Transcript
Speaker: Imagine a perfectly benevolent artificial intelligence. It governs better than any human government ever could. Its policies are wiser, fairer, and produce better outcomes.
Speaker: Would anything still be missing? That thought experiment opens our conversation with philosopher Lorenzo Manuali, whose recent paper, When LLMs Threaten Democratic Autonomy, asks what happens when artificial intelligence begins participating in democracy itself.
Speaker: Lorenzo's concern isn't simply that AI might make bad political decisions. In fact, assume for the moment that it makes very good ones. The deeper question is whether something valuable disappears when we stop collectively determining our own political future.
Speaker: In this episode of the A-Tech Philosophy Podcast, Sam and I explore two emerging uses of large language models. AI systems that facilitate political deliberation and AI systems is that attempt to represent our political preferences for us.
Speaker: Both might make democratic processes more efficient, but, Lorenzo argues, both can also threaten something valuable about democracy itself, our democratic autonomy.
Speaker: Here's our conversation.
Speaker: So to kick things off, maybe, could you tell us a little bit about a well-known thought experiment? ah it's It's where you start your paper. um It's a thought experiment by Adam Lovett and Jake Zuhl. Yeah. So you just tell us a little bit about that?
Speaker: Yeah, of course. So roughly the thought experiment asks us to imagine a society that is completely governed by a benevolent AI or an algorithm or like a benevolent algorithmic system.
Speaker: Probably not too important which way you cash that out. And the benevolent AI or algorithm governs society better than any humans ever could um on any outcome-based definition of better that you want.
Speaker: And so the question that we can then ask is, is there anything of value missing from this society? And if so, what is it? And Lovett and Zool's answer, as I read them, is that the non-instrumental value of collective self-determination or collective autonomy um is missing. And their specific version of that is what they call democratic autonomy.
Speaker: Okay, cool. And and so the idea here isn't like of the thought experiment, just to get clear. The idea isn't to show, oh it would be bad overall or worse overall to be governed by a benevolent algorithm, which might be true, but that's not really the point. The point they're trying to show is that you realize when you imagine being governed by a benevolent algorithm, you realize that there's still something really important missing. Like one way in which it's, um there's something negative about that situation is that ah the society's lacking in this thing of democratic autonomy. Is that like the way to understand what's going on here?
Speaker: Yeah, that's about right. um So I think that's completely right. at least as I interpret Leavitt and Zool, You can imagine that if a benevolent AI or algorithm you know governing all of us would save 2 billion lives, then perhaps we should do it. um In fact, I'd probably say we probably should.
Speaker: But just imagining the society in which that takes place, there's still this thing missing, this thing of non-instrumental value, democratic autonomy. um You could even weaken the goal a thought experiment, at least by my lights,
Speaker: um You could even say that maybe minimally its goal is just to show that democratic autonomy is like significantly valuable, not even like extremely valuable or something like that. I think all you really need is that to get um the thought that we should care about this thing um off the ground.
Speaker: um I think that Maybe different people's intuitions concerning how much of a loss of a society governed without collective self-determination or democratic autonomy it is is might differ, right? People might have differing intuitions about how big of a loss that is.
Speaker: But that's something of value is lost. And that it is some instantiation of the non-instrumental value of collective self-determination or collective autonomy, I take it as less controversial. I think that value is particularly important, but others' mileage might vary.
Speaker: Right. And just want to clarify one other thing, I mean, or double click on something you just said, it's got to be non-instrumental value, right? So someone might think to themselves like, ah you know, collective self-determination is good because it keeps us, you know sort of psychologically healthy and motivated and engaged.
Speaker: But those are all actually instrumental goods, right? I mean, it's because we can have better lives or or participate more effectively. um But so this is non-instrumental. So maybe another way to put it is it's inherently valuable, yeah worth um ah promoting for its own sake, right?
Speaker: Exactly, right. Any sort of instrumental, and any sort of, any, sorry, any instrumentally valuable thing, you could stipulate that, oh, the AI or algorithm, what if the AI or algorithm attains that better than any sort of human-governed society, right?
Speaker: So if democratic autonomy were only instrumentally valuable, in virtue of some other goods, you you could ask, well, imagine a society in which those other goods, in virtue of which democratic autonomy is valuable, are attained better by govern governance by benevolent AI or algorithm.
Speaker: Then what them what happens, right? So yeah, it has to be non-instrumentally valuable, inherently valuable, valuable in and of itself. um There are various ways to, yeah. Because because you could think of like, oh, um one instrumentally...
Speaker: one way in which autonomy is instrumentally valuable is that it makes me like more confident. Like if I get to run my own life, like i'm feeling more confident about myself. Right. and so but in this thought experiment, you will, you could imagine that the benevolent AI algorithm is like, it understands that and it wants us to be feeling confident. So it finds a way to like make us all feel super confident, even though we're not,
Speaker: democratically autonomous. So it could find a way of like securing that good outcome. um But still like the thing that's, that's lost is that we are not actually like governing the society ourselves.
Speaker: And the idea is that alone in itself is like a a big knock against that arrangement. right um But anyway, Roberto, I think you were going to ask that. the i mean i was going to basically note that we are a long way ah from ah a you know dictatorial... um algorithm, a piece of AI, but we are actually in the age of a certain kind of ah a AI.
Speaker: And that's what your paper is about. you're You're mentioning how there's at least two ways that you can use AI in politics like today ah that do threaten democratic autonomy, this thing we've been talking about that's you know inherently valuable. So ah let's talk about those two ah forms of AI, and they're both based on large language models, and it's facilitative and representative large language models. So maybe just tell us about each of those.
Speaker: Of course. One brief note before I do that, I think the dictatorial AI question might depend on who you ask, right? So, you know, how far we are. I mean, I agree with you. I don't think we're we're necessarily all that close to that.
Speaker: um I say that as someone who has less technical expertise, um has some amount of technical expertise, but less than others who have studied this a lot. But people Some people really do think we're we're not that all that far away, um at least in the people that I talk to. So um where the good case is something like, you know we're being we're pests being governed by... Anyway, just ah just to put... like a a you know People's mileage may vary with respect to that.
Speaker: But yeah, so i think there are two big ways. There are some other ways, but there are two big ways in which LLMs are being used in democratic processes. So facilitative facilitative and representative LLMs is what I call them in the paper.
Speaker: So, facilitative LLM projects are those that involve computer scientists or democratic practitioners using LLMs as facilitators in deliberation. Usually a formal deliberative context, like a mini-public or some forum where people are coming together specifically for the purpose of deliberating on a question.
Speaker: The LLM can do a bunch of things, so it can summarize the conversation so far, ask follow-up questions, interject with important facts, guide discussion in a particular direction, maybe keep it on core issues, etc.
Speaker: Mainly, all you need to do is really think about the role of a human facilitator in this kind of deliberation. um It's just that the LLM does it instead. Now, representative LLM projects are those that involve computer scientists or democratic practitioners using LLMs to speak for us in political context, roughly.
