Shrey Jain: Applied Scientist at Microsoft Research Special Projects
Shrey Jain, an applied scientist at Microsoft Research Special Projects, speaks with Matt Prewitt on a very timely and topical subject: AI and – more specifically – the dangers it poses to the nature of natural human communication (“context collapse”). They take a deep dive into the current threats to privacy by expanding beyond the often discussed cryptographic sense into “privacy as contextual integrity”, and the immediate opportunity to embed ethical guardrails into this ever-changing realm of generative AI through possible solutions of designated verified signatures in “plural publics”.<br><br>Shrey’s recently published paper co-authored with Divya Siddarth and E. Glen Weyl “Plural Publics” is linked in the episode notes.
Top Keywords
- context 0.014
- communicate 0.007
- communication 0.007
- shrey 0.007
- information 0.006
- mean 0.006
- norms 0.005
- privacy 0.005
- another 0.005
- knowledge 0.005
- group chat 0.004
- publics 0.004
Transcript
This is a RadicalxChange production. Hello, and welcome to Radicalxchanges. In today's episode, Matt Pruitt speaks with Shrey Jain, an applied scientist focused on AI security and cryptography at Microsoft Research Special Projects. Shrey's concentration on information integrity takes us on an exceptionally compelling and timely journey in the era of generative AI. Together, Shrey and Matt cover privacy and context collapse, one of the issues of AI undermining natural human communication. One possible solution being Plural Publix, highlighted in a recently published paper authored by Shrey, Divya Siddarth, and Glenn Weil. This is a captivating and salient conversation, and we hope you enjoy it. Without further ado, here is Matt Pruitt and Shrey Jain.
Shrey Jain, great to great to see you. Great to have you. Thank you for thank you for joining today.
Thank you for having me. I'm really excited to chat.
Super. So let's begin by, can you tell us a little bit about, about who you are and and what you do?
Yeah. For sure. So I'm Shrey. I work at Microsoft Research. I'm currently an applied scientist there with a primary focus on epistemic security research, which means how do we protect our communication infrastructure. And a lot of that relies on to how we send information and how we verify the authenticity of that information as well as the identities from which information is being sent from.
Super. And I you and I have talked, talked quite a bit over the past year or so about, about your work and about these these interesting questions about, information provenance and and common knowledge. And, I know that we'll get into some of that stuff at some point here. But, recently, you've been thinking and working a lot about, the the topic of the day, which everyone's thinking about, which is, which is, AI. Would you you wanna sort of set the table a little bit on what you've been what you've been thinking about? And we'll jump in.
For sure. Yeah. I mean, maybe we can go a little bit before AI and talk about the problems that are already there with the Internet, and then we can talk about how these problems will be intensified by AI. And so, like, to start at the Internet, it's interesting because I think the Internet is often viewed as this, medium that has allowed this global connectivity and has allowed communication to grow in a way that humanity has never seen. And what I mean by that is, like, so long as you have an Internet connection to to some sort of network around the world, you can kind of communicate with whoever you want geographically, instantaneously geographically, like, independent, but instantaneously. But what's not as well discussed is the harms that that has had to communication. And what I mean by that is, like, when I communicate with you in person, there are certain properties of our methods that, or, sorry, of our in person communication that are lost when we start communicating online. And the first concept I wanna try to illustrate is what's called context. So context is what is lost and has been lost in our digital communication today. And so, like, what is context? And it's defined differently by various different academics. So I'll be as precise as I can for what I mean by context in in the context of my research at least. And so what I mean by that is it's the information or background knowledge that is known to be known and known to be known to be known and known to be known to be known amongst the people in a given space. And so an example of that could be language, meaning we all have an understanding or we know the set of rules or symbols that we must follow to speak with one another. Another example could be, religious traditions or culture milieus or, religious, or sorry, like, nationals national memories. These are all examples of things that are common knowledge amongst each of us and give us context towards, the communication we have with one another. Like, the only way I can communicate with you to the depth that I am is because we have context around, one, I'm entering this podcast with a given understanding of what we're gonna talk about. I have some understanding of you and your background and the shared experiences we've had with one another, and that influences the way that I communicate with you. And so context is actually, like, essential to communication. It's, like, fundamental to the way that we communicate with one another. And that's why when we speak to our loved ones, we communicate in a very different way than someone random on the street because there is that shared history that we have with these people. So that's kinda like the first concept that I wanna illustrate is, like, very important, and how the Internet has hindered our ability to both establish and preserve context is that now you can kind of just share information wherever you want. You can put it onto Twitter and have not just your loved ones or your coworkers or the people around you, observe that information. But you can have virtually anyone, observe that information. And they may not have that same context or understanding to understand what it is that you're saying. And so as a result, their expectations, their dispositions towards that information are very different than how you interpret it. So that's, like, the Internet argument. I'll pause there before maybe we can go to the AI. I just wanted to see, what how you maybe feel differently on that as well.
No. Let's, let's peel back, what you're saying a little bit. So the the idea is that in order for us to communicate with each other, we need to have some common ground to stand on, basically. So for for example, when we when we use a language, one of the reasons that we're able to use the language to communicate is that we have some shared understanding of what each of the words mean, for example. Right? And there's all kinds of things like that. Lang well, language is just one example, but there are there you know, another I think another, it might be useful to, think about this in the context of sort of media history and stuff. Like, for example, in the one thing that often gets talked about is the twentieth century, there were, like, three news stations in The United States for or something like that. Right? So everybody was listening to Walter Cronkite. And when people, met each other in the in the bar or in the workplace and talked about the news, they'd all seen the same news. And so they were therefore able to able to communicate, in in a certain in a certain way. Right? And, I mean, can you give can you give a few more examples of of context maybe and help drive the intuition that more communication can cause a loss of context. And and and just to be just to I think one more interesting thing to pull apart here is that sometimes I think we have the thought that that that common ground is problematic. For example, so for so, like, if we have if we have certain common ground in our communication because we both listen to Walter Cronkite, but Walter Cronkite is giving us government propaganda, then, like, maybe maybe they're even though we're able to, you know, communicate and understand each other in a certain way, we might be standing on shaky foundation. Like, there might be something wrong with our contacts. Right? So So I think a lot of these a lot of these different, dimensions get, you know, people people bring them in when they hear about this stuff, and it it makes the conversation quite complex. And I wonder if you can yeah. Maybe a few more examples. Does that make sense?
