[Music] Welcome to another episode of Tech Unhinged where technology meets human. I’m your host Ravi Javeed and today we have Anand Rao an expert in communication education and debate with over 20 years of experience as a chair of department of communication and digital studies at University of Mary Washington. Anand has led groundbreaking work in AI pluralism and its role in education and technology. From co-founding the Millennial Speech and Debate Institute to shaping debate curricula at Harvard, his work reflects a strong commitment to critical thinking. Recently, Anant has focused on AI pluralism, offering fresh perspectives on the ethical challenges of AI. Also, congratulations on your new book and the center for AI and liberal arts. Welcome to the show, Anand. Great. Thank you so much, Ravi. It’s a pleasure to be here. Well, let’s dive into our first question. Anan, uh, you’ve dedicated much of your career to exploring the intersection of AI and communication. What inspired you to focus on AI pluralism and how does it align with your background in communication and debates? Great. Well, thanks Robbie. You know, as you mentioned in the intro, so much of my life has been dedicated to academic debate. Uh, I started uh debating in high school. I joined the high school debate team. I debated through college. I coached as a graduate student. I was even a debate coach as a faculty member for the first part of my career. And ever since I’ve spent over three decades in higher education and much of that has been tied to teaching argumentation and debate as well as courses like small groupoup communication and decision-m all of them looking at how do we test ideas, how do we develop, you know, human-based knowledge. And so throughout my career, I’ve observed that the most effective ideas pretty much always come from productive tension tension between diverse perspectives. Um it’s rarely just one person developing that idea. So small group communication, it’s not just one person doing the project. It’s working together collaboratively to be able to bounce ideas off of each other to develop new ideas, new perspectives. And I found that always results in a more fruitful product at the end. Now my interest in AI has been pretty consistent for a long time and been very interested in AI digital studies and what I’ve noted is that most AI systems are really single agent models. They take a very monolithic approach to development. Now the problem with this is that it contradicts most of what we know about effective human reasoning. You know we need and we really thrive off of an exchange of diverse viewpoints. Now, I’ve been really fortunate for the last several years to work with a team of other scholars on this and we have a research team that’s been meeting pretty consistently uh for the last two or three years. I’ve learned a great deal from them and this includes Devon Gier who’s a computer scientist, Stefan Bowser, John Hines, Alan Coverstone. What brought us together is that we all have an interest in AI, but we also have backgrounds in debate. We all debated and or coached and have studied argumentation. And we’ve really explored how you might apply some of those principles of rhetoric and argumentation to create AI systems that are more pluralistic. Now, I really need to give credit to Devon because he introduced me to the concept and the phrase AI pluralism as a as a term. And he’s our team’s computer scientist. He was a debater when he was young and he’s been very intrigued with the idea of charting out what AI pluralism can offer. So most of our work really explores how debate frameworks can enhance AI systems and they do that by enabling them to represent and engage with very diverse perspectives. U and that’s what we’ve been really working on and that work is now culminating in our forthcoming book. We’re working finalizing it and hopefully it’ll be in print later this summer or in the fall. Um and all of the contributions in the of the book uh also include you know these experts across computer science, rhetoric, education, argumentation and what we really explore is how these pluralistic approaches to AI development can create systems that reflect the complexity of human life, human values, our reasoning processes. How is it that we are able to work collaboratively? And it’s not just one perspective but it’s multiple perspectives. So you know my background in debate, it always comes back to debate for me. it naturally aligns with this focus. You know, debate fundamentally presumes that if you engage with different viewpoints, it will lead to better outcomes. And what’s interesting about that is that even in a situation, you know, where debaters can be very competitive. They like to win. So even if you’re in a situation where one person is really only debating because they want to win and the other person is debating because they want better ideas, you still get the benefits of engaging with those diverse viewpoints because you have to test ideas. You can’t win if you don’t have the better idea that’s better supported, that’s better developed and that really improves our thinking through that exchange. So similarly AI pluralism the intent is really to create systems that can engage in that kind of dialogue um both between AI systems and humans and between AI systems so that instead of just having one answer or having a monolithic approach you can have this more transparent and hopefully aligned AI that kind of reflects our own human diversity. Moving to our next question, Anand. So, you’re presenting at the AI4 conference this August on the subject of AI and the humanities through the lens of AI pluralism. This topic fascinatingly bridges the technical with the human side of AI. For our listeners today who may be unfamiliar with the term, how would you define AI pluralism? Why is it such a crucial concept especially in the context of rapidly evolving technologies like artificial intelligence? Sure. Yeah, that’s a great question. You know, this is something that I’ve been developing my own understanding of what AI pluralism can mean. Our research team is continually developing how it might be applied. I think fundamentally it’s an approach that embraces the differentiation of AI agents. And so much the same way that you know you and I have different perspectives and capabilities, experiences, we want to recognize that that should be reflected within AI systems. And the goal of this is that it should be able to enable genuine dialectical exchange. Um, and we need to have meaningful diversity in that kind of exchange. And part