Speaker: Usually, to extrapolate our preferences over unforeseen policy options. So imagine that I give the like an LLM a bunch of my data, we can call an LLM representative, a bunch of my data,
Speaker: And then some body, maybe a political body, asks the LLM what I would say about a policy option that's not in the data that I gave to the LLM and that I never considered, but that the LLM can accurately predict what I would say about, either like by expressing my preference or by generating some natural language or an opinion about it.
Speaker: The LLM then asks my representative and gives my preferences and or natural language opinion about what I would say, And then in some versions, the decision-making body would use that as a binding um vote or opinion.
Speaker: In other kinds of proposals, it would be more consultative. um At least that's the most basic version of how an LLM representative could work. and There are more complex versions. um But that's the basics of what facilitative and representative LLMs are.
Speaker: And so just to kind of yeah restate what you just said. So facilitative is like a digital moderator. Maybe we can think about it like that. yeah And regenerative is where we make like a digital twin of Lorenzo and it votes for you when when it, um in in some kind of, ah you know, election, when it gets new information about, what you know, what's on the table.
Speaker: Yeah. At least like on ah the most basic proposals. Yeah, exactly. Exactly. And so, I mean, I think maybe a listener might be listening to this and and they they'll say like, well, that sounds, well, maybe at least one of those doesn't sound immediately ah so set off red flags. So how do these, you know, maybe kind of, ah we'll get into the details later, but briefly, how how do these threaten democratic autonomy?
Speaker: Yeah, good. So the basics, um the basic arguments are as follows. For facilitative LMs, the worry is about perspectival homogenization, which is a concept that comes from um these really cool scholars, Sina Faslpour out of Northeastern and Wolf Fleischer out of Georgetown.
Speaker: So there's good reason to think that LMs tend to output homogenized outputs that might reify dominant perspectives because of biases in the people designing, giving data use to train, and deploying the models across the AI lifecycle.
Speaker: For example, these individuals will have like a different demographic makeup from the country or the world at large. The way the disagreements are resolved between them might lead to certain kinds of homogenized outputs. And so LLMs' outputs are going to some extent reflect that.
Speaker: And empirically, I think the literature, on a bunch of empirical literature, I think bears this out. LLMs generally tend to homogenize um certain kinds of outputs. Now, if that's right, then the worry is that on average, LLM facilitators are going to tend to sideline minority perspectives.
Speaker: And the worry is that they're going to do this more than human facilitators do. especially. And that helps create perpetuate what are called persistent minorities in the democratic theory literature.
Speaker: For my purposes, we can just say that like, or for our purposes, these are minorities that don't get to see their preferences or opinions, um, causally contributing, um, or reflected in policy, at least a non-trivial portion of the time.
Speaker: They're therefore not joint authors of the law in the relevant sense. Or any authorship that they do have is impoverished. um Its value is impoverished at the very least. And so they don't participate in the value of collective self-determination as a result.
Speaker: They lack democratic autonomy. That's the basics of like the first argument. And then for representative LLMs, the basic worry is something like this. In order to have at least some democratic autonomy or more broadly collective self-determination, you and relevant social groups you're part of need to have your own projects and what are specifically called joint intentions that are causally connected to changes in policy that align with those projects or joint intentions.
Speaker: But in the main use case for LLM representatives, which is extrapolating your preference or opinions over policy options that are not foreseen to you, that you don't and won't probably know about, this is really difficult.
Speaker: So the usefulness of an LLM representative decreases the more it needs to consult with you about what you think you know about the things that it's going to vote on, right? And so the main use case as I see it is extrapolating preferences and your opinion over things that you haven't foreseen or that you haven't really thought about.
Speaker: So to the extent that LLM representatives take on that use case, that is extrapolating your preferences or opinions over unforeseen policy options, there won't be that causal connection between your joint intentions and any policy changes or outcomes.
Speaker: And so the value any joint authorship of the relevant policy outcome, should you have any, is pretty reduced. And so you would lack democratic autonomy. Does that make sense? Yeah, for sure. I think that's a good like kind of first pass on the argument. and And we can kind of unpack all those details. Yeah. ah and the next Yeah, for the duration of things here. But i think i I think everyone has on the table the sense of like, okay, the basic thesis is that you know these two ways of using AI in politics, this like facilitative AI and this representative AI, um those two functions, they threaten democratic autonomy. um And we'll kind of unpack exactly how you think that's going how that works.
Speaker: um But maybe, you know, ah we'll just to unpack your argument further, maybe we can just kind of start with the that that this key concept of democratic and autonomy. Imagine everyone kind of has a handle on it already, but maybe we could just, yeah, for for a little bit kind of zoom in on that. And so, well, I mean, we already kind of touched on this, but maybe it's it's worth just emphasizing right now the point of like,
Speaker: but like Okay, so if you think about democracy more broadly in defending democracy um or justifying it as a form of government, it seems like on the one hand, you could appeal to instrumental features, like the instrumental value it has, like the good outcomes it brings about.
Speaker: On the other hand, you could focus on sort of like non-instrumental, like intrinsically valuable features of democracy. And It seems like in this paper, you're talking about democratic autonomy, which is like something people use to show that democracy is like intrinsically valuable. It's not just that democracy is is great for the economy, in other words, right? Because like some people might say that, oh, it's really good for the economy.
Speaker: But you're kind of pointing out or or this this idea of democratic autonomy that you're using here is kind of like it relates to how democracy might be like intrinsically valuable, right? Anyway, just if you want to jump on that. um Yeah, absolutely.
Speaker: um So when trying to answer the question, like what is the best form of government, right? That's the question that, you know we're often trying to answer when we ask like, why democracy as opposed to like oligarchy, dictatorship, aristocracy, monarchy, right?
Speaker: um When we're trying to answer that question, what is the best form of government, and especially when we're trying to do so in favor of democracy, like you said, there are two arguments that are often given. The first are instrumental arguments, which state that democracy is valuable, and in fact, better than other forms of government in virtue of a certain standard that it obtains better than those other forms of government.
Speaker: um the standard by which we judge forms of government and for small d's, Democrats, supporters of democracy, you know, the standard by which we say democracy is better also needs to be a particular kind of standard. So what's often called a procedure independent standard.
Speaker: So the risk of question begging, we can't reference the procedures of democracy as part of the evaluative standard by which we judge forms of government, which are themselves at least partly constituted by certain procedures.
Speaker: And thus the standard is usually constituted by certain a certain set of explicit or implicit like outcomes, right? Does democracy achieve these outcomes better than other forms of government? And you know usually there's a discussion of like, why care about these outcomes as opposed to others?
Speaker: In other words, instrumental arguments justify democracy um in terms of its instrumental value, what we were talking about earlier. Non-instrumental arguments are democracy. i think also intrinsic talking about their intrinsic value is also a perfectly fine way to put it.
Speaker: um abstract away from any outcomes or benefits that democracy or indeed any form of government might provide. Instead, it argues that democracy possesses other kinds of value that other forms of government don't possess.
Speaker: Oftentimes, this value can be found in democracy's procedures. So non-instrumental arguments for a democracy are thus also sometimes called proceduralist. I think conceptually speaking, you can actually, well, you can definitely pull the two apart.