Yeah. For sure. So let me come back to how we can create a plurality of context and why maybe then we become less scared of the foundations that we're we're resting on. But first, I give a bit more, onto, like, why context is important. And there's this very rich social theory here from Georg Simmel, who has this canonical paper called the sociology of secrecy and secret societies. And it's a very elegant piece of which outlines why having context in a relationship is what gives you confidence to speak to someone else. And so, like, as an example, the way the reason I go up to someone and reveal more about myself or my life is through the shared context we have with one another. And in his paper, he, refers to this as reciprocal presentation. And so reciprocal presentation is how you build context with one another. It's through these, like, lived experiences. It's through these inside jokes we have with one another that allow us to communicate in a far more efficient way than maybe we would if we didn't have that context with one another. And, like, if we wanted to go to a very, like, real example that's not just between friends or or, your family members, like, even in very specific fields of academia, we see this, where to become a doctor, you there's this canonical set of textbooks that you read that doesn't reading those textbooks doesn't make you a doctor, but allows you to speak the language that other doctors speak. So when I say a medical term to a doctor, I assume that they know that what that term means, and I have this understanding that they also know what it means, and they know that I know what it means if I'm a doctor. And it's like this rich common knowledge of terms that allows us to speak so efficiently and work and solve problems very effectively. Effectively. So context in those scenarios is, like, actually critical to to to communication and the success of that occupation. And another context is just, like, how we live our life and the way that we are able to communicate with one another. Like, there's a beauty almost in having inside jokes with people because you don't actually have to speak a different language, but the inside jokes themselves is a way of, like, encrypting your words in the physical world. Because even though you're speaking in English, this cup of tea to me may mean something very different to you because of some inside joke I may share with someone else. And it's all depends dependent on the ways in which we present ourselves. So that's, like, a bit about context. And then you're right. Like, there is a real fear of, like, what if the shared knowledge that we have with one another, as an example, the field of medicine for all we know could be dependent on axioms of science that are completely false. But I think it's important that we have ways to present new ideas and present new forms of knowledge that can become part of this base of which can eventually become common knowledge amongst all these participants. And so what we'll talk about maybe later in the in this episode is, like, how do we get to that ability to bring new knowledge and then not just bring new knowledge, but make it common knowledge amongst everyone.
Yeah. One more idea that, is is in this mix is the idea of signal and noise. So, for example, I think tell me if tell me if you think this is the right way to think about it. Sometimes when we get a signal, you know, we might hear it with, like, no background noise. Right? So in a way that sort of silence around the signal is part of the context. And, so when I think about the when I think about the kinds of ideas that you're getting at here, context collapse and sort of, you know, contexts, you know, these sort of separate contexts between people becoming, becoming harder to maintain as separate contexts. It's like the signals become noisy in in a way. Right? It's, like, harder for people to understand what's being said amongst amongst themselves because they've lost that sort of clean context in which they're able to communicate. Is that is that how you think about it?
Yeah. And and if I if I may, I'll, like, extend on that. I guess to give two more examples of why context is collapsing today. First, let me go into, like, why do we even care to connect with one another, and, like, why do we care to communicate in this very, like, broad form? So I think here, I'd wanna reference a bit of John Dewey's work from The Public and Its Problems. And in that in that, work, he references this idea of what are known as emergent publics, which are, these spaces in which share people with shared interests can communicate with one another, promote the ideas that they have, have ways to deliberate, discuss, and solve problems. And there's, like, these very for to hold up democratic processes, there's very, there's reasons why you wanna have these publics support that form of communication to bring about new knowledge and deliberate in ways that, was not otherwise seen prior. And, JCR Licklider, who was one of the early Internet pioneers, basically took this idea of emergent publics and said, okay. What are the ways in which we can actually build this system? And had the initial vision for ARPANET, the precursor to the modern day Internet, be based off of many of these many of the similar principles of emergent publics, have ways in which people can connect with one another in ways that they can deliberate, discuss, and create new forms of knowledge to hopefully solve more of the problems that humanity is facing. And I think what although JCR Licklider's vision was what I think we should we have yet to still solve, the reason why we haven't solved it is because of in part, by context collapse, which is when I share information, it's not preserved within the context of people who understand it. And so what do I mean by that? It's when I say something, I wanna have the confidence that you know the context under which I'm speaking on. And so not only do I wanna know that you have that context, I want you to know that I also have that context. And so it's like this rich form of common knowledge book context. And in prime example that we kind of alluded to earlier with Twitter or other social platforms is that if I send a picture of, you or any really, like, out of you like, some something even from this conversation I shared outside of the context under which I was assuming I was speaking under, that information can be misjudged, misinterpreted, and create all these different issues, which in specifically in teens today on social media, you're seeing causing anxiety, isolation, because people are misjudging the way that information is being shared. So I hope that gives a bit more context as to why, one, we care about context, and two, as to why context is still not being preserved on the already existing Internet platforms. And we haven't even got to AI yet. Like, it's already a significant problem. Yeah.
Yeah. So so there's this so we set up this idea of context collapse. I wonder if you can say maybe a few more words about because I actually I wanna make sure that in this conversation, we respond to a particular worry, which is that which is that context is problematic. So, for example, if we have if you have a, if you say something on Twitter, right, you're saying it in a really, really wide context. Anybody in the world, anybody in the future of the world can read it. And, and so in a way, that's your audience. You're speaking to this kind of you're speaking to everyone at the same time. And when you're speaking to everyone at the same time, there's this kind of, lowest common denominator in terms of, like, the language you choose because there's you know? And, and, and, yeah, if we imagine something different, if we imagine sort of, you know, smaller context, private circles of people talking to each other that aren't in that global context, there there are there are worries about, about that. Right? I mean, there there are worries about, you know, well, okay. If we if if we have, like, imagine if we, you know, fragmented into five different political camps, and then within each of those five camps, people are sort of developing their own, their own language. You know, one camp from the perspective of another might look like just a total, like, dangerous cesspool of of, you know, or something. Right? So, so how do you now I think there's, like, a real synthesis between these two points of view. There's, like, a real, bridge between these two ways of looking at things that feels really, really important for me, to to get across. So I wonder if you can say a bit about that.
