right. Second, well, this work you're doing doesn't get rewarded. And third, you might even be
perceived to be annoying, right, like a fly in the ointment or something like that, right. And sort of say, like, “Listen. You're annoying. Can you get away? Like we have some business goals here, right, to achieve.” I lost my train of thought. [laughter] [01:00:12] KAREN: Yeah. Well no. I mean the thing that I love about—The thing that I love about this particular anecdote that you're talking about, or these reflections, is the fact that specifically, that you decided to take on that triple whammy as a white man. Because typically, the reason why women or people of color are the ones that take up the triple whammy is because, for them, it's an existential thing. Like they have to do
it. Otherwise, they're not going to survive at the organization if it doesn't get better on certain fronts, or they're not going to progress in their career if these things don't get better. But for someone where it's not an existential crisis, to actually put their weight behind that, and lend that, I think that's—yeah, it's really incredible. And it's a very good demonstration of how to be a good ally in those situations. We've spent most of this time talking about like what companies can do. I
do want to just briefly touch on, what are some of the things that people who are not inside of these companies can do, whether that's civil society, or the press, or regulators? We've already touched on some of the things that you think regulators can do. But what are some of the things that should be happening outside of companies, very explicitly, to help us push and topple these problems around social media recommender systems? [01:01:30] JOAQUIN: Yeah, a couple of things. Let me start with the job seekers, right. We are seeing a pretty dramatic revolution, I would say, in the labor market, where I guess people talk about the great resignation. Others talk about
the great reshuffle. People are choosing to leave their current job and thinking about where they want to go. Well, one of the first things to do, is inquire about the values and principles of the company you are going to, right. Like job seekers have more leverage now than
probably ever in history, right. So I think this is a great time to do this. [01:02:14] And it's working, right. LinkedIn published this really interesting study, which is, within 12 months—and I'm sorry if I'm misquoting the numbers—but they approximately go like this. Within 12 months, we've gone from
one in 67 jobs offered hybrid employment, where you could work from home some part of it, to one in six, right. That's crazy. It works, right. So markets are very powerful, for sure. Speaking of markets, I forgot to mention this. You know, I guess another one is to further encourage competition, right. There is something we will have to talk about some other
day, but I'll just plant it here in case people want to look it up. And probably most people at the Belfer Center will know about this work. Stephen Wolfram proposed this idea of saying, “Hey, listen. Maybe we should force companies to open up, to a market of ranking providers, right, where maybe you could have” --And, by the way, the engineering design of this would be extremely difficult, right. And there's a lot of privacy concerns. But bear with me for a second. [01:03:28] I know Twitter is pretty excited about this. I have good friends there who have been vocal about this, right. So where you could almost have
like an API or way to plug in your own ranking provider. And again, right, I don't know. I love the outdoors. Maybe there's even like the REI ranking provider. I'm kidding. But, you know, you could choose like the Fox News ranking provider, the NPR ranking provider, right. Like I don't
know. I listen to German radio, Deutsche Welle[?] ranking provider, whatever it might be, right. So I think competition on marketplaces could play a big role here. There's probably more ideas. Maybe I'll throw one more, which is, let's have more courses, like CS-182 in Stanford. And I know, with all respect, I know there's the equivalent in most schools these days, where you see philosophy, political science, and technology converge, and really create a forum where we just really imagine how society should function.
KAREN: And CS-182, it's their ethics and computer science class? [01:04:47] JOAQUIN: Yeah, that's the ethics, public policy, and technological evolution. I don't want to get derailed and start typing. It's almost that, the title. KAREN: Yeah. This is a good segue into this last question, because you were talking about some studies on just the future of work. And we started on a personal note about how you made your way to Facebook. So I just wanted to end on a personal note, of where you're going
to head next. Because you've left Facebook. And you recently went on your own journey to figure out what you want to do, continuing to work sort of in this responsible technology space. So what was the process that you went through? And tell us where you've landed. [01:05:25] JOAQUIN: Yeah, yeah. Well, I've just joined LinkedIn, as it happens. And this was—And I'm super excited. At LinkedIn I'm going to be a technical Fellow focused on AI.
But what really excited me was, on one hand, the mission of the comp