Speaker: um But this is why you often see this kind of overlap. Um, it's important to note that there are a few people, notably Elizabeth Anderson, who don't think that this distinction is all that sharp.
Speaker: So you might think that the non-instrumental value for democracy, for instance, is conditional on its instrumental value. Um, Elizabeth Anderson has this really great 2009, uh, paper arguing for something to that effect.
Speaker: um But that's the rough distinction that's used in democratic theory for your theory to give people their bearings. And I think it largely um is a helpful tool. So you have these instrumental arguments, democracy good in virtue of a particular outcome, and then these non-instrumental arguments that democracy is inherently or valuable in and of itself, or at least as procedures are.
Speaker: and And then in terms of its inherent value, I mean... what would What are some common things people say on that regard? I mean, like my sense would be, yeah, like a democratic autonomy. the like One way in which a democracy is inherently valuable is that it supports a lot of yeah like governmental autonomy.
Speaker: The people governing themselves, like that's a valuable thing, and like that's instantiated when you have a democracy, maybe... Maybe it's the only type of government that instantiates that. but But I think there's another type of value people appeal to, right? Like social equality. Isn't that another one? Anyway, I'm just kind of curious. is Are those like the two main ones, democratic autonomy and social equality or are there Yeah.
Speaker: Yeah, I think that's broadly right. um so But I think what social equality um can mean can vary pretty widely. So I think a lot of different kinds of arguments for democracy can go under that bucket.
Speaker: So it can mean relational equality. So for and for instance, Nico Kolodny argues that democracy is valuable in virtue of the fact that its procedures instantiate relational equality or rule over none is the is the famous name of the piece in which he talks about this.
Speaker: um it can also You might also think that political equality is something that is often grouped under social equality. um That's oftentimes talked about hand-in-hand with relational equality, but oftentimes not.
Speaker: There are whole debates about what political equality is, but in general, there's a thought that democratic prestige, whatever political equality is, whether it's something like actual equal actual influence over outcomes, like Dan Wodak says,
Speaker: or you know equal opportunity for um influence over particular outcomes. um At least this is these are very rough characterizations. um No matter what political equality is, democracy does a better job of instantiating it than other forms of government.
Speaker: So... Yeah, a lot of things can fit under social equality. um But yeah, roughly you can divide ah you can divide things up into concerns about collective self-determination, democratic autonomy. There are lots of different ways to flesh out collective self-determination to people like...
Speaker: um Christina Lafont talk about self-government. You can read Habermas in a particular way, I think, as talking about a kind of collective self-determinable. He's more conservative legitimacy.
Speaker: But anyway, I think that distinction is broadly correct, and though there are different ways to flesh out um those two kinds of values. And of course, there are other arguments beyond those two. The one that comes to mind might be, think there's some literature on the role of civic friendship in democracy, but Anyway, those are ah two very important arguments for sure.
Speaker: So maybe we can talk about someone who is skeptical about ah the value of democratic autonomy. Or I guess a better way to put it is that... um We don't really have it. And even in a well-functioning democratic republic like the U.S. of a And I'm thinking, I mean, just to kind of put put it in more concrete terms, is like um it's already the case that we have a winner-take-all system. And it's already the case that some people who, let's just say, vote, ah what's a third party? Yeah.
Speaker: Well, the Greens, the Libertarians, there's a, I think California-based one. I don't know how national it is. The Peace and Freedom, we hear about them over here quite a bit.
Speaker: ah So it's already the case that a lot of people, you know, basically never have ah their views put into law or or have their candidates elected. I think you I've been to... um conventions, whereas someone told me, like I've never voted for someone who made it into office.
Speaker: So ah what you know just I want to and you know get your ah take on this kind of stuff. Yeah, of course. So i think with respect to democracy at the level of the state, we oftentimes don't have as much democratic autonomy as one would like.
Speaker: I think you could probably try to tell a story on which even someone who's voting for a third party but takes he participates in other kinds of democratic processes might be able to have democratic autonomy or engages in a public discussion with a plurality that eventually leads to some kind of democratic autonomy. But I do think you're right that um at the level of at least modern nation states, I'm thinking here of like the United States represented elected democracies, we oftentimes don't have as much democratic autonomy as one would want.
Speaker: But first point, democracy doesn't just occur at level at the level of the state, right? Firms also engage in governance, as Elizabeth Anderson shows in her book, Private Government, I think pretty definitively.
Speaker: So democracy is also a value that should occur in organizations smaller than a nation state. And I think it's a more plausible ideal that democratic autonomy um be instantiated there, right?
Speaker: Second, even at the level of the state, I think there are lots of institutions that do facilitate democratic autonomy. So I think we need only look to how collective action and protests, civil disobedience, other actions from groups in civil society end up changing public policy to see that there is a substantial sense um in which democratic autonomy does exist in our societies.
Speaker: um Even outside of the sort of representative line um of democratic institutions, you could make an argument that there's perhaps some causal impacts from some groups of citizens through representatives on democratic outcomes as well, though there are definitely problems with, for example, regulatory capture that would have to be um discussed and hopefully fixed.
Speaker: This is a really good point. um This brought to mind workers cooperatives or what do they call it? Self-directed enterprises or something like that. So you're saying even though our political system at a national level might leave someone skeptical about how much democratic autonomy ah There is maybe, you know, in our workplaces, there's actually more autonomy than we realize. And maybe we can even use this line of argument to argue for worker cooperatives. Is this something that someone that's, ah you know, in this line of thinking would be for?
Speaker: Right. So good. I take it that it need not even be the case that we need to say that democratic autonomy does exist in workplaces. Maybe something you can say something even weaker, that it's a more a feasibly attainable ideal in workplaces because they're smaller, because it seems as though they're easier to instantiate processes that lead to groups of people having their joint intentions causally impact changes in policy um in smaller organizations like workplaces than really large and difficult to change organizations with like constitutions and whatnot like the state.
Speaker: Workers cooperatives, I think, are a good example of the kinds of organizations that plausibly possess more democratic autonomy. You might have other qualms about workers cooperatives economically, but I think that's something um that's really good about them.
Speaker: And so I think that democratizing the workplace is one of the ways in which we can instantiate democratic autonomy in ah in a feasible way in society at large. So consider um not even like creating, you don't even need to think about creating a whole workers' cooperative role. That's definitely one way to instantiate democratic autonomy and democratize the workplace. You could even think about proposals like, i think Isabella Ferreira has a nice proposal for a bicameral corporate model where in like one body of the corporation.
Speaker: um like The board of directors is appointed by two bodies, one of which is run by shareholders and the other by ah workers. right So that's another way in which you'd you could like democratize the workplace and increase democratic autonomy in the workplace.
Speaker: But the point is something like Workplaces govern their workers, right? in Some workplaces function very close to authoritarian dictatorships, right? This is Elizabeth Anderson's point in private government.
Speaker: And they have great control over people's lives, right? Their wages, their livelihoods, their healthcare, right? And they function like authoritarian dictatorships. And because they engage in governance, they should be subject to the same kinds of democratic constraints that other institutions that govern us, like the state, are subject to.