Yeah. It's interesting. So I think, like, the point here is that, like, by heightened context and communication with one another, insular groups or bad actors could potentially have power in a way that maybe we don't currently allow for today because of the broad form of communication. I'm not sure I feel as strongly about that because if you do wanna communicate in a non public and private form and form these insular groups, you can if you want to. And we're seeing that with a lot of the messaging platforms today. Like, as an example, WhatsApp WhatsApp and Signal are probably the best prime examples of that, where they have made strong conscious efforts to not break their end to end encryption policies despite various different governments, most recently being The UK, wanting to, disrupt those, you know, policies due to fears of anti terrorism or child sexual assault material. And so I think that WhatsApp and Signal have taken on the ideology at least that we have these properties we wanna preserve in human communication. We understand the repercussions that come with that, I e, enabling insular groups to communicate in this heightened and strengthened way, but we're willing to uphold our principles at the cost of being used in The UK. And so they said, we're willing to be banned from The UK. WhatsApp has made that statement now. And so I think that's, like, one interesting part. But I think, like, you said something earlier that caught my eye, which was, is full context needed? And I I I guess I wanted to try to address that, which is for we never have complete knowledge of someone, I guess. Like, I I don't know even if I live with you and I spent my entire life with you, I don't have complete knowledge of how you think. But I think that, like, there's a certain amount of information asymmetry, which has allowed life to still exist despite that information asymmetry. Right? So, like, I still somehow the NFL game still plays at 7PM despite us not all having complete knowledge of each other's intentions or dispositions. But I think what's happened is that that asymmetry, as a result of the Internet, has started to grow to a point in which maybe we're not able to control and actually understand and function in a way that we have for the past x number of years. And so I think that we're basically increasing that asymmetry to a point it has become, like, stable. Humanity can still grow and progress and build new health care infrastructure education systems, and we can communicate with one another online. But we're reaching a point of asymmetry where now we no longer have the context to communicate effectively with one another in these very broad forms, meaning, information networks like social media platforms in particular.
Yeah. No. I'm I'm I'm I'm with you. I just think there's a there's a point that I sort of want to come across, which is that I think sometimes people hear about people hear the worry about context collapse, and they think that on the other side of the spectrum is, like, fragmentation. And that's really not what this is about. Right? I mean, what what this is about is more about building these sort of, like, rich structures of context that are partly open. Right? So if you think about a if you think about, for example, a language, you know, I don't speak Turkish. So in a way, there's a context there that I don't have access to, but Turkish isn't close to me. I can learn it. It's like this it's this sort of, rich, like, structure of, like, the right amount of friction between communicative contexts that allows a diverse, interesting society to happen, essentially. So I just wanted to sort of Yeah. That across. So
I think then, whenever it feels right, we can start talking about plural publics, which is addressing this precise problem, which is not only do we want to establish and preserve context, but we wanna do so in a way that allows for opportunities like this Turkish example that you just gave, but without losing the context that was initially established amongst the people who developed that language in the first place. So, yeah, we can dive into that if it feels right, or we can wait until we get more to, like, the AI argument as well.
Yeah. Well, I'd love to hear, I'd love to hear how this connects to what you're thinking on AI.
Yeah. So I guess let's take a step back and look where are we in the in the AI timelines. One, there's no answer to that, but I think that we're starting to see sparks of I'm gonna avoid using the word AGI, and I'm just gonna use sparks of AI that can really damage social norms. And, like, that's what I think I care about personally right now. AGI is just, like, much, much more complicated, rich problem that, I think requires a separate conversation. And so right now, we're just seeing with things like chat GPT, GPT four. This week, the the time at which we're recording this podcast, we see Google's launched their language model, Facebook's launched their language model, and various different other, iterations of these tools are being released in a way that they're able to produce very persuasive human like content at the ease of someone's fingertips of type typing a prompt. And today, it's in the form of language, and we're starting to see multimodal models evolve to start to become more related to image and video. And what is starting to occur is some very dangerous norms in human communication. So let me outline what some of these norms are starting to become. The first is that you're seeing in education systems a deep reliance on GPT three and chat GPT already. And it's become this, like, actual crisis for educate educators today to determine whether or not a student is using GPT or not. And from a purely technical perspective, it's almost impossible. In fact, if you there's a tool that OpenAI, the developers of GPT and GPT four GPT chat GPT and GPT four called text classifier. And what text classifier does is it says, take an output of text, and we'll take that text in and determine whether or not an AI created it or not. And if you run some experiments for yourself, you will see it is incapable of determining whether something came from an AI or not. I copy pasted outputs from chat GPT, from GPT four, put it into the text classifier, and it was unable to detect that it came from an AI. So if the creators themselves can't determine what's coming from their own models, and the technical research is showing that it may even be an intractable problem to technically prove what came from an AI or not. Then what I'm going to get to here is very quickly a place at which I don't know where information came from. I don't know who created it. Was it GPT? Was it a human? And so information here has become decontextualized from its origins. And so this is, like, one of the first problems. We can't verify who created what. The second problem is that as we become more dependent on these tools, the ways in which we do communicate is gonna start to shift the distribution of of the natural language that we normally do. So let me rephrase that. Today, these large models are trained on a corpus that was created by humans. So I know there's a siren in the back. I don't know if it's picking that up. But,
right now just pause for a moment and then we'll okay. Now it's better now. So where I was was I was describing how the