of what we’re also advocating for is that you need to integrate that into artificial intelligence development. We need to build AI pluralism as a foundation for the way that AI is developed and the way it’s deployed. Um, so we we talk about this in terms of kind of three key principles. Uh, the first is differentiation. As I mentioned, you really need to be able to create distinct AI perspectives. Um, you might see that if you go to a chatbot now and you can ask, “Hey, play devil’s advocate and can present some other viewpoints and it it’s pretty good at that.” Um, but ideally what you’d be able to do is have a different chatbot that could actually interact with you in that first chatbot. They could each have very differentiated perspectives, differentiated understandings, perhaps even training systems so that they can fully represent that diversity. Um so the second key uh principle of this is kind of this egalitarian approach. We really need to value different perspectives equally. Now this doesn’t mean that every perspective has equal import or impact. And that’s the same way with humans. You know we have some perspectives that are shared by some humans that we find aborant that we don’t think uh has a place within civil society. And so it’s not to say that you allow all perspectives to be equally represented and have equal impact, but it is important to recognize that there are different perspectives and we need to value that there are different perspectives in the foundation so that then you can have that dialectical exchange. You can have that debate to figure out which ones are the perspectives that should be represented in a decision that’s made for instance. And then the last principle is diversification. You really need to increase the range of perspectives represented. It’s not just two perspectives as maybe we think about in academic debate sometimes or in parley debate maybe you even have four different teams that are competing and that represents other perspectives. When we really think about uh democratic deliberation in civil discourse we’re looking at more than one, two or three or four perspectives. We really need to be able to have a diverse set of perspectives that are represented here. Um so in the book what we’re exploring is that this is not just a design function that should be toggled on or off. It’s not just allowing an AI chatbot right now to sell it, hey, you’re a monolithic AI. I want you to pretend that you’re representing four different viewpoints. Um, instead, what we’re really looking at is a foundational orientation. This has to be a structure of thought that fundamentally reshapes how we approach developing and deploying AI systems. Um, so part of this comes into even the way that you deploy them and train them. So, you know, I think a lot of AI models are being developed in a monolithic way as I mentioned. Part of that is because, you know, the systems and the platforms are gathering a lot of the data. They’re trying to train that model on something. What we really need to be able to do is recognize first that as some of the models have already talked about, some of the platforms have talked about different models have different applications and different uses. So, we already see a little bit of that diversity creeping in even unintentionally, right? they they know that there are going to be different uses for models in different ways. You know, I was listening to an interview, I think it was Mustafa Sullivan the other day that was talking about how when we have an AI agent maybe as, you know, an assistant, we’ll probably have a different one for at home or personal life than we’ll have for at work. That recognizes that we have different applications. We have different viewpoints, maybe even different values that we try to represent in that work that we would have at work versus what we would do at home. That is a very superficial understanding of it, but it starts to to develop into um really where we’re going with AI pluralism that we should recognize that you and I might each have different agents that take over or work with us in different ways and your agents are going to be very different from my agents. Uh and they’ll be very different from somebody else’s agents. So when we incorporate this into this architectural diversity and and I think that’s something that could really be built out and you know member of our team Devon has has charted out a lot of this and talked about some different ways this can work. What that’s going to allow for is very diverse perspectives to be represented. So you know an AI agent can then have a debate with another AI agent and they can have different training data. They can have different value systems. they can different perspectives and that allows for a better test of those ideas than if you have the same model pretending to play different sides because they aren’t able to do it as as authentically. Um, and I think this really reflects the way the humans interact as well. AI ethics is an ongoing conversation especially as AI becomes more integrated into the society just as we discussed. How do you believe embracing AI pluralism can help address common ethical dilemmas like bias, fairness, and accountability in AI systems? Yeah, an excellent question. This is something we really need to to delve into. You know, there’s a lot of work being done on AI safety. Some would say not enough is being done on AI safety and we need to address these areas. You know, our perspective on this and what we’re arguing for is that we think that AI pluralism has the potential to provide a really powerful framework for addressing some of these challenges. And part of that is because, you know, unfortunately when when AI models are being developed or at least in the past in some instances, they’re developed first and then there’s a question about safety and alignment and ethics. And so instead of treating ethics as an afterthought or attempting to create really a single unbiased system, which I think is ultimately impossible, you’re not going to be able to do that. Pluralism and this perspective embraces the fact that all perspectives have some biases. They have limitations. But by bringing it to the forefront, we can address that. We can identify those biases and we can deal with some of those ethical concerns a little more fairly. So, you know, for instance, for addressing bias, um, pluralistic systems make their values and assumptions explicit rather than hidden. Um, if you’re trying to claim objectivity within a model, what would typically happen is that they’ll take all