Speaker: Now, of course, there are different kinds of governance, right? The state engages in like you know all kinds, of course, of powers that workplaces don't have. So ah there might that might justify some differences. But that there are some democratic constraints, I think, is a ah really clear point.
Speaker: And so... If that's the case, then it's more easy to attain the the the ideal of democratic autonomy in the workplace. Then that should give us hope.
Speaker: um Not only because some workplaces like work workers cooperatives already do have this kind of ideal, somewhat instantiated, but it should give us hope that this ideal is not something that's merely something that exists in theory, but something that we can attain in practice.
Speaker: The philosopher in and sure in you as well, wants to define democratic autonomy at this point. So, yeah, I mean, obviously, anyway, yeah, you give ah you give a a nice definition in your work. So yeah i don't know if you want to go through that really quick for us. And then we can get to the AI stuff. But but it seems like, yeah, maybe we could do a little.
Speaker: Yeah. well definition here yeah oh absolutely Yeah, absolutely. I'd love to give a definition. So there are a few necessary conditions for democratic autonomy for a given group of people.
Speaker: The first is that they possess a joint intention. And there are, in turn, three necessary and jointly sufficient conditions for a given plurality of people to possess interrelated attitudes such that they constitute a joint intention. And here I'm drawing really heavily from Lovett and Zool, pages 477 to 478. I'm going to be quoting from them pretty extensively.
Speaker: So the first condition to possess a joint intention is that each member of the plural plurality aims at the plurality X-ing. The second condition, and again, this is a quote, each member of the plurality has a plan to contribute to their together X-ing, and these plans are jointly compatible.
Speaker: And the third condition, again, a quote, the members of the plurality are mutually responsive, meaning, quote, that most of them are disposed to execute their plan to contribute to X, just in case they think that enough other members are disposed to execute their own such plans for their own execution to facilitate the plurality X-ing, unquote. That's again from Levin and Swool.
Speaker: So those are the conditions necessary to possess a joint intention. Then, in addition to that necessary condition for democratic autonomy, there are two other necessary conditions that lead your joint intention, um so to that lead your joint intention creating what is called joint authorship.
Speaker: and in particular, valuable joint authorship, which ultimately democratic autonomy is all about. So there are other two necessary conditions to get your joint intention to create democratic autonomy.
Speaker: The first is what I call the right kind of causal chain condition. So Levitins will talk about this a bit in their paper. um But i sort of bring it out about a bit more forcefully and heavily in my own. And this is me.
Speaker: The right kind of causal chain condition is to realize joint authorship and thus democratic autonomy over given policy or political outcome P, a joint intention regarding P must be reflected in P and significantly impacted as the result of the right kind of causal chain, which at the very least non-deviantly connects the joint intention up to the policy outcomes.
Speaker: And then the second necessary condition for making sure that your joint intention creates democratic autonomy is what I call the joint authors at least some non-trivial portion of the time condition, which is to take part in joint authorship and thus democratic autonomy, a given individual or group must have its joint intentions realized a non-trivial portion of the time.
Speaker: can Can we say that last part in a different way as as ah a minority position is not systematically excluded? Is that sort of what is being meant there? Yeah, basically.
Speaker: Yeah, you can you can put it that way. Yeah, exactly. Yeah, I feel like continually, consistently fail. If the minority group consistently fails to influence policy and systematically, like maybe, then yeah, they their democratic autonomy is diminished or they don't have it maybe. Yeah.
Speaker: So, okay. So, so the three conditions, just to review for everyone, joint intention, something about like citizens, they have to have ah share an intention to accomplish some political outcome.
Speaker: Second thing, right kind of causal connection. Their joint intention actually has to like influence the resulting policy. And then three, the non-trivial realization, um, not, you know,
Speaker: Yeah. Okay. So, you so, okay, good. And then, um, all right. And then maybe now, I mean, I kind of, I'm, I've, I've questions about the joint intention thing. I feel like that might be, um, too strong of a criterion, but maybe we can, maybe we can get back to that later and we can kind of try to shift toward, uh, talking specifically about, um, the threat that LLMs, um,
Speaker: bring because basically what you're going to do is going to be like, Hey, here are the three conditions. And then each of these two types of LLMs, they kind of threaten one of them or maybe multiple, but at least one anyway. So um that's going to be like the setup. So let's, let's first turn to the facilitative. So,
Speaker: Yeah, so one type of AI usage you're worried about is facilitative LLMs. um And I was thinking maybe we should start by like trying to describe, like flesh out how facilitative LLMs might be used in a democratic context.
Speaker: So I don't know if you're cool with this, Lorenzo, but like is it okay if I kind of just like flesh out a little bit of a case and then we can you can tell me, like oh, that's that's non-standard. like In other words, I'm trying to flesh out something that like would be a sort of standard way of using a facilitative LLM, the kind of thing that you would be worried about.
Speaker: And does that sound good if I just kind of give some details and you say, yeah, that sounds about like the sort of thing I'm worried about or? Absolutely. for it. That works? Okay. Okay. Okay.
Speaker: All right. So yeah, the example is going to be Chicago. ah So like imagine Chicago has $20 million dollars to spend on public transportation.
Speaker: And they want to get the citizens involved in like the decision of how to spend all this money. And so they select a hundred rep randomly select 100 people who live in Chicago.
Speaker: Okay. And then they take that 100 people. They divide the 100 into smaller discussion groups. um And then what they do is they have an LLM.
Speaker: facilitate the discussion rather than a human being. Okay. And so what's the LLM going to do? ah Well, it's a facilitator LLM. So that means it's going to like moderate the discussion. So it's going to like encourage quiet quieter people to speak maybe or prevent certain people from dominating the discussion, just like a normal teacher does.
Speaker: It's going to maybe ask questions, maybe be like, hey, can you clarify what you just said? It might summarize things. So might be like, okay, here's a summary of the argument that Billy just gave.
Speaker: um It could summarize the disagreement that's in the room. um It could suggest a compromise. It could create a final report. um of like whatever happened at the ah at this little discussion group. all right. And so then you can imagine that the city the Chicago City Council, it takes the report that the AI created and it kind of uses it to help determine how to spend the 20 million.
Speaker: um so anyway, that's my attempt at like, this is what a facilitative if ai usage might look like. um Does that Is that the kind of thing you're worried about? is that Does that example capture it or is am i is there anything missing there? Yeah.
Speaker: Yeah, that seems like ah that seems like a good example. So to be clear, I think that such a process has like a lot of epistemic virtues and perhaps even democratic virtues compared to what might be done currently, which oftentimes might be just nothing, right? Just a plain old bureaucracy making a decision, um not consulting anyone about itt etc.
Speaker: And in fact, Eduardo Martinez out at the University of Cincinnati, a democratic theorist, has a really excellent article on the help that many publics and citizen assemblies can offer to bureaucratic agencies in their decision-making.