distribution of human natural language is actually going to shift as a result of people using these tools more. And so as an example, right now, what happens is is that, these models are trained on humans' natural language as it relates to them typing online. So it's trained on Reddit data, Twitter data, and all these other platforms. What happens is is when I start using these models and communicating through the outputs of those models itself, then what I've done is I've basically poisoned the Internet now with outputs from the language model itself. And so when they go and retrain these models, it's trained on a distribution that's just slightly different from the human natural language. And if you take the limits of that, you get very far from human natural language actually. That is very scary. Right. That we are now no longer in control of our natural language. So there's, I guess, some very this is, like, some of the, like, higher level problems. And then there's just, like, very real attacks already occurring right now with these models, which is scammers are getting very intricate with how they're using these tools. So there was an instance very recently where a scammer took a video from TikTok of this, girl who was creating a video, synthesized her voice. They only needed ten seconds of audio. Synthesized the voice, called that girl's grandmother as though they were her, and said, oh, I'm in I'm in dear I'm in need of, like, help and all this other, stuff. And then the grandmother sent 21,000 US dollars to the scammer because they sound like the loved one of the of the grandmother. And this was an article written in the Washington Post, I think, on 03/05/2023. And if you read that article and you're not scared about what's gonna happen to your loved ones that they can no longer trust your voice, especially as someone who's probably had YouTube recordings of them already out there, talks of them somewhere made to the public, You are susceptible to these threats, not just through voice, but through image. And the limits of what is about to happen in the next year or two is quite scary. It's not just, like, scammers using your voice. You're no low you're no longer in control of your reputation. Who what you said where, who you it's gonna become very persuasive. Unfortunately, also probably very graphic because of how cheap it is to create this content in a way that we can control, and we can't assume that it's only gonna be OpenAI's models or Facebook models. We're seeing the UK government is put a put a bid up for to invest 900,000,000 into creating their own version of GPT. It's called Brit GPT. And so what the limits of this is all showing is that it's not just gonna be big companies. It's not just gonna be nation states, but everyone in their communities is gonna have their own version of these AIs that's gonna have their own set of preferences that may be really good for humanity in progress, but also may be very harmful and enable a lot of damage. And so this is, like, very scary for the way we communicate with one another online to the point where now maybe we can talk about how do we combat that issue.
Well, yeah. I mean, I I I do just wanna sort of sit with this for a second because I know there's we we are in this kinda crazy moment where people are there's so much, like, we're all just sort of coming to coming to grips with something, that is really, really, really pretty crazy. And, I mean, my my, personal philosophy about this is that I don't I don't, there's a really difficult balance to to to try to strike between being, like, too alarmed and not alarmed enough. It's, like, really, really easy to miss by a lot, you know, in terms of being too alarmed or or or not alarmed enough. And, I don't think any of us exactly know where we should where we should fall on that right now. But the problems that you know, the potential problems that you just raised are are are enormous, and there's more of them. Like Yeah. You know, for example, I mean, even if you just think about this, you know, you gave a couple examples of of of scamming, but it doesn't take too much creativity to kind of multiply those upon themselves. And, and, we could we could find ourselves in a world where, basically almost all information is is hard to or even impossible to interpret. Right? And that's what kinda comes back to to context.
Precisely. Yeah. It's, so, I mean, I guess the way I was speaking, I'm not even really speculating, and I that's why I was trying to I'm not really I'm speaking on what's possible today. Like, these are what these tools are capable of. And for context here as well, Reid Hoffman, who was the former, CEO and founder of LinkedIn, had this post about how he wrote this book with GPT four and how he started using GPT four last summer. So it's now almost ten months post when he started using this tool that is now available to the public. I can imagine what's already running internally at these others' teams that they're just trying to get these tools out there and ready. The capabilities today are actually behind what reality is at. Right? So I think that does speak to that. And I think, like, what you said, earlier, which is, like, what does the world look like as this all becomes more, proliferated? Well, you operate in, like, a zero trust environment. You no longer trust what is real or what is fake because the cost of deception tends to zero as these models become more and more accessible. And so how do we like, what what happens there? Like, what happens when people don't trust? Well, you get into this more, like, Web three blockchain ideology of you don't trust. You only need to verify. And so this is kind of where, like, the intersection of cryptography may become very powerful, but I'm still, can't underscore, like, how scary some of this stuff actually is. Not for, like, the AGI. Like, we're all gonna necessarily die from the AI killing us. Maybe that is possible. But even just, like, the very real, we can't communicate online anymore. We depend on that right now. Right? And so yeah.
Yeah. Yeah. And I have no, I have no words of comfort. I'm not in the words of comfort business. I mean, I think it is I think it's it's it's really it's crazy. It's it's important for, it's important for us to understand, the the environment, basically. And I I think what we can what we can do is is try to articulate, try to articulate new norms within a new environment and new ways of establishing context. Right? And, I know you've thought a lot about that. I mean, what what does that what does that look like? How can we, I mean, what you know, so one possible reaction to a kind of a full blast understanding of the world in which we now find ourselves is privacy. Like, maybe the like, basically, the I I mean, I'm old enough to remember the very early days of the Internet when, there was absolute panic about, about privacy and the people who are worried about privacy lost. And, they lost so completely that the conversation shifted, you know, about thirteen to ten years ago or so, into a new ways of dealing with the world in which we found ourselves because the old norms of privacy had been had been have been so completely obliterated that it it they they had very little relevance. And, and yet today, it's almost like we're back to that, you know I mean, one possible reaction is just to go back to a pre Internet notion of privacy, basically. Right? We want to want to keep certain information away from away from the public, away from the open Internet, away from, sensors where it can be fed into, into models that can interpret it in in ways that are essentially impossible to foresee. Or I shouldn't say interpret it. That's the wrong word. But it said process it and allow it to be used in ways that are impossible to foresee is a better way of putting it. So is that is that how you're thinking about this? I mean, like, how are you thinking about the, you know, essentially, the the reestablishment of of of context? Are you thinking about it through the lens of privacy? Are you thinking about it through the lens of cryptography and shared secrets? Like, yeah. What are you what's on your mind?
