Tech Brew Ride Home - Using Tech For Good? With Jigsaw's Dan Keyserling

Episode Date: April 27, 2019

This one requires a bit of explaining. Remember that segment I did about Change A View, that subreddit that was becoming its own site to try to create a platform for healthy discourse and debate online? They got backing from—and technical support from—an Alphabet subsidiary called Jigsaw. So, I went down a rabbit hole with Jigsaw, because they weren’t on my radar, and I learned that their remit is to… seemingly… try to make the Internet not suck as much. So, long story short, I reached out to some people and got put in touch with Dan Keyserling of Jigsaw to see what they’re up to. Combatting radicalization, online censorship, trolls, bias in AI… there ARE still some folks who believe technology can make the world better… not just break it. Huge, if true!  But seriously, this is a bit of a palate cleanse for me, and for you… but really… mostly for me.  Sponsors: Wix.com/podcast Everykey.com, promo code RIDE20 to get 20% off. Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to a weekend bonus episode of the TechMeme Right Home. I'm your host Brian McCullough. This one requires a bit of explaining.

Remember that segment I did recently about Change of View, that subreddit that was becoming its own website to try to create a platform for healthy discourse and online debate? They got backing and technical support from an alphabet subsidiary called Jigsaw. Jigsaw actually wasn't on my radar, so I went down a bit of a rabbit hole. And I learned that Jigsaw's remit is to seemingly try to make the internet not suck as much, or at least use technology to make the world a better place,

which is bizarrely kind of feeling like an old school thing at this point. So long story short, I reached out to some people and got put in touch with Dan Kaiserling of Jigsaw to see what they're up to, combating radicalization online, online censorship, trolls, bias in AI, all of that is what Jigsaw is working on. It turns out there are still some folks who believe technology can make the world a better place and not just break everything. Huge if true. But seriously, this is a bit of a palette cleanse for me and for you, but really mostly for me.

Please enjoy. Well, I'm here to cheer you up. I'm going to give you a positive, hopeful vision of the role of technology in the world. listen, if you can do that, by the way, let's, let's, let's, let's, let's, let's endeavor to do that. That's all I'll promise. All right. Well, then, can we consider ourselves recording officially? Sure, absolutely. Okay, well, because if you can do that, then, uh, then you'll make my week here. Um, so I, I got turned on to, to jigsaw specifically, I, I, I'm ashamed to say, I wasn't aware of you previously, but, um, because I did

a segment recently about Change of View, that site that's coming over from Reddit. I believe you guys are working with them. So just real quick, can you tell me how Jigsaw got involved with Change of View and what you guys are trying to do with them? Sure, and thanks for having me. Yes. So that story starts with a tool that we've been working on for a few years called Perspective. And perspective, put simply, uses machine learning to spot toxicity online. And one of the most common applications of that technology is to help publishers and platforms

all across the internet host better conversations. So that's everything from helping human moderators moderate conversations on, for example, news websites like the New York Times or other platforms that have slightly different formats and slightly different standards in terms of service, they can use that technology to help basically enforce the terms of service. So, you know, Reddit is one of the most in the Internet. It's a place where people go to have conversations exchange a lot of ideas. Reddit has a really strong tradition of sort of having the community moderate itself. And we, in a number of conversations with Cal from Change of View, we wanted to really help him use perspective as he stood up his

new venture and integrate perspective to help his moderators do their jobs more effectively. Can I ask possibly a dumb or maybe even insulting question because maybe it's more complicated? But is it more complicated? Toxicity, especially in things like message boards and things like that, is it really more complicated than just having moderation, like leading into moderation even just a little bit? Or like what's the secret sauce that change of view, in your opinion, has found? It's a good question. I guess I should explain a little bit what I mean by the word toxicity, and that's a very deliberate and specific term in how we think about how to train the machine learning models and just stop me if I start talking tech jargon. No, go on.