of the different perspectives in the training data. They’ll try to represent them. Um, but it really gets compressed into one perspective and it’s difficult to represent those different perspectives. So, you assume that this is a more of an objective perspective presented by this monolithic AI. When you have pluralistic AI, you can recognize, hey, this perspective from this AI agent um has this bias. it has this political perspective for instance or it has this kind of cultural perspective and value set that it’s representing and it’s upfront about that you know it’s it’s kind of like talking about politics in class and I know this is always a concern um you know when it’s a college classroom or even uh maybe a high school classroom is the teacher able to be objective you know one thing I try to do is let students know I do have a perspective on this there are things that I favor and I don’t favor um that’s why I like to have debates about it because I’ll learn sometimes that maybe what I thought was the right approach is not the right approach. But I need to be upfront about what some of that bias might be. Now, this isn’t to say that we need to just then pretend that all bias can be managed through pluralistic models. Um, and it’s okay to have bias in the system. There are lots of things being done to make sure that we address bias in the training set. That should continue to be done. Um, they should make sure they’re looking to to ensure that the AI systems are representing that bias fairly and explicitly. But I think that could be done a little more easily within this framework of pluralism. Um then if you move on to something like fairness um there’s a real question as to whether or not fairness you know as a contested concept um can be identified. You know you have different communities you have different traditions they each define fairness a little differently. Is it equality of opportunity? Is it equality of outcome? There’s sometimes procedural questions various other frameworks. And so pluralistic AI can represent those different conceptions of fairness rather than trying to say that there is one single definition. this might allow for a more nuanced uh discussion about it, maybe a more contextsensitive application of what fairness might mean. Um, and then finally, when you get to a question about accountability, I think this is really central. And we we talk about maybe we’ll talk a little bit about explanability and and alignment. Uh, what accountability is really important for is that AI systems need to be held accountable for the actions that they take. And especially if we move into more advanced systems or even embodied AI, um if you have a self-driving car, for instance, if it makes a mistake, um we need to be able to figure out why did it make that mistake, how did that mistake occur, and what’s the accountability for that system. So within a pluralistic system, we can be offer a little more transparency and and that sense of explanability, which maybe we’ll talk about in a minute. Um it allows for a different kind of accountability. It’s a little more open. Um, and it’s a little easier to be able to say, hey, there are different ways of getting to these answers. It’s not a single correct answer. Let’s let’s dispel with that myth, but let’s make sure that we understand that there are different perspectives that can be represented and make sure that all of them are fairly represented and open so that they can be held accountable and we can contest those ideas. And you know just sort of related to this is my next question where Anand I was going through a Google research paper that discusses how AI systems often reflect majority views uh which can sort of unintentionally marginalize minority perspectives given the significant role AI plays in tasks like offensive language detection. How do you believe AI developers can balance the need for diversity in training data while maintaining fairness and accuracy in AI systems across different social contexts? This is an excellent question and I think what it does is it really highlights one of the fundamental tensions in AI development. Uh you know current approaches try to resolve this through statistical methods when they’re going through training collect more diverse training data maybe reweight some of the existing data to represent different p marginalized perspectives. But while those methods have some value and they should incorporate that in the training data, what they still ultimately do is they collapse all those perspectives into a single model. Um, and so even with the best of intentions and even with some great techniques for developing that training training data, when they ultimately collapse it all together and and some of it gets lost in that representation, then I think that you find that there will be more unintentionally marginalized perspectives uh that are there. So pluralism, I think, offers a different person possibility here. there’s another another approach that might be able to address this a little bit better. So instead of creating a single model that might try to represent all perspectives, it’s a little better if we can design systems that explicitly maintain the distinctiveness of those different perspectives. So if you have marginalized perspectives um that are currently marginalized these minority perspectives instead of trying to find the middle ground between them maybe what we need to have is a different AI pluralistic model and agent that represents each of those minority perspectives and so that way we don’t worry about hey how am I being represented within this um it’s not to say that they will always get their way or get the answers that they want but they at least know that there’s some transparency in their representation within the interactions between these agents. And I think that’s an very important first step for this. Um, you know, when you also talk about something like offensive content detection, what’s challenging about this is that, you know, you have to wonder how does what gets counted offensive in one context get counted as offensive in another. Um, there are a lot of different, you know, perspectives on this and, you know, we might have different thresholds for what we find acceptable or offensive and it might differ beond um on context, not just between individuals. And so a pluralistic approach would recognize that. And I think instead of training a single classifier that’s going to just reflect dominant cultural norms, we might develop multiple classifiers that will represent different standards, different