Yeah. So let me take a step back because I think I've sound a bit too pessimistic on this podcast. So I wanna shed a bit of, like, optimism here, which is that I think that although AI is very scary in terms of the impacts it could have, it could be the precise attack we've needed to fix the Internet since its creation. Like, we get this do over moment now. We actually get this do over if we can fix it. Right? Because no it's no longer, like, privacy, as you said, is like this, like, ideal thing. It's like we want privacy. It's like you actually need to have privacy preserving technologies if you wanna continue to use the Internet. And so, like, this is, like, our do over moment because no longer can you actually function. Like, we've been able to again, like, I talk about this, like, asymmetry or stability, whatever graph you wanna model. We've been able to operate without having these, like, foundations of privacy or a complete common knowledge. But all that goes away as it as the rates in which that occurs and, like, attacks to context and privacy go out the window because of AI. And so now you actually become, as a user of the Internet, in need of privacy preserving technologies. And so that is actually quite exciting because maybe we could actually fix the Internet in the way that we've never been able to because we finally have the threat we've been looking for. Now sounds a bit paradoxical. We're thinking of AI as being this, like, very bad thing, and it's gonna, like, help us because it's gonna cause all this damage. I wanna be clear there as well, which is that I think there's a lot of promise for that what AI can have for human productivity. And there's, like, ways in which we can actually really benefit from this. The ways in which we lose out is from when it starts starts to harm our social norms. And to allude to, like, the ways in which I'm thinking about, ensuring that we can benefit from these productivity gains, but not lose out on social norms that are very important to our relationships and the way that we function is with us, like, a set of various different types of technologies. So privacy is one of them. And, like, often what we mean when we say privacy, and Helen Nissenbaum has articulated this much more clearly than than I can, but it's we actually mean context when we say privacy. And she has this, book called privacy in context, which talks about contextual integrity, which is the ways in which we preserve and establish context. And oftentimes, when we talk about privacy, we're not actually talking about this, like, public versus private, like, atomistic view of saying, like, how our data should sit. We actually mean a much more fluid form of communication, which is information should be shared in under the context in which it makes sense for it to be shared. Because what's private to me and you may not be private to someone else, but it may be semi private to another group of people, And we should have these ways in which we can control the privacy or context on which information is shared. And so I think, yes, privacy is, very closely related to how I'm thinking about, we can not solve, but at least move forward into, improving our communication systems while both human generated and artificially generated content coexist. And I think privacy as we know it today is anchored on a lot of these cryptographic techniques that, have really built up our understanding of privacy from a technical perspective. And so, yeah, I think we can talk a bit more about that, but I think I just wanted to highlight, a bit of Nissenbaum's framing here to, like, context and why I think AI actually is almost hopeful to the we can actually build some of these ideologies that Nissenbaum has talked about for years, but we've never actually done anything substantial.
Yeah. I mean, one of one, thing that I wanna on the topic of, like, maintaining social norms that I wanna focus on for a second before we get into some of the more technical stuff about how to do this is, like, when we one of the most interesting things to me about the way that we relate to each other socially is that so much of the way we relate to each other socially is instrumental. Right? So there's, in other words, people use each other for various different towards various different ends. And one of the most important ends towards which people use each other is information processing. So in other in other words, we go to each other to help to help help understand things. Right? I mean, I I I talk to you to learn about information that I don't know that much about. You know? And we do this with everyone we know. We do this with everyone we know. We we, people we like, human social networks are, like, networked information processing devices almost. And, they part of the way that they relate to each other is they need each other for that Yeah. For that information processing. Right? So and so if we have suddenly machines that can do global information processing on, like, a planetary scale, there's an immense amount of this of that sort of instrumental value that human beings are giving to each other, which is an important mediating thing in social relationships that just sort of that just sort of goes away. Right? And, you know, it's such a paradox. It's such a bizarre paradox because what we want, like, the amazing thing we want, the the I mean, the the dream, like, the dream of a better world is a dream in which our relationships are less instrumental. Right? In which we're not using each other for narrow self interested things like, you know, I want to understand something, so I'm gonna treat you, Shrey, like you're a book, you know, and pull the information out of you. Right? That's actually not, like, the sort of full respecting of your subjecthood that, that that our higher norms really point towards. Right? And yet, if we weren't able to give each other these sort of these just this, like, basic basic instrumental sort of stuff about explaining things to each other, telling each other things that we didn't know. Yeah. It it like, the ground on which our human relationship, stands would would would would radically shift. You know? Yeah. And, so I set that up as a as a frame for for this larger question of, like, you know, the social relationships that we want to that that we care about, that are, you know, one of, if not the most important parts of what, you know, creates meaning in our lives and what gives us joy and what we what we care about, what we want to flourish. You know? Those are going you know, that's what's at stake here. Right? I mean, what's at what's at stake here is, is our ability for those relationships to to have a ground to stand on and to in order to be able to to to flourish.
Yeah. I I I couldn't agree more, which is I think, I think it's interesting, like, how we even trust one another now has changed. And I think it's so interesting when I read, interesting when I read, Geric Simmel, who wrote this piece in in nineteen o six, talked about how trust has become much more objective than subjective. Meaning, I can trust a person based on their financial status or job and not about their moral principles or dispositions about certain people. And I think increasingly online, we're seeing something very similar, which is, like, our trust in one another is changing and becoming much more objective. And I I can only imagine how some of the mundane, if you may call them relevant to, like, intelligence finding tasks that you just commented on are, but are actually crucial to our relationships. As an example, is me being here on this podcast helpful for my and your relationship? Well, probably because we're both having this, like, very this discussion that gives me energy to talk to you and, like, gives me the sense of, like, excitement that you're excited about my research and that we can contribute to this to one another. But from, like, a productivity point of view, from, like, Shrey's hours in a day could be better spent if an AI version of himself that was trained on all of his research just showed up as a deepfake and did all the podcast for him and did all the meetings for him. And Trey just went and touched grass all day with his friends in Toronto. And, like, that's it. Because he just needs to, like, hang out with the people, like, proximal here or, like, maybe he just does, like, whatever. He maybe he's just auditing all the AIs in his life that are, like, talking to whoever they are. This sounds, like, super futuristic, but I would not be surprised if it exists in the next three to five years, where you have an AI version of yourself just showing up to these mundane tasks as you call them. But maybe, like, participating in these, like, mundane tasks is actually important in some relationships. And it's like, how do we how do we know, like, when it is and when it is not appropriate? I think it's, again, relationship dependent. It's, like, context dependent. Like, I probably don't need to go to a meeting with 50 people in it, and I'm not talking. I could probably just have a summarizer, you know, give me all the notes back and have myself participate in my virtual, you know, AI version of myself if I need to and just audit that once a week to make sure I didn't do anything stupid. And, like, I think it's, like, figuring out the ways in which we use AI and the social norms we need to make sure we don't cross is gonna be very hard. But I'm hopeful that we can find ways in which we can do so so we don't lose the beauty of the feeling that I get when I speak to you about my research because I value our relationship, and I value the relationships with others. And so there's, like, this weird productivity also, index that I think maybe we're indexing maybe a bit too hard on when it comes to AI as well.