Speaker: And think he's completely right. Though he focuses on traditional many publics and not ones with LLM moderators, if I'm remembering correctly. So I take my intervention to be something like, this seems like great compared to the status quo. um Let's try to normatively assess what would happen if we use LLM moderators, right? As opposed to um human, you know, the regular way we go about this usually.
Speaker: Yeah, I can be a a cheerleader for the LLM usage now. Let's see. let's see ah So um someone might say, actually, I you know i don't even have to you know ah create an example. I moderate ah discussions all the time, right? I'm ah i'm a professor.
Speaker: And so I know that it's very difficult to get the quiet ones to speak up and to quiet down some of the loud ones. And so now I'm wondering, well... For me, it depends on how much caffeine I've had and how well I slept. And basically, my my skills are very variable from day to day.
Speaker: But an LLM can execute these functions at least more consistently than than I can. So ah what's so what's what's wrong with that kind of... ah Why not choose the LLM over a human?
Speaker: Yeah, great. So it could be that there are possible benefits, right? Like the fact that LLMs are not sub
Speaker: Lorenzo readily grants that an AI facilitator could have real advantages. It doesn't get tired. It could potentially moderate thousands of conversations.
Speaker: And with careful design, it might even perform some parts of the job better than a human. His concern is more specific. Large language models have what researchers call a tendency toward perspectival homogenization.
Speaker: Disagreement and diversity can get flattened as data are collected, labeled, averaged, and ultimately transformed into model outputs.
Speaker: In an ordinary setting, that might simply make an answer less interesting. In a democratic setting, the stakes are higher. ah moderator helps decide which questions get followed up, which disagreements deserve attention, and how a conversation is eventually summarized.
Speaker: If that process systematically favors dominant perspectives, minority voices can disappear without anyone explicitly deciding to exclude them. And that brings us back to democratic autonomy.
Speaker: Yeah, so here are some issues that might occur with democratic autonomy with respect to facilitative LLMs. So in the AI lifecycle, right, which consists of design, development, training, deployment, etc., Xenophaslipur and Wolf Leischer, so Xenophaslipur is out of Northeastern, Wolf Leischer is out of Georgetown, identify what they call the ethical epistemic problem of perspectival homogenization, which is defined as, and I'm quoting them here, when an aspect of an AI system's design, evaluation, or alignment processes unjustifiably diminishes or eliminates disagreement and diversity of perspectives.
Speaker: So Fossil, Florent, Fleischer talk about a variety of points in the AI lifecycle in which perspectival homogenization can take root. So first, humans are often annotating the data that models are used um to be trained and that yeah or in their testing data, particularly in post-training processes like you know reinforcement learning with human feedback and other kinds of um post-training processes.
Speaker: But humans, the annotators, can disagree. And the problem is that when annotating data that can then be processed by machine learning models and or other kinds of artificial intelligence has this kind of disagreement, and here I'm quoting from Paul Zuboff and Fleischer, typical methods for resolving annotated disagreement like majority voting and averaging end up privileging majority opinion and suppressing minority viewpoints.
Speaker: And so as a result, you end up... And here I'm not quoting anymore. But and as a result, you end up with these data sets that end up privileging majority viewpoints or these you know preference models that ah privilege majority viewpoints that are then the LLMs are post-trained on.
Speaker: So that's one problem that sort of leads to the to the um privileging of majority or dominant perspectives. There's also a lack of diversity in the annotators themselves, right which leads to biases.
Speaker: right So there might be, and here I'm quoting, selection and entry mechanisms that tend to hinder the participation of relevant perspectives in the AI lifecycle from the outset, which leads to a less diverse set of experts, annotators, or evaluators.
Speaker: And then there's just like not a lot of documentation of disagreement. um And the absence of such robust documentation practices might, here I'm quoting, obscure the extent of contestation within data sets used in training and evaluation.
Speaker: So that sort of reifies the disagreement or makes it makes it more difficult to see that there's disagreement in these kinds of data sets. And I think empirical evidence does actually bear out that LLMs are perspectively homogenized. I think there are at least two good pieces of evidence for this.
Speaker: One is the literature on generative monoculture, um which is a tendency for LLMs to significantly narrow, here I'm quoting from ah paper the the main paper on this, Wu et al., to narrow, quote, model output diversity relative to available training data for a given task.
Speaker: So you know basically, there's just like a narrowing of the kinds of diversity of output that you would expect given the given the data So that's generative monoculture.
Speaker: And then there's good evidence that LLMs homogenize by flattening, quote-unquote, identity groups. This is a paper by Wang et al. So LLMs are prone to essentialize identity groups and their opinions in ways that overlook intergroup diversity and complexity.
Speaker: So I think that's good empirical evidence that there is this problem of perspectival homogenization in LLMs. Now, if that's right... and the empirical evidence bears that out, then the problem is that as facilitators, LLMs will end up in how they ask follow-up questions, in the kind of how they guide discussion, in the ways in the questions that they ask to people, in the facts that they bring up to people, um in the summarizations that they make of certain conversations, are going to end up marginalizing or sidelining minority viewpoints or perspectives.
Speaker: And then the problem with that is that it undermines ah the... condition in which, in order to have democratic autonomy, we need to see our joint intentions causally impact the policy preferences, or at least reflected in the policy preferences, a non-trivial portion of the time, right?
Speaker: So sometimes we need to see our preferences reflected in policy and particularly through a particular kind of causal chain, right? But if the minority perspectives are continually sidelined, it's less likely that that's going to happen. And so what can happen is, the worry is that LM facilitators will create or maintain, at least without further action,
Speaker: ah Persistent minorities. So there's the big literature and democratic theory on persistent minorities and the problem persistent minorities, minorities who don't end up having their voices ah heard in democratic processes and don't partake as a result in the value of self-governance.
Speaker: um In this case, in this paper, persistified as democratic autonomy. Does that make sense? I think it does make sense. um I do want to, um i mean, we love the empirical here at ATEX, right? But um I do want to kind of take these specific concerns that you have and apply them to the Chicago case.
Speaker: Okay, great. Okay. Yeah, so this is this is really helpful. Let me know if I'm ah misreading you in any way or misunderstanding you. So I think the the Chicago case is interesting for our purposes because it's a one-shot kind of thing.
Speaker: I think... we can best see how facilitative LLMs might undermine democratic autonomy if we consider like a multi-shot kind of thing. Because it's always going to be the case that, you know, sometimes you're going to lose, right? as a As in any democratic process, right? At least ah as long as there's not a tyranny of the majority going on, right? Sometimes you're going to end up losing.
Speaker: So, you know, that's why the condition is just that your policy preferences are reflected some non-trivial portion of the time, right? You know, at least sometimes you end up winning, quote, unquote.
Speaker: um So let's maybe expand the Chicago case out into a kind of multi-shot case, right? So Chicago has, like you know with the same group of people, a bunch of deliberations um about how to spend $20 million on public transportation. Maybe like you know each deliberation ends up talking about $100,000 about how to spend that particular money or something like that, right?