cultural contexts. You know, when we even think about what might be offensive content when I think my kids growing up of my kids growing up when they were little, there was some content I didn’t want them to hear. I didn’t want them to be exposed to. As they get older, I think they’re they’re it’s okay for them to hear some of that content. So we have rating systems on movies and TV shows that just represents that there’s a different context in which we accept what’s we we identify what’s uh acceptable what’s not acceptable and so that insight about you know aiming for a single universal idea of offensiveness that’s inevitably going to break down that only represents one perspective I think pluralism in a similar way will be able to say here are different perspectives on what’s offensive here are different perspectives on minority uh opinions and contexts and By building out that structure and if you really build it into the architecture of AI so that you have these different agents representing those perspectives I think you have a better chance of maybe limiting that sense of marginalization or at least giving everybody a fair shot at representing their own perspectives. But then Cand when we talk about linguistics and feeding AI linguistics right that’s there’s a lot of room that goes to subjectivity but then you know when we talk about real world examples there have been you know higher rates um of facial recognition system failures when identifying individuals from minority ethnic groups right uh which is always leading to potential harm and misrepresentation given you know AI pluralism is here and we are talking about it, how the development of it can ensure that the voices would be heard or need to be heard and not only being heard but represented in AI technologies as well. Yeah, that’s that’s a very important concern. And you know, when we’re thinking about different minority perspectives or just even different uh you know, diverse perspectives within human communities, that’s part of what’s driving our our interest in pluralism for AI because, you know, conventional approaches to AI tend to reinforce a lot of the existing power balances that we see in human interaction. But we think AI pluralism can provide some really important mechanisms to help ensure that these marginalized communities and you know they’re not only marginalized within AI representation often they’re marginalized within human interactions. We don’t want to just repeat what’s happening within human interactions within AI systems. So we think pluralism can provide for more meaningful representation. Um and part of that is because it changes how we think about representation. um you know traditional approaches seek to include diverse perspectives and training data and that’s really important and they should be doing that but it often results in those viewpoints being statistically overwhelmed by majority perspectives. So if you train one model and you have to include every perspective and you include every different context well when that model then makes a decision or represents something or produces content um it can’t represent all of it within that content. it will try to find some common ground or find some middle ground between them. And I think that’s what happens when you start to marginalize those different perspectives. So as as we discussed in the book and one thing that we’re really exploring is that pluralistic approaches will instead try to maintain the distinctiveness of these different perspectives. Um so that way they can ensure that these minority viewpoints, these various viewpoints aren’t diluted or erased. Um you know and I think part of this is an element of contestability. So you know at a foundational level we think that you should be designing AI systems with a pluralistic p perspective. Different agents representing different viewpoints. Not every agent needs to include all training data in the same way. And even if you include all the training data you can say this agent is to represent this perspective maybe a certain political or cultural perspective or uh you know a social perspective representing a group of people. If you if you’re thinking about that for policym or applications um when that’s an important starting point, you also need to get to the point where there’s some contestability because it’s one thing to be included in the training and in the setup, but what happens when you actually employ these models? What do you do when you’re using these models to develop content to make decisions? And we’re going to get to the point where we’re making more and more decisions based upon the input and and really the work product coming from AI systems. So we need to be able to to develop AI contestable systems. Um and I think pluralistic systems and especially with this dialectical exchange can enable genuine contestability. So when an AI system makes a decision that affects marginalized communities if it’s within a monolithic system, it’s really hard to figure out if that happened, when it happened, why it happened, and what to do about it. But if you have a pluralistic model where you have different agents representing different perspectives, then those communities can meaningfully challenge those decisions and talk about, hey, our representation um occurred in this way within the setup of that dialectical exchange within that debate and we want to see when the other model argues against it. Why? What are the arguments made against it? How did that other model respond to the arguments made by our AI agent? So this moves beyond a real shallow notion of fairness toward more substantive ideas of kind of a procedural justice um democratic participation. It’s not just giving everybody a seat at the table. Um and that’s not true only for humans but also for AI agents but also making sure they have the ability to fully engage in that process. So when we get to a question of whose value should AI systems reflect, it shouldn’t be all values boiled down into one perspective. it it’s itself a deeply contested political question. It can’t be resolved just through technical means. We think that technology should be built to represent those perspectives and then you need to be able to build in the framework for that kind of dialectical exchange that exchange of ideas that debate so that you can really review and deliberate over those different perspectives and those ideas. So Arand given your role in education and it has been good more than 22 years now. how can integrating um this particular ideology into curriculara help students