Right. Totally. Like, I mean, what another way of another way of, of calculating this or thinking about how we are calculating what we're getting from each other is that there's this there's this sort of complex, like, time discounting things going on. So, for example, when we when we, if we have if we're having a very instrumental kind of conversation, like a like a work boring work conversation, it just, like, has to I need to convey something to you or whatever. You know? I could I could send my AI to that. I could send the, you know, the I could send the artificial map to that conversation and just skip it. But the what what will be lost there is is some possible gain through our relationship in the distant future. So in other words, we have to start sort of basically pulling, you know, put pushing our our calculating minds into the, like, impossibly distant future to understand the utility, to understand the, you know, dry instrumental value of our relationships with each other. And I think that's gonna be, like, a huge challenge. It's gonna be a huge challenge for people to see why they care about each other, why they relate to each other when the obvious feedback about why they care about each other and relate to each other is gone and just push into this ultra abstract sort of space. Right?
Yeah. Yeah. I couldn't agree more. And I think it's like, for me, I've been thinking about this very closely as it relates to my relationships in my life, and I've realized, like, spending more time with someone outside of a professional setting, whether it's related to academia or professional work, it changes my perception of them when I do work with them because I see how they interoperate with their family or how they act around them or have just, like, more context to their life. Like, this is their story, and they're showing up to work. And we all have to do, like, similar tasks, but this is, like, their story outside of work. And the way I communicate with them is dependent a little bit on that story that I've had the privilege to know because of that context we've built up. And so the way we interact with one another is so, it's so, how do you say it? Like, fragile because it's like, if the minute you, like, cross that boundary or don't, abide by these norms because you didn't pick up on, like, a very personal thing I told you in confidence or an inside joke that is very common knowledge amongst all of us because you decided to have an AI show up today, the minute is then when I, like, don't trust you anymore. And, like, that trust doesn't come back easily. Right? And so it's very fragile too.
Right. And you I mean, you can imagine, you know, you can imagine, you know, very strong norms emerging here where it's, like, absolutely unforgivable to send your AI in place of yourself without disclosing it, for example. And I'm pretty sure that that is absolutely unforgivable. Right? I agree. I could not agree more. Yeah. Yeah. Like, and, you know, it's it's important to have these kinds of conversations honestly to help work out some of these norms, which we've never had to work out before. Right? I mean, I think, you know, the, you know, hopefully, we can do it fast enough.
Yeah. I think Hopefully we can. I do think I do think the norms the norms have we've gone through some of these norm changes as we have in the Internet. The way we sign off on emails or, like, the way we have a subject title and the way we, like, do these, like, very obscure things in human communication if you integrate over all time. We figured out ways that was, like, kind disregard stray or, like, all these other things that are, like, now the norms of how you send an email. And if you don't do that, it's, like, weird or you're being rude or, like, there's these expectations we have of, like, what something looks like. And I think we're just yeah. You're right. Like, we're just gonna have a new ways of, like, what that looks like with AI.
Right. Right. But we've gotta move quickly. And what I mean, one of the challenges, like, I one of the, challenges that I've worried about in different contexts for a long time is that, you know, norms take time to evolve, and, technology moves really quickly. And the the technology that we're talking about now is moving, like, unprecedentedly quickly. And and so, you know, let's just hope it's not like a losing battle to keep you know, to to get our norms into shape. You know? Yeah. But, you know, there what but what can you do? We need to fix our norms. We need to think about all this stuff much more, and we need to, develop some of the, you know, the other ideas that you've been working on. We'd love to hear about, about your thoughts on Yeah.