Okay, perfect. One of the challenges was to build machine learning models that were nuanced enough to provide useful information to human moderators who, of course, have. a very nuanced understanding of language. People can detect things like passive aggression, sarcasm, personal attacks, misogyny, racism, et cetera, all of the different sort of facets that make up toxicity online, which a lot of publishers and platforms want to either, if not filter out on some level, be aware of or be able to detect.

And different platforms have different standards and preferences for what they want their communities. to be like. So the first challenge was creating machine learning models that were adept at detecting all of those things and of putting that power in the hands of the people who are running these platforms and publishers. So in the case of Reddit and in Change of View, Change of View is really interesting, and I was actually rereading the blog post that Change Review published when they launched their new site. And, you know, there are two different things that Change of View really focuses on. They're interested in, obviously, sort of making sure that their communities don't have harassment

and attacks and things like that, but also measuring what they call the delta, the difference, basically, between a person's initial opinion about a subject and their opinion after they engage the community and they hear for stimulation. So in the case of perspective, what we did is we provided sort of a tool that allows change of views moderators to have comments flagged so that they can review them more efficiently. And that's not an uncommon application for perspective. A lot of news organizations use it similarly. New York Times, as I mentioned, uses perspective to help its team of human moderators sort through the comments on their site so that they can do their job more effectively.

But then again, it's just coming down to, it's just having more humans in some sort of a moderator capacity, and then it's just providing them with more tools to do that at scale. Exactly right. The keyword there is scale. So human moderation is obviously not super scalable. to hire people. It's expensive. It's labor intensive. It takes a long time. A tool like perspective just makes them able to work more effectively by sorting comments according to toxicity, for example. And this isn't something that a user or a reader would necessarily notice.

But if you're trying to manage a community and you're trying to engage with thousands, tens of thousands, even millions of comments, machine learning can really help make, help you accomplish that at scale. And its perspective is an API that in theory anyone could ask you guys to use on their applications as well? Yeah, exactly right. It's free. Well, so again, I apologize for not being familiar with Jigsaw until hearing about Change of View. But real quickly, just tell me about Jigsaw how it got started and what the mission is. Sure. Jigsaw is a unit within Alphabet, which is Google's parent company that uses technology to address some of the most pressing global security challenges. That's everything from radicalization online to hate and harassment, to defending news

organizations from certain kinds of cyber attacks, defending against censorship, for example. We are an interdisciplinary team, so we combine researchers and software engineers, product managers, designers, legal experts, people with a history and policy. And we sort of take an interdisciplinary approach to these issues. We try to forecast out threats years ahead of time. We try to really understand the role that technology is playing in these threats. And then we build technology to address them. So, yes, this is why I wanted to talk to you. Because, again, the general consensus over the last few years increasingly seems to be the internet and technology in general is breaking society in a lot of ways. So let's kind of, let's go down some of the things that

you just mentioned and some things that come up all the time on our show. And, you know, at the risk of putting you on the spot or, you know, giving bullet points on something that's a complicated issue, let's just go into some things and tell me some solutions that are possible. Let's start with, like you mentioned, radicalization. That's obviously been in the news a lot lately. What are some ways that the Internet and technology can fight radicalization as opposed to seemingly just fomented? The way that Jigsaw approached this issue was to ask the question whether it was possible to provide information to people who are considering joining a violent extremist group, in this case, in our case, ISIS. And whether that information could change something.

So could you reach people who are interested in joining ISIS and provide them with information that would undermine ISIS's mythology that would thwart their ability to recruit new members online? And so Jigsaw conducted an experiment with something called the redirect method where we created an advertising campaign, a targeted advertising campaign designed to reach people who were considering joining ISIS or who were displaying an affinity for ISIS's propaganda and mythology online. And we directed them towards video content that addressed their questions and addressed their queries,

but provided a viewpoint or a series of views that undermined ISIS's mythology that didn't conform to sort of their version of the world. And what we were interested to know in that case was, first of all, could we identify people who had an affinity for ISIS propaganda? Could we reach them in some way? Are there terms that people search for that suggest that people have more than just an interest in ISIS or the subject matter, but are really displaying an affinity, a strong sort of desire to subscribe to that ideology? and then if we showed them videos, would they be interested to watch them? Because sort of the obvious conclusion is that if somebody shows you a video that isn't what you want to see or doesn't address you question, you just click back and you don't bother watching it.