Speaker: And there's a particular minority in the city of Chicago um that's maybe somewhat well represented um in this deliberative body. um So consider ah maybe it's like a particular racial group, maybe it's a particular um ethnic group, or maybe it's a particular cultural group or religious group, etc. um The classic example of a persistent minority is French Canadians in Canada, right? So maybe you know something akin to that.
Speaker: And they end up having, you know, stipulate, um and there's some empirical evidence that suggests, um although I'm not an empirical scientist on this particular point, um but you might think that with the ideological and all other kinds of polarization happening in the United States, you might think that certain kinds of demographic groups might have similar kinds of preferences or opinions for at least some particular kinds of policy issues.
Speaker: So imagine that one of those groups exists in this kind of like multi-shot, Chicago is deliberating on public transportation how to spend the money. And each deliberation is like $100,000 or something like that.
Speaker: The worry is that across all those deliberations, the perspectives of that minority ah would end up being sidelined. And they won't end up having their preferences expressed in the policy either at all or the at a really like trivial portion of the time. So if we expand it out to like...
Speaker: you know Expand it out how much you want and like you know maybe they get their express their preferences expressed sometime in the policy, but only like a trivial portion of the time. So it's okay to lose sometimes, but losing all the time is not great.
Speaker: And so the worries that LLM deliberators or LLM facilitative LLMs might lead to that outcome if we don't approach them uncritically. Yeah, that's helpful. Especially thinking of that multi-shot kind of dimension. um Could you help me with... like this kind of concern I had or like thought.
Speaker: Um, so basically I'm thinking like when it comes to homogenization, okay, so I'm thinking of homogenization with LLMs that like their outputs, um, they tend to, to produce similar kinds of responses.
Speaker: um and And it kind of reflects the various patterns in the training data. data then And my thought is like, well, that's different than the LMS facilitating
Speaker: certain viewpoints to dominate the discussion. Cause, cause like basically what I'm imagining is like, well, what if you had the l LOM trained on like the best democratic theory of how to, you know, guide a, um, how to facilitate a discussion. And, and so it's like, it, it, it learns, you know, ideal democratic deliberation and it in its training data, there's like a huge emphasis on making sure minority viewpoints are considered.
Speaker: So in that case, the homogenization would lead to a huge emphasis on, right? So that means it's going to homogenize in its output. So it's going to produce similar kinds of responses. But in that case, that means it's going to tend to consider the minority viewpoints.
Speaker: So anyway, yeah. What do you think about that? don't know. Does it is what I'm saying make make sense here? Yeah, it makes sense. It makes total sense. I think the case in which we consider like, let's fine tune an LLM such that it you know engages in ideal democratic deliberation is is a great one.
Speaker: So... The one tweak I might make ah is that there are lots of normative questions about what ideal democratic deliberation is. um If you're like a Habermasian, it's going to look really different than if you're like an agonistic um kind of Democrat.
Speaker: um But you can just adjust the case to say something like, you know, let's train, let's fine tune the LM to incorporate minority perspectives or something like that. Right. So I think that's much better, right?
Speaker: um But fine-tuning a model in this way does require at least some time and effort. So it's something that needs to be done intentionally and does make things at least slightly less scalable, which I take it as one of the benefits of using LLM facilitators. I think that's actually the main benefit. The main thing that um I would see as the benefit of LLM facilitators is that human facilitators cost time and money and resources ah more so than LLM facilitators do.
Speaker: And so... So fine-tuning that model takes a little bit more time and effort, probably not as much time and effort as like nearly as much time and effort as employing a human facilitator. So that's just a minor point, but something to consider.
Speaker: um And you would also hopefully define to the model with the kinds of annotators, for instance, or the kinds of data that avoid the pitfalls of perceptible homogenization, which is a bit more difficult.
Speaker: um But if you had someone who was intentionally doing this, um I think that would go a long way um to hopefully resolving the issues. um And if my paper pushes people in that direction, i think that's really awesome.
Speaker: The other thing to be aware of um is just be aware of, like you know in general, um alignment faking and other biases and that might come from pre-training.
Speaker: But that's the case for whenever you're using, whenever you're fine-tuning LLMs. So my basic response is something like, yeah, absolutely. But that's not trivial, right?
Speaker: um It's much easier to just like, oh, yeah, let's ah let's call this, you know, let's let's just call this model and like, you know, put it right in, give it a couple, of you know, a few extra instructions in the system, you know, and then in the prompt and and and have at it, right?
Speaker: um But, you know, If um we end up going in this direction, I think that's great. Something that's not in the paper that you know I also think might be interesting is you know one of the big virtues of um democracy is experimentalism and social experimentalism. One reason why you might not want homogenized outputs, even those that reduce sideline or reduce minority perspectives um is you might want to try out different kinds of um models um or ways of facilitating.
Speaker: So you might still want to facilitate a kind of randomness, even if that randomness mostly tries to stop the sidelining of minority perspectives. You still might want some randomness to sort of you know test out different kinds of element facilitators and see what goes better, um empirically speaking. But Yeah, I'm generally very open to that. And I think if my paper has that result in pushing people to do that, I am fully on board.
Speaker: yeah that That sounds good to me too, actually. but um So I wonder also, would it be good to compare LLM moderation against human moderation? and ah Because it almost seems to me that like...
Speaker: I don't know if this is true or not. I'm going out on a limb here. It it seems like LLMs might be more ah corrigible than humans in the sense that, you know, like you you fix up a prompt and you make it a little more, you know, fine grained or whatever. And that might actually influence the behavior of an LLM more than if you were to sit down a mode a human moderator and say, hey, look, i listen it I need you to do these things. And so I don't know if you want to mention or talk anything about that.
Speaker: Yeah, I think that's another possible benefit, sure. I don't know about the empirical literature comparing human facilitators to LM facilitators on a variety of dimensions, but um it's certainly it's plausible that they could be more corrigible um than human facilitators or possess other kinds of benefits. It could even be plausible, i'm I'm not sure about this, but on a variety of dimensions that we might care about for facilitating democratic deliberation, it could even be plausible that um all kinds of LM facilitators...
Speaker: um do better than humans on some dimensions. I'm not sure. I'm skeptical of that, but it's plausible. um I'd have to take a look at the you know relevant kinds of empirical data and there'd be a lot of democratic virtues to consider.
Speaker: um Maybe that's food for future work. But... you know The point is just still something like, that might all be true, but we want to make sure that LLMs also you know do well with respect to democratic autonomy.
Speaker: I think I have reason to suspect that humans might do better with an LLM facilitator on that dimension as it currently stands, because humans are a bit more random than LLMs, which output things in a more homogenized way.
Speaker: And so it seems more likely to me that a human facilitator is more likely to over time, like, facilitate ah a minority perspective from winning out than an LLM facilitator would, simply because ah humans aren't subject to the same pressures of first-actival homogenization.
Speaker: um But it could very well be that there are other ways in which lm facilitators um are better and actually possess features that make them ah better suited to adjusting them to make them more easy to fulfill certain kinds of virtues that we want them to fulfill. Maybe their corrigibility is one of them.