I I I'd love to share with you, work that we've been doing called Plural Publix. This is work, Yeah. Joint joint work with people who listen to this podcast probably knows Glenn Weil and Divya Siddarth. We coauthored a paper called Plural Publix, which should be coming out around the time this podcast, I assume, is around this, like, time frame of coming out, which tries to answer this problem of how do we preserve and establish context in the era of AI. And the core of it is, first, like, what are we looking for in a communication channel? And we try to outline what exactly that that is. And we would call these communication channels publics. Publics, again, emerging from the John Dewey esque, influence here, but also publics here being this, like, academic word that people use to describe a space in which people communicate. We define a public to be a communication channel in which everyone knows that everyone knows that everyone knows that they have the context to communicate in a given space, and they all have that context to communicate. And so it's common knowledge that they show up into a group chat or that they show up into a given platform, and they know that the other person they're talking to has that context. And so they can communicate with one another. Okay. So that's, like, the goal, and plural publics here is, like, the goal of having many such publics. It's like, how can we ensure that not just one publics or one insular group or one really positive group has this communication channel? How can we ensure that they can intersect in all these, you know, really rich ways that Dewey imagined emergent publics could have such that we can get new forms of knowledge, but ensure that the knowledge that is created maintains its integrity from where it was created in the first place? And so in the paper, we reference various different types of technologies that we believe could help us get there. We're so early in this understanding of trust, deception, and authenticity. The tools we suggest may even undermine what we're trying to achieve, but we believe with the research we've done and our strong set of collaborators, that for right now, this is what we're gonna be working on, which include, group chats is the first place. And so Jared Lanier, who is one of the researchers also at Microsoft, has this piece called how to fix Twitter and all of social media, and it's written in the Atlantic. And it's just, like, great piece that talks about why screw all this, like, social media broad based communication. Everyone should just get into a group chat. Why is that the solution to human communication online? Well, when I get into a group chat like, let's assume here the group chat is, like, less than to eight people, like a smaller size group chat. When I say something that's mapped back to my phone number, which today we can assume most people only have one of, There's a sense of reputation cost to when I communicate. So if I communicate in, like, a very bad way that doesn't abide by the norms of that group chat, everyone's like, who's this guy? What's he doing? This is so weird. Get out of the group chat. I don't wanna talk to you here. Leave. There's this sense of cost to communication that we don't get in these other broad form communication networks. The other part is that when you enter a group chat, there is this sense of, like, assumed context or imagined context that you have with one another that we wanna try to double down on in the ways in which we can communicate in the future, which is, like, let's preserve some of these ideas that when we enter a chat, we we don't just assume, but we can, like, verify that you have that context. And, like, how do we verify context? Some ways, it's gonna be, like, very easy. It's like, are you a doctor? Okay. Verify that you have, like, an identity certificate issued by a medical school to enter this group chat. And therefore, now you can, like, communicate in this chat. Or it may even just be like, did you go to dinner the day we talked about this inside joke about the cup of tea? And, like, can we have it attestation through the form of of a certificate that's, like, very easy to issue from all of our friends? And so what we like we first reference group chats. We then talk about identity certificates and, like, I certificates more broadly, which is why we need to get more into this, authenticity and verification realm as the cost of deception increases. And then, some other types of technologies we allude to are blockchains, why blockchains are specifically well suited to create common knowledge as, there is some very rich, like, theoretical proofs there, from written by Joseph Halpern and Rafael Pass that talk about why blockchains are uniquely suited to create common knowledge on information, and they can be done in a privacy preserving way. And the final set of technologies we reference is, deniable messaging, which is there's this, like, really amazing property of human communication that we get online when I show up to a coffee shop and talk with you. It's, like, kinda awkward if I said, hey. Can we record this if I'm just catching up with you in a coffee shop? Like, I don't wanna have to record my, like, everyday conversations that I get in person. I kinda wanna ensure that what I share is just between me and you and nobody else, and that is what the a deniable message is. And so we've been working on different types of deniable messaging techniques because we believe that the way we preserve context is only have the people you want to understand and verify a message be able to do so, and everyone else can assume that they're fake as well. And so you preserve context amongst the people who should understand it, and no one else should believe its authenticity. So we in I guess, to summarize here, I've outlined, like, a bunch of different technologies that we believe are gonna become critical to the foundations of human communication. Us at Microsoft Research are working on building out these tools, and we're working with a set of collaborators from Harvard, Berkeley, MIT, people at Radical Xchange and the community writ large to help both whether it's from the technical perspective, whether it's from writing about these tools or doing podcasts on it, find a set of people that can, like, jump in and build on these products so that we can, like, embed them in various different ways. We may even be wrong. We don't know, which is why we need to, like, keep testing it with with everyone.
So can we talk a little bit more about, sort of, deniable messaging? Because I think that I think that's I think this is very interesting. I would like more people to understand this. There there's a or why don't I let you set it up? Yeah. I mean, can you what what what is that?
Yeah. Okay. So, I'll give the, higher level example first, and then I can give, like, the cryptographic example after. So deniable messaging or, what an an instance of this in cryptography is called designated verifier messages is when the only person who's convinced by the authenticity of a message is who it's sent to. An example today is if I text a message to Matt, then I only want Matt to believe that I sent that message. So if Matt goes and shows his phone to, let's say, Glenn, then Glenn should not believe that Shrey sent that message in the first place because Shrey only wanted Matt to believe the authenticity of that message came from Shrey. So once again, it's like the authenticity of a message should only be able to be verified by the designated verifier. And anyone else outside of the designated verifier should not be able to prove the authenticity of that message. And so that is what, like, a deniable message is. There's many ways in which you can implement it. One of which we've been looking at is through designated verifier signatures. And so we've released some code as to, like, how people can fill this out. But maybe for now, we can focus on, like, why do we care about that and, the challenges that exist with it.
Yeah. I think it's it's really it's really inter I mean, it's I I think it has well, I I find it real fascinating and and also wildly counterintuitive in a way. It's sort of like an inversion of how information has always worked. Right? Because what it is is that and let me let me try to say some things about it, and you and you need to sort of correct me if I'm if I'm getting it wrong a little bit. Like like, the problem that that this kind of messaging could solve is that when messaging when information is shared within a context, it only takes one prick of the contextual bubble for that context to be destroyed. Right? So anyone in a group chat can just take the whole group chat and and share it with with you know, post it online, and now poof, context gone. So this is why it's really, difficult to maintain context in our current environment because it's only as strong as its weakest link. It's Exactly. It's just really hard. But this is a potential solution to that problem. Right? This is a this is a now it it it whether it's a silver bullet, I'm not sure, but it it it it it's a it gives us traction on that problem in a way that, I'm not aware of everything else that gives us traction. So and what it what it does is it means that it means that if I send something if I send information to you and I want you to be able to verify that it came from me, You can do that. Somebody else can get that same information, but they won't be able to verify that it came from me. Now I think what might be confusing to somebody thinking about this for the first time is they'll think, well, okay. But won't that third person be able to know that it came from Matt anyway? Because it's written in Matt's style and it pertains to Matt and it's, like, stuff about Matt. So won't they be able to know anyway that it came from Matt? And the answer is no. But the reason why not is that, basically, I will be spamming the entire universe with, with false information from me at all times so that the only way for anyone to discern signal from the noise from me is for them to be able to verify particular messages in the way that you're verifying that the message came from me. Yeah. You you said it better than you said you
you you said it better than I can explain it. But I think that, the one thing I would add to the end there is, like, it's not just you spamming the Internet. It's, like, you and all the villains of you who are creating deep generated content about you. And, like, as such, there's, like, a bunch of, noise. And, yeah, you articulated that very nicely. So, yeah, you explained that well.