understand the societal impacts of AI and sort of equip them with the tools to critically engage with the AI technologies. Yeah, certainly. You know, I think traditional AI education often focuses on technical aspects and that’s certainly very important. Um, especially building up to the point of generative AI and really advanced technologies that we have now. And unfortunately, what happened is that they often treated those ethical questions as secondary add-ons. So, okay, we’re building it over here and then over here in a philosophy class, we’ll talk about some of the ethical considerations. a pluralistic approach. And the way that we’re approaching this is that we want to really integrate those ethical questions, the reasoning, the critical thinking throughout the educational process um in much the same way that we think they should be integrated within the development of AI technologies. So in our book, for instance, Alan Coverstone, who is an education expert and a background in debate, he has a really great study where he’s examining how pluralistic approaches align with Howard Gardner’s theory of multiple intelligences. So he argues that traditional educational systems have emphasized what he calls intelligence singularity. You know giving preference to a form of intelligence focused on mathematics and logic for instance. And similarly AI has largely been developed to excel in those domains. Um but they don’t really look at other types of intelligence. So embracing intelligence diversity in education we can help students develop a full range of cognitive capabilities. And that is you know kind of parallel with this discussion of AI pluralism. But by kind of expanding an understanding of how AI should be developed, how we should be discussing its architecture, its application, I think it opens up the possibility to discuss different types of intelligences, um, and to consider some of those ethical questions right up front. Um, I think pluralism also encourages students to engage with AI critically rather than just deferring to it. You know, I think it’s very easy for students when they start working with an AI agent, a chatbot for the first time. you know, they play around with it, they get it to write a haiku or a poem, uh maybe an email, and then they realize, wow, this is really capable, and they start to think that it won’t make mistakes, uh, and they’ll think that this is necessarily right, and they they’re very differential to it. And I think that’s that’s more than a little dangerous. Um, we want to make sure that we’re maintaining critical thinking skills. So when we’re teaching AI through AI pluralism, um what we’re talking about is that different perspectives, different value systems are represented within these models and that recognizes that not all of them are right, that there’s no one right answer, and that obviously they’re not all going to be um you know, completely consistent with one another because they’re different perspectives. So that means when you’re engaging with one, you should employ a lot of critical thinking skills. You should be engaging with it. you should be looking for arguments in favor of whatever is being proposed, thinking about arguments against. And so when students start with that as that as an understanding about how AI should work and how it it really is working, um I think they’re able to think more critically about that. You know, one of the the papers that we wrote uh in the research group, we referred to an augmented debate centered instruction and that’s really a model that will demonstrate this AI pluralism and it’s meant to work both for discussion of AI and also just within the classroom. Um what it allows for is that students start to engage with multiple AI agents that represent different perspectives and and they can work through this on very complex issues. So instead of just hearing from one perspective of the professor, they can hear from other perspectives of other students as well as other AI agents. Um that allows them to learn how to evaluate arguments, think about the evidence that’s being used and it’s much better than just accepting an answer coming from an AI chatbot. So can you sort of share any real world examples where a pluralistic approach to AI has been successfully applied particularly in education or public policy right now? Sure. Yeah, that’s a good question. Most of the discussion of AI pluralism is uh still theoretical in terms of development of of systems and employment in different ways. I think there are a couple of different ways to to look at this and to answer this. The first would be that within a number of AI models and platforms now there has already been some application of AI pluralism even if it’s not called as such u you know when we think about you know AI bots that are adapting to different contexts for instance so you know when Microsoft has co-pilot and they say well you’ll have one that you’ll use at work and another one you might use in your personal life they’re recognizing diversity of perspectives and different contexts that might be employed now they’re probably still working with more of a monolithic understanding of an AI model that it’s developed as one model that gets employed different ways, but that still starts to scratch the surface on what it means for AI pluralism. So more platforms, more models are recognizing that there are differences between models and that even when you deploy one model, you’ll deploy it in different ways in different contexts. And so I think that’s an example of how AI pluralism is in part naturally developing. Uh and I think there’s a a recognition that there is a diversity of applications, diversity of contexts. Um, and so what we want to be able to do is to be able to bring light to this to say of course that’s going to be necessary. We’ll be even better if we fully embrace this as a model to be able to identify what that might look like in terms of architectural diversity, what the benefits might be if you take that a little bit further and if we’re more transparent about what that means for representing different perspectives and different value systems. Um, now another example of thinking about this application would be uh something that a couple of the members of the research team that I’m on have been working on. They’ve developed a platform called debaterhub and it was meant to to teach argumentation and debate. But within this u Devon and John who’ve worked on this um Devon in particular has worked on some experiments where he has had um AI agents built where he’ll have multiple