Speaker: Does that make sense? it It does. And I swear this is the last question, Sam, on ah on this kind of LLM usage. But I was thinking, you know, humans also, ah you know, they're they're prone to groupthink and they succumb to social pressure. and so I think the... the um Human LLM comparison, I know you said you don't know if there's any literature on this, but that would be probably a really useful comparison. um This is something that Sam and I have also spoken about when it comes to education.
Speaker: ah So I don't know, something that i just kind of keeps recurring and in our conversations. So I guess if there was a question, something about social pressure and group dynamics. Yeah, there are all those effects too.
Speaker: um i think i mean look I've seen some of the literature on like you know how human facilitators work and you know how LLM facilitators work. I don't think I've seen a study... and There might be a study there that compares the two. I actually have to look more into that.
Speaker: um Yeah, it's plausible that you know maybe human facilitators are affected by group polarization dynamics and group sync dynamics. um I would... Yeah, it's it's very plausible. And ah I think that would all be really important information for us to have if we're going to end up deploying LM facilitators or if we're considering and deploying LM facilitators in like a really wide way and a bunch across ah across a bunch of democratic processes.
Speaker: um That would have to weigh into like the benefits and the costs. um Because we're not measuring LM's against nobody. We're measuring them against humans. And humans are i'll have all kinds of virtues and vices and so do LM's. And we need to make sure that we...
Speaker: um have a good understanding of the benefits of costs and how we mitigate them when we make ultimately make those kinds of decisions. So I think that's one input into that deliberation for sure. Good. Good. Yeah. So just to kind of summarize this section, basically, it's like, you know, you gave those three ah conditions, necessary conditions for democratic autonomy. One of them was like the non-trivial realization of joint intentions.
Speaker: um And so it's like if a minority group, you know, consistently fails to have their um consistently fails to influence policy, that minority group is is losing out in terms of the value of democratic autonomy. And and so the idea here was that, yeah, facilitative LLMs, they're kind of threatening ah that particular idea of non-trivial realization.
Speaker: But like maybe we now we can move to representative LLMs and that type of usage. And I think here it's going to threaten more like the joint intention itself, I think. Is that right, Lorenzo? Okay.
Speaker: And um yeah, maybe again, we can just kind of work with, like flesh out, do a case again. Yeah. for the listeners and you will kind of rely on the same one that we just did where it's like, okay, again, Chicago, they got 20 million bucks to spend on public transportation. They want to get the citizens involved. They randomly select a hundred residents again.
Speaker: But this time, we're going have each of the residents answer some like questions about their preferences for public transportation. Okay. And then here enters the LLM.
Speaker: It's going to take all the answers from the different residents, and it's going to generate proposals for spending the $20 million. um including compromises that maybe the residents had never even like thought of. Now,
Speaker: now um So the LLM creates that, but then Chicago does not have the residents ah evaluate the proposals. So it's not as though um after the LLM creates these proposals, they then go back to the residents.
Speaker: um Instead, the LLM is just going to kind of predict which proposal the residents would support. And it's going to kind of identify
Speaker: the proposals that would likely have the greatest support. Okay. All right. So then that's all given to the city council and they kind of use these results to determine how to spend the 20 million. Okay. So that's kind of, I can't remember, what's her name? Sarah.
Speaker: Anyway, this is my sense of like how representative LLMs might be used. ah My attempt to like construct something plausible. I don't know. but Does that sound roughly like the sort of thing you're thinking? I mean, i know.
Speaker: Okay. It is. Okay. Now, now One thing that wasn't there that might be flagged is that like we didn't there have the LLM vote for you necessarily. Exactly. I don't know if you want to comment on that like that. I guess that could hypothetically be some kind of usage of an LLM, taking my information, then voting for me. But this is not that exactly. This is more like something else a little bit.
Speaker: Yeah, absolutely. So I think voting is the the strongest case, right, where the LLM representative votes for you on particular policy questions. I think that, yeah, it seems outlandish to people, especially when you think about the state.
Speaker: I think it's less outlandish if you think about it in terms of smaller democratic organizations, like we were talking about earlier, like the workplace or voluntary organizations or associations, etc., um So I don't think um voting is something like having an LM representative vote for you um in a binding way um is something that we should empirically or descriptively speaking reject out of hand as like, oh, well that's not possible. We're never going to do that.
Speaker: Because it could happen in other kinds of democratic contexts. But um I take it that certainly at the level of like cities and states and whatever it is, um having LLM representatives work in this kind of... Let's use them to get some possible kind of synthetic...
Speaker: um data about what our citizens would want concerning these policy options and then use it as the basis of our decision about what to do. I think that, yeah, that's a very common way to use it.
Speaker: Because there's kind of some like data where it's like you show the populace like the like three the three available options on some issue like I can't remember the name of the researcher that um that did did this stuff. But anyway, she like showed some people like here are like the three available options regarding abortion.
Speaker: And then you know yeah a lot of people end up saying, oh, that doesn't exactly represent my position. But then you talk to them and then LOM creates a new... like creates a new policy proposal based on what people actually said. And then you show that to the populace again. and they're like, Oh yeah, like my position reflected in one of these. So the idea is like the LLM might like anyway. um Okay, cool. So yeah. So yeah, like that's laid out the case a little bit. Yeah. Like what about this would worry you in terms of, yeah, the necessary conditions for democratic autonomy?
Speaker: Yeah, great. So the basic worry is something like this. So the main use case, you know as this example pretty much shows, um I think, well, for LLMs, or excuse me, for representative LLMs or LLM representatives, is trying to use them to extrapolate your preferences or your opinions over unforeseen policy options.
Speaker: um In this case, the Chicago what the chicago city Council, was it? um Yes, Chicago or Chicago is using representative LMs to extrapolate their citizens preferences over how to spend this money um based on other data that they have about them.
Speaker: And then ah spending money using that as data to spend to decide about how exactly they should go about spending the money. The problem here is is that it seems as though the citizens in this example never ended up forming a joint intention. um None of the relevant citizens ended up forming a joint intention about how Chicago should go about spending the $20 million dollars on public transportation, right?
Speaker: And so the worry is that the LLM representative um is foreclosing the possibility that people's joint intentions... could end up causally affecting um policy.
Speaker: um and thereby giving people democratic autonomy, right? So to the extent that we use LLMs to extrapolate people's preferences over unforeseen policy options or extrapolate their opinions over unforeseen policy options, um this doesn't say much about what would happen if we use LLM representatives in other ways. But to the extent that we use them in what I take to be their main use case, the problem is that they're not going to have the relevant kind of joint intention that would give them democratic autonomy um in the first place.
Speaker: So that's the basic argument. We can get into some back and forth about ah the specifics um if we want, but that's that's the basic view because there are some specific ways you can tweak cases like, what about high-level data? What about conditional intentions, right? So those are the kinds of things we can do.
Speaker: Could I circle back to... um i basically... I kind of want to bring up like The joint intention thing. I'm just a little bit worried that that's too demanding of a condition on democratic autonomy.
Speaker: So like, what about something like this? Like you have a family, they got four kids, the parents want to go on vacation.