Yeah. I just oh, I think it's I I I think it's an amazing it it it's a potentially hugely important technology for all the reasons that we've we're talking about and, and kind of an amazing backwards upside down vision of the future of the information landscape, basically. Right? Where Exactly. We're all going to you know, everyone's going to be, like, just flooding the airwaves with, with nonsense precisely so that they are able to communicate anything to anyone. Like, it's Exactly. It's totally bizarre.
It is. Yeah. No. I'm, I'm nervous, but I'm also excited because I think that if we can get these technologies embedded into our systems, we do have promise to to hopefully fix what we've never been able to create so far on the Internet.
Yeah. No. I'm I'm with you, and I'm I'm wishing, I'm, I'm I'm wishing you the best in all the all the all the projects that you're that you're engaged in. I think you're doing I think you're doing extremely important work, really. And, yeah, I mean, there there's a there's kind of a vision of the future that is coming into focus for me a a a little bit of, you know, where we have to work really, really hard to create these these kinds of contextual bubbles. Or bubbles is really isn't the right word. I'm I mean, we're still kind of looking for a better vocabulary here, aren't we? Like, the, a a bubble is not the reason a bubble is not the right word is that it's too absolute. Right? There's, like, a clear inside and outside, and it fragile. It can pop. What we want to do is, you know, maintain more durable, more open textured, contextual, communication channels. And Yeah.
Yeah. I think we basically want this, like, social graph that we've kind of, fantasized over almost. Like, the social media industry at least has is, like, the world's social graph will let us be connected to one another, Facebook. But, like, I think it's really interesting and exciting to see that social graph now is, like, very, like, not durable. It, like, doesn't preserve context, but, like, we can rebuild this social graph in a way that's, like, it's gonna have a lot of bad nodes and a lot of bad actors. But maybe we can, like, draw circles around some of these nodes in a way that it doesn't, like, prohibit nodes from coming in or leaving. It just does so in a way that preserves the integrity of information within it, and people can feel confident in their communication that Gerhard Zimmer, like, wrote about because of that context that is preserved. And, like, maybe we don't maybe we're, like, not even afraid to post online anymore because, like, we feel as safe as we are when we get to sit in a coffee shop with our friends. Like, that'd be that'd be amazing. Right? So yeah.
Yeah. It's, it's also you know, it's kind of interesting to contemplate that the the the direction that you are articulating is a little bit a little bit different from the, well, what what what you're articulating is basically a way of maintaining more normal, more familiar kinds of contextual human communication and human relationships in a world where unprecedentedly powerful decontextualizing information processors are gonna be out there. And, that vision is attractive to me. I'm I'm in I'm in your camp. There there there's others there's other views of the future that are out there. Some of which are unambiguously horrifying to me. Others of which are, maybe there's something interesting there. I'm not quite sure. It's really, you know, it's really a wild, a wild intellectual, moral, spiritual, existential landscape right now. Yeah. And, I wonder if, like, you know, one fear I have is that is that some people will be captivated by some of these other visions and, like, won't you know, we'll we'll just sort of decide, well, okay. Let's just sort of surrender to a machine or, you know, something like that. You know, I I I think that those sorts of, there's a there are a lot of oversimplified reactions to the technology that are pretty dangerous, and I'm I'm hopeful that we can do, you know, everything we can in the next, you know, really months and years, to, to kinda win the argument about the way that the way that we should develop our, the way the civilization should should should develop and, relate to to technology. Yeah. But I wonder if we I wonder if you have any sort of, you know, final thoughts or topics that you wanted to make sure we got we we got on the table that we didn't get to yet.
Yeah. I I guess, I mean, I'm not too worried about not winning this argument because I'm I know everyone's gonna be affected by it, and they're gonna be looking for a solution. And I think that, like, a solution to this problem is one where people can be in control again. And so I think that everyone's going to realize these, like, very harmful effects. It's the dangers now come in from, like, how do we ensure that, like, we can govern these technologies and build them in a way such that we can promote the ideologies? And I think it's precisely what you said. It's like, how do we create this conversation with a group of people to create the narrative that this is what we want our relationship to be with technology? Like, we're actually dependent on it to communicate now because we, like, can't do so otherwise. And, like, all these other assumptions that maybe we need to write out, like, is the nation state a sufficient provider for humanhood because we trust a passport more than anything else? Like, what are these assumptions that we're gonna have to make about personhood, identity, and verification, that that I think are, like, very scary to think about that we can't trust some of these sources or we can. But I think that the more we can experiment with it, the better. Like, we don't really need a zero to one technological innovation. This cryptography has been around for thirty years. We just need to implement it in the ways that we are capable of and do so in a way that we all feel good about instead of, like, maybe over indexing on these, like, other aspects of AI. And, like, really focus on making sure that we can get these productivity gains gains without the cost that we talked about throughout this this episode. But I think, yeah, that's pretty much, like, the last thing I maybe I'd I'd add to that. Yeah.
I think it's a great place to to close. Shrey, I'm super grateful for your time, and, always a huge pleasure to talk to you. I always learn a lot, and today is no exception. So thank you.
Thank you. Yeah.
An immense thanks to Shrey Jain for joining us on this episode of Radical Exchanges and enlightening us on such momentous and ever evolving topics in the areas of digital communications, AI, privacy, and sharing his hope for building a better, more ethical future. Thanks again for listening. If you enjoyed this episode, it would help us if you could rate and subscribe to Radical Exchanges on your favorite podcast platform. This episode was produced by G. Angela Corpuz and edited and audio engineered by Aaron Benavides. The Radical Exchanges podcast is executive produced by G. Angela Corpuz and Matt Prewitt. If you would like to learn more about Radical Exchange, please follow us on Twitter at RadXchange or check out our website at radicalexchange.org. We also invite you to continue the conversation on our discord where we have a variety of channels discussing topics like what you heard today, as well as partial common ownership, community currencies, soul bound tokens, and much more. There will be links to all these in the description. Thanks again for listening, and stay radical.