instances of AI uh models represented. So you’ll have one that represents the affirmative on a debate topic, one that’ll represent the negative. And they’ll each give a speech. Um they’ll produce the content. They’ll have evidence to support their arguments and their claims. they’ll engage in cross-examination between each other and then he has a third agent that’s built to judge and evaluate that debate. Um, and so he’s been able to build this out. He even has published and made available on GitHub and on the site some of the materials from some of these debates. So you can see how those different perspectives can be represented with those AI bots, those AI agents. Um, and how that pluralistic perspective allows for more meaningful debate. Uh and what’s really interesting about that, he’s also explored how you might even allow for belief revision. You know, in the may same way that if a debater has gone through a debate topic for a period of time, let’s say they’re a policy debater, they have a topic all year, the debates they have at the beginning of the year are probably going to be very different than the ones they have at the end of the year. And part of that is that they learn how to deploy those arguments in a better way. They find new evidence. um they’ve also probably changed their perspectives on what’s a good argument, what’s a bad argument, um what’s going to be uh the best thing for them to advocate for and the judges the same way. So, one thing that Devon is experimenting with and and is advocating for is to employ within those bots a belief revision mechanism so that if you have a bot that is um engaging in a debate, maybe evaluating a debate, evaluating perspectives or proposals, that the first time it goes through it, it’s going to learn about it and make a determination and maybe even judge one way. The next debate, it doesn’t just start over. Um it doesn’t just start fresh. It doesn’t totally reset. It’ll remember that first debate. So much the same way that a human would have that interaction, what it’s doing is it’s building out that training data to allow for the beliefs and the values that are being represented to be revised based on their interactions. Uh, and I think this will better represent kind of that full dialectical exchange um, and that experience that we want AI systems to be able to represent that humans are able to experience as well. I I see that all of these ideas they sound pretty compelling, right? But I believe that it’s not not without challenges. So Anand what are some of the hurdles organizations face when trying to implement a pluralistic approach to AI and how can they work to overcome them given that you know we we live in the real world and we know how it is and not everything is fair out there. Yeah, of course. You know, I think everything comes with an opportunity cost. And so, if you’re thinking about this from the organizational perspective of maybe an AI platform, you know, an open AI or anthropic or or one of those, there are certain processes that they’ve they’ve gone through, certain ways that they’ve implemented AI technologies and gone through that research. Um, you know, and I think that will take a little difference in perspective to figure out how this works. Now I think one of the benefits is if there’s an understanding that and maybe to explore this idea that by developing a pluralistic model they’ll end up with a better outcome you know better better represent human perspectives better alignment and safety issues that can be addressed um as well as producing better content that they’ll see that there’s a benefit there but it’ll take a few things so I think that there are a few challenges that I probably could identify the first would probably be technical you know when you’re designing systems to maintain the real genuine diversity of perspectives that’s a little more challenging than just be able to present one model that’s kind of a monolithic approach. Um you in the technical chapter in the book Devon outlines several approaches to this different ways that this can be done where you’re able to deploy different uh representations of perspectives uh you can talk about how different agents can be developed and and handled um and that their interactions much the same way that we were talking about he’s done with debaterhub can be deployed so that you know it will change the the way that they’re going about this uh in a technical way um but they’ll be able to see some real benefits in terms of what the work product will be and what the outcome will Now the second challenge is likely economic. You know when we’re thinking maybe not just about AI models and the companies that are producing those models but even those that are deploying them. You know the current models incentivize really a more authoritative AI system. Um and this is going to require a bit of a mindset shift because you know I think it depends on the philosophy of the organization. If you’re thinking about deploying a pluralistic model versus a monolithic model, I think the initial appeal for somebody that wants to run the their organization in a very authoritative way, the monolithic approach might be appealing because they’re thinking we want to make sure everybody gives the the same type of answer and that we’re able to develop work product that’s going to be consistent. The problem with that is that that’s not the way humans really work and that’s not the way most companies actually operate. Uh and so it’s important that that they develop an understanding for how pluralism um can allow them to develop better ideas on better products. So if you just have one person working on, you know, engineering the the design of a new product, it’s probably not going to be as effective or as good of a product as if you have a team of engineers. The same thing happens with AI systems. And so economically it might look like it costs a little bit more to run multiple agents, but really if they have better products, better outcomes, then I think they’re going to b they’ll recognize the benefits there and they’re more likely to deploy it. Um that also kind of bleeds into a cultural challenge. You know, some organizations uh have a really ingrained assumption about what a single best solution might be. If it’s a very top- down approach, it’s probably not going to be as appealing. Uh but I think this also comes back to a recognition that this is reflective of the way humans work more effectively. If we are able to collaborate, if we can test ideas off of each other, we tend to work better, we have better work product. The