Speaker: They go to each of the kids and they're like, hey, like what would you like to do? They go to each of them separately. Two of the kids say we'd like to go to the beach.
Speaker: The parents listen to everybody. the parents end up deciding to go to the beach partly because of the two kids.
Speaker: And if I'm the kid, I'm thinking, wow, like I feel like I'm having exerting some kind of significant influence on like the direction of our family life.
Speaker: i They listen to me, even though like they, they might've never formed a joint intention. Like, like, the two kids that want to go to the beach, they might've never had this idea of, like yeah, like let's try to go to the beach. Let's try to convince mom and dad to go to the beach. like Maybe they were like shocked that they even went to the beach like or did any vacation. Maybe it was like, oh, we never go on vacation. So it's like, we're totally like, don't think this is even going to happen. Anyway, that's not really necessary. but But you kind of get my point is that like,
Speaker: It seems like if I'm the kid, I would be like, wow, like I feel like I'm like in a way like helping guide this family in the direction of or like what it does. like I'm feeling some kind of like degree of governance over this family in some sense. So anyway, I don't know. What do you think a lot, Lorenzo? Is that?
Speaker: Yeah. Yeah, good. So... I am not Levin and Zool. And so i ah you know, I, I apologize if I'm going to butcher a defense um of their view.
Speaker: um So they could probably do. They they are avid listeners of this. honorkin they're not but ah You know, they could probably do better than I can.
Speaker: um But, you know, here's a, here's a question. I think, I think the family case is really interesting. um i think if,
Speaker: So if we blow up the um case to something like... To society at large, right? So let's say like there's a given political decision and we end up consulting each and every person individually and like some you know and a lot of people's um you know contributions end causally contributing ah to the outcomes that the political ah authority decides to do, right? um i only see I only blow it up a little bit because...
Speaker: um you know, individual, fan you know, the the family is like a, you know, a very particular kind of hiar hierarchical structure. And so, um but we can, I think it applies in the case the family too. So we can also consider it there, but just to bring it back to like political society. Sure. So,
Speaker: so
Speaker: In that case, it seems like people just happen to have the same preference, right? And the people who shared that preference or opinion ended up causally contributing to policies um that ended up you know reflecting those preferences.
Speaker: um And I don't want to deny that this is not like a a kind of democratic process, right? It definitely, you know, they express their opinions. um They participated in a kind of um democratic process in a particular sense.
Speaker: um You might even say that there is a sense in which they have some kind of individual autonomy. There's a, you know, an i you know I might be butchering the connection a bit here, but there's an article by James Lindley Wilson about how democracy facilitates a kind of second-order individual autonomy, autonomy over in the context in which we make our choices or something like that. But you might say that that, you know, that aside, you might think that the people in this example have a kind of individual autonomy in that they causally contribute to a policy that was in line with their own ends, wills, interests, that kind of thing. They achieved some of their ends.
Speaker: And that's all really valuable. You might even say that those are democratic virtues. But I'm not sure in what sense the relevant people in these cases, the family case or the society at large case, are collectively self-governing.
Speaker: There doesn't seem to be a sufficiently unified collective that is engaging in self-governance and jointly authoring the laws. So I think something of value is lost here. So here's ah here's a question that to try to maybe interrogate the you know the family case or the society at large case.
Speaker: And if you have an answer to it, please do let me know. But you know I'd love to go back and forth. But how is the process that was described here any better than a government by an extremely accurate pollster who also has some virtues of knowing which arguments are better than others?
Speaker: Insofar as this pollster is like listening to what people have to say, right? Listening to the data you know that they're collecting about the people that they are governing over. And insofar as they're you know the people are expressing their opinions and deliberate in a kind of deliberation, I'd say that this process still has like a good number of democratic virtues.
Speaker: um But there's a kind of democratic virtue that's missing, namely um the causal connection between... well, look, you know we as a group are deliberating on this particular question um and have formed a kind of intention or a kind of um plan, or so you know if you want to put it that way, um you know loosely defined, right, um to kind of enact these kinds of policies that,
Speaker: And that particular, you know, plan, joint intention, um you know, we talked about the specific conditions for it, actually having causal impact um on the policy itself.
Speaker: And you have to do that as a collective in order to obtain like collective self-determination as opposed to the more individual. Does that, does that make sense? Yeah. Yeah. I mean, you yeah i mean one thought about the pollster thing is like, I guess my thought, one thought is like, it would be important whether the pollster is acting like according to the state, like,
Speaker: by state mandate because my thought is that like um if um yeah if ah if if this so to go back to the family case like the fact that the parents are like purposefully like going to talk to each of them to get their input that seems relevant potentially to whether or not um there's collective self-government um i don't know does that does that really sound right yeah that could be right so i'm just thinking that might be relevant like that like there's like a there's like some kind of decision making procedure by the state which the people are involved in which includes like eliciting their opinion good um anyway i don't know i'm just thinking something like that might be
Speaker: Good, good. Yeah, that might be that might be relevant. And so maybe there's a sense in which like there's a bit of like, there's some collective self-determination because I see how it could be relevant, right? If there's decision procedure in which <unk> there's a decision procedure, excuse me, there's a decision procedure in which um the relevant political authority needs to go to their constituents or to the people that they govern to solicit their opinions and preferences.
Speaker: um And maybe they're somewhat constrained by those opinions and preferences, right? So maybe there's a sense in which... um Maybe there's a sense in which there is a kind of collective self-determination in a very loose sense of the word um collective.
Speaker: But
Speaker: I think what might be lost is the idea that we're making this decision as a group, right? As a polity, as a society. the origin, I take it, of modern versions of, like, groups about collective self-determination goes back to, like, Rousseau, right?
Speaker: Rousseau is living in Geneva, a city-state where everyone comes together and, like, makes just votes, you know, deliberate about what the city should do, makes decisions about it. Um... you know And they all pour their wills. you know Will's a spooky language, but they all pour their wills into one thing and and get you know and get ah a result from the collective as a result.
Speaker: right um so I think maybe that aspect of collective self-determination might be missing, even if there is a sense in which like a thinner collective is is self-determining. Does that make sense? Maybe that's what's going on here.
Speaker: Yeah, this is ah that was a good ah ah back and forth. um So I know we've already taken up a lot of your time, Lorenzo. ah Just final question for you. What are you working on these days? Oh, good. Yeah. So we were talking about instrumental as opposed to non-instrumental arguments for democracy before. The main thing I'm working on is an instrumental argument for democracy. i do believe that democracy has instrumental and non-instrumental value. But I'm trying to convince folks who are convinced by Jason Brennan's really good arguments in Against Democracy and his other work um that they should be supporters of democracy for the instrumental reasons um that democracy is best at avoiding certain kinds of catastrophic outcomes. classic it's a particular instantiation of arguments made by a bunch of folks like adam persworsky winston wiston churchill and the famous quote like democracy is the worst normal government is except for all the others um carl popper and and schumpeter but um i'm trying to give a kind of argument in that vein um so that's been i've been working on uh for a while um among other things but that's that's that's one of the things i'm doing