But in our experiment, we were pleased and surprised to see that not only could we reach hundreds of thousands of people who were searching for things that suggested they had an affinity for ISIS, but they consumed this content. They watched these videos that provided everything from citizen journalist footage from inside the so-called caliphate, footage from inside hospitals, depictions of breadlines and civil society breaking down in Syria, or sermons from more moderate Islamist scholars. They watched that video content for much longer and at a higher percentage than comparable searches. And for us, that resulted in a methodology, a way to go about reaching people who were vulnerable to being recruited by violent extremist groups.

And we published the results of that research so that other groups could learn from it and can think of ways to apply it to other violent extremist groups. Okay, online censorship, I think you also mentioned. And I've been doing a lot of segments recently. I came up in the 90s, and so my assumption about the web and the internet was always like, well, this is the silver bullet against totalitarianism. The internet's always going to be open. Information will always be free. But it's almost starting to feel like it's the opposite, like the internet, and technology

is like the perfect mechanisms for creating a panopticon. And you have, like, sort of the surveillance internet. in certain countries becoming more appealing to a lot of governments. So what is some of the work that you guys are doing to combat censorship and things like that online? So censorship is a broad word, and I think it's worth sort of unpacking what that looks like all around the world, and also some of the ways that governments restrict access to information that we might not necessarily. call censorship at first glance. So one of our programs is something called Project Shield, which put very simply, defends news organizations and political organizations around the world

from a very common form of cyber attack, a distributed denial of service attack, which is a type of cyber attack that overwhelms servers with traffic and takes websites offline. This is a very common type of cyber attack. And it's actually often used as a very crude form of censorship. So instead of creating some elaborate censorship apparatus, a government or a non-state actor can direct traffic to a website that it doesn't like and overwhelm it and take it offline, historically defending against these kinds of attacks can be quite expensive. There are commercial providers of DDoS mitigation technology, but small news organizations and small political organizations, human rights groups, NGOs, they often can't afford these

services or they don't necessarily have the technical staff to implement them. So we created a service that offers that for free for certain categories of organizations. And what about things like just you and me? like your average internet user or whatever, like what are some ways that, I mean, and I'm not even getting into the surveillance of like, you know, our data being gobbled up by various companies and things like that. But what are some easy solutions for your average internet user to feel like they're not being watched by the Panopticon all the time?

Yeah, well, we could talk about the subject all day because it's complicated. and it exists in a bunch of different forms that are all worth addressing. But I'll tell you about a couple, a few of the things that our team works on. The first one is a tool called Intra. And it basically encrypts your connection to the phone book of the Internet, the domain name system. Because another common form of censorship is what's called DNS manipulation or DNS poisoning. And it's when really a government intercepts the connection between individual devices or IPs and domain name system that is the backbone of the Internet, the phone number that connects URL to server addresses somewhere, and manipulates it or blocks it, or in some cases distorts the web pages in certain ways. We saw this technique being used in Venezuela, most recently, during the violence and the recent political activity there.