same thing is going to be true within uh an AI pluralism model. So it gets to this challenge of how do they evaluate? Um what is it that they’re really evaluating? Is it the simplicity of deploying something in which case maybe a monolithic approach looks like it’ll be easier but if they’re really looking for is what the outcome is going to be then a pluralistic model is one that’s definitely going to reap more benefits I think. Yeah. No definitely. So Anand when you know you look ahead how do you envision the role of AI pluralism evolving over the next decade? What impact do you think it will have on the development of AI technologies and their integration into society? Yeah, that’s a a great question and it’s hard to tell hard to be a fortune teller to know exactly where it’s going to go or what’s going to happen. I think at the very least we’re already seeing indications that most AI researchers, most platforms are understanding that there are diverse contexts, there are diverse applications, diverse perspectives. And so there are a number of of these concerns that are kind of driving an understanding of AI pluralism, even if they’re not talking about it as AI pluralism yet. The first is just thinking about how AI can be deployed. Even if you’re just thinking about it from a business perspective, you have to think about how it can be adapted to, you know, if you’re working with a law firm, that’s going to be very different than working in healthcare or working in education. That changes the context in which these AI agents are able to work. Uh that starts to give while superficial some understanding that you have a variety of perspectives that need to be represented within the AI agents that are being deployed. Um whatever it is that they’re going to be working toward. I think the second uh really force the the second pressure that’s coming in has to do with AI safety. Um and if we’re looking at alignment concerns, how are we going to make sure that AI systems can align with with human values and human perspectives? U part of that is a concern about how do we represent those human values in those perspectives? You can’t represent all human values within a monolithic model. It just simply won’t work. in the same way that that’s why we have uh you know exchanges between different cultural and governmental groups because we have different perspectives about what our own human values are. How do we resolve human alignment just within humans first uh let alone with AI systems? And so I think part of AI pluralism and understanding of representing those diverse perspectives is going to really push AI safety forward because it’ll develop a more robust understanding about human value alignment uh with AI systems. It certainly expands the idea of contestability and explainability. Uh it’s easier if the AI systems can explain through debate and argumentation what they’re doing and why. Uh and I think that pressure and once they see that there’s a mechanism that’ll help address some of those concerns, that’s going to help drive a lot of AI pluralism. Um even if they don’t talk about it in those in that same sense, there will be a drive to be able to exp increase explanability, to increase debate. Um, and you know, I think that helps support some of the transparency efforts that they have within AI safety. So, I think technically we’re also going to see the development of more sophisticated multi- aent systems in part because the technology is advancing so rapidly and there are the capabilities that it’s not just one AI that could be deployed. You know, for instance, if you hear some theorists and and some people talking about future, you know, business deployments, they don’t talk about just having one AI system that does all of the work. They talk about, you know, maybe you’ll have an AI system that can do coding, for instance. Well, you might have a hundred of those or a thousand of those or a million of those that are then each doing the coding. They’re each working separately. So once you get to an understanding that it’s not just one AI doing all of the work, it’s really an AI that can be it could be um you know uh copied, it could be um added to the workforce and you have multiple AI systems. Then I think you start to develop an understanding that hey while I do that each one can represent a slightly different role within that workforce. Maybe it also can represent a different context. Maybe it also can work in a slightly different way. And once you understand that idea of diversity representing greater work product, I think then you start to build into the technical systems that you don’t just develop one and then you give it different instructions. You start to develop different AI models that have different training and different context and different value systems. Um so that you get better work product, but more importantly you come up with better ideas. Um you’re able to come up with better outcomes and I think it’s going to be better aligned with humans uh in the long run. So I think AI safety, I think some of the technological developments are going to lead to this. And we’re certainly hoping that if we can advocate this and make sure people are aware of some of these benefits of how debate and group decisionm really benefits not only humans but AI systems that we can help propel that as well. Yeah. And I I believe that this particularly sort of puts responsibility on us that this sort of um you know data is accessible or uh being given awareness to the people who are at the core of AI implementation perhaps right so have you guys you know strategized or thought upon the next steps of after you know the book has done that how the message is going to get across and how those people who are already at the core of implementation they are probably just done with their studies at this point and now you they’re working in the industry, they would get to know about it and responsibly be taking care taking care of this aspect of AI. The I I think you really hit on something that we’ve talked quite a bit about. You know, I I think we stand at a really pivotal moment in terms of AI development for a number of reasons and it’s not just about pluralism, but just the choices that we make now are going to shape how these technologies are developed and how they’re deployed and it’s going to impact generations to come. So, we really need to be concerned about some of those decisions that are being made now. Um, that also is driving our incentivizing us and really driving our interest in getting the word out about this. And we I really