And so we built this app, and we had been testing it for a couple of months in Venezuela. And it's a really simple app. It's one button. And when you press it, your DNS connection is encrypted, and DNS manipulation is no longer a threat to you. The truth is there are a lot of different versions of a similar tactic, ways to disrupt people's access to information. And part of what Kixot does is, first of all of those threats, all of the different ways that censorship manifests in the world, how that technology changes over time, and

what the most effective ways are of creating more access to information. If I can squeeze one more in real quick, because again, these are stories that I... I can talk about our VPN, too. Well, tell me about the VPN real quick, and then I've got one more to ask you about. Okay, perfect. Another tool that XR developed is called Outline. And many of the listeners I'm sort of will be familiar with virtual private networks, VPNs. A lot of people in the world, hundreds of millions of people use VPNs all around the world every day. The issue, and a lot of people use VPNs to make their connection to the Internet more secure

or to circumvent government firewalls or restrictions on their access to information. A lot of businesses use them just to manage who can be on their network and who can't. The technical challenge that we were trying to overcome there is that a lot of organizations want to create their own VPN because they might not always trust the providers of VPNs around the world. But it until very recently was very hard to do that. It required technical expertise to set up your own VPN. And even then, it wasn't always clear whether you were doing it right. Specifically, we had talked to a number of human rights groups from around the world who were very interested in really having their own. technological infrastructure so that they could make sure that it conform to their

standards and that they could trust it but it was a hard thing to do and so outline makes it dead simple to create your own VPN it's as simple as that so any organization around the world can create their own VPN and choose from a number of cloud providers or set up their own servers and then they can offer access to that VPN to anyone in their network for for organizations organizations and news organizations that are operating in places where the government might be hostile to those groups or where the government has a strong interest in restricting access to information, those tools can be invaluable. All right. The last one I want to squeeze in is seemingly everyday stories about bias in AI. is, again, is it simply down to the fact that we're still at the stage with AI where it's humans that are teaching the machines and thus we're bringing in the human biases inevitably?

Or is there an elegant solution or maybe an obvious solution to introducing biases into AI? It's a really good question, and I'm glad we have the chance to talk about it. It's a really interesting subject because where we have encountered issues with bias in machine learning is in the context of training machine learning models on large quantities of online comments. So let me explain what I mean by that. So the way that we taught the machine learning models for perspective to understand toxicity in language was to show it a lot. lot of language. We showed these models millions and millions of comments, public comments online from newspapers like the New York Times. And in processing the tag data of millions and millions of online comments, that's how it learned to spot toxicity. So the question of how does bias

sneak into machine learning models is because all of those comments contain human biases. And And you often don't know what those biases are or how they'll manifest in machine learning models until you start to test it. And so a big effort that our team invests really heavily in is testing all of our models and working with outside groups to really try to, first of all, identify bias and then mitigate it and fix it. We really believe in publishing our research and showing our work and engaging in the broader AI research community on these issues. We publish a blog called the false positive, and we've written a couple of research papers on this subject. Because as you rightly point out, we have to make sure that we get this right. A lot of these technologies are experimental, and it's well worth our time, sort of spending time

thinking about how we can better identify and mitigate biases in machine learning. Final question. I don't know if this is like an unofficial motto or something, but I was reading up on jigsaw and like there's this phrase that I think you guys use, how can technology make people in the world safer? That's sort of like a mission statement for you all. It's written on the wall right behind me, yeah. Oh, okay. Well, there you go. So I've, I've I've sort of joked before that maybe one of Silicon Valley's problems is there's the tendency to believe, well, for whatever the problem is, if we just get a bunch of smart engineers in a room, they can solve it, right?

I'm curious, to what degree are you guys aware of your own sort of biases? Like, how much are you thinking of the need to get non-tech points of view to fix some of the issues that you're working on? It's absolutely critical. to get not just a non-tech point of view on all of these issues, but a really diverse set of views from different industries, different groups, different countries, different parts of the world, different ideologies. A big part of how Jigsaw works is trying to get those viewpoints. So a core part of our methodology is sending people out into the field to understand

how technology actually works in places that are facing oppression and violence. It's to speak to local communities, speak to advocacy groups, speak to experts, the academic community, to speak to governments around the world who are also confronting these problems, to really understand as many perspectives as we can to inform our own. and lead with that sort of interdisciplinary, inclusive approach to problem solving. I know that that sounds even a bit idealistic, but we find that our technology is better when we incorporate non-tech points of view in its inception, its design, its testing, and its deployment.