appreciate the opportunity to join you for this podcast to help get the word out about it. Um, we’re trying to share a lot of the research that we’ve done with broader communities both not only in academic settings but also in you know technical settings so that we can talk to other technologists and computer scientists and researchers about how can this be developed. So, so as we wrap up, what message would you like to leave our listeners with regarding the importance of AI pluralism, especially as we continue to shape the future of technology and its impact on society? I think really the central message I would want to leave listeners with and and any of these other communities with is that AI pluralism isn’t just a technical approach. Um, I think that this is a framework. It’s a foundation that could be used um as a commitment to developing technology that’ll enhance human representation um that’ll enhance collective intelligence rather than diminishing it. Um and we face a really great risk that if AI becomes too powerful without developing some of these safety issues and mechanisms um allowing for explanability for allowing contestability um then we really are going to run some risks in terms of how that might align or not align with human values. Um, if we can approach this as as something that technologically can be done, but also the way that it can be engaged with and deployed, then I think we have a better chance of of using this in a way that could really benefit everybody. Um, and it would help from the technological perspective. It’ll help with, you know, the outcomes that’ll that’ll be developed coming from AI systems as well as human AI interactions. I think it’ll make it a lot easier for humans to trust AI, to be able to engage with AI if we view AI as understanding that it represents a perspective. It represents a context. And in much the same way that as like you and I have a conversation, we should recognize that our interactions with AI could represent a similar interaction, similar dialogue uh between diverse perspectives. And I think that we’ll have a more critical perspective but at the same time a more fruitful one so that we aren’t just differential or we aren’t just dismissing it um but that we can have real engagement with AI in the future. Yeah I know you know uh you’ve kind of left a message for our listeners as well but when if we particularly talk about the you know the engineers out there what particular message would you leave for them? For the engineers, I think what what I would probably leave for them is that it’s it’s important to think about how if you’re a developer, you could employ deploy pluralism as a guiding principle. Uh and it’s not just one for for the technology. I think it’s important to think about um especially as we have more uh technological capabilities now than we probably did even just a year or two ago. Um it’s possible to deploy this in ways that you probably couldn’t have even foreseen before. Um, and as you do this, it’s really about keeping an eye toward that interaction with humans. I think this is where when somebody makes a reference to a human in the loop. Um, I think that sometimes superficially just kind of means we want a human to make a decision before AI does anything. I think it’s more robust than that. I think what it really represents is that we should we should really employ a humanistic perspective on not just interacting with AI but also taking so much of what we’ve learned about a humanistic perspective in developing AI because I think that’s going to allow it to develop in in new and unforeseen ways uh and we’ll probably be more productive. I think it’ll probably develop in ways that that we’re going to be happier with the results. So you know for an engineer it’s important to think about the technological aspects of this you know multi-agent language systems um and also having different uh agents that can be include you know belief or vision and different training sets in different contexts all of that’s really important but I think engineers should also be thinking about how are they going to interact with each other and how are they going to inter interact with humans because if you’re thinking about that a little bit more expansively it’ll make the safety issues a little easier to grapple with um and it’s certainly going to result in I think AI systems that everyone’s going to be a little more excited to work with and I think that you’re going to be happier with the outcomes and I think we can say that within next 5 to 10 years AI pluralism might be a discipline of its own within the academia given certainly I I think that AI pluralism is something that is another opportunity that one thing that I love coming from my perspective in communication and rhetoric is that it provides an opportunity in academia to bridge some of these disciplines you know we need to be more interdisiplinary ary and one thing that I’m always excited about with AI is that it provides opportunity to use AI tools in an interdicciplinary cross-disciplinary way. It’s not just held in computer science. Uh it’s now at the point where we can incorporate it into daily lives. And so we should think about how do we deploy AI? How do we use AI? And AI pluralism, I think, is an opportunity to kind of further that interdicciplinary work. So perhaps as a field of study, maybe as an area within the study of AI, I think it certainly works well and and one thing we’re doing at our university for AI literacy to talk about AI pluralism as an understanding and a starting point for students to think about this is the way humans interact. This is the way humans think if we could use this framework to understand AI. I think that it makes it a little easier for them to understand it in a critical way. Um and it’ll also prepare them to be able to interact with AI in a more fruitful way in the future. Well, Anand, this has been an enlightening discussion. Thank you for joining us today and we are excited to see how your work continues to shape the future of AI and communication. Great. Thank you so much Rabbi. I really enjoyed it and I I appreciate the opportunity to meet with you and to talk to you about this. Uh, you know, it’s always great to have these discussions and I think that’s what’s central to what we’re doing with our study of AI pluralism is how do we take that opportunity for engagement, real engagement and testing of ideas to be able to further AI and to make sure that it’s something that will serve and work with humans rather than something that some humans might end up fearing. [Music]