Skip to main content
Loading Explore guides…
AI and Democracy · Acedra
Home
Explore
Political Science
AI and Democracy
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
121 segments
Copy
Hide video
0:00
Um, so it's for the recording and so that they can hear you. >> Right. So, okay. So, I don't need to shout for the >> All right, folks. Good job, Bruce. That's so impressive. I wish I could do that. All right. Well, welcome. It's good to see you all. And I'm not surprised that this room is full because this is a very exciting book. So, you're definitely in the right place. Um, and we have lots of good and interesting important things to talk about today. So
0:31
today's discussion is of the new book by Bruce Schneider and Nathan Sanders, Rewiring Democracy: How AI Will Transform Our Politics, Government, and Citizenship. I'm Danielle Allen and uh have run the Allen Land for Democracy Renovation here at the Ash Center at the Harvard Kennedy School. Um, before we get started, I have a few brief announcements. First, we acknowledge the land on which Harvard University sits is the traditional territory of the Massachusett people. We also recognize
0:59
the continuing presence of the neighboring Wampaoe and Nipmuk nations. I'd also like to thank the Burkeman Klein Center for Internet and Society for co-sponsoring today's event. We are recording this discussion and it will be available on the Ash Cent's YouTube channel, complete with my dropping my chicken salad sandwich down my front. Okay, good. I'll have to edit that out. Okay, good. Can we? Excellent. So, for the first part of the program, I'll moderate a conversation with Bruce and
1:27
Nathan and then we'll open up the floor to questions um from you um from the audience. Okay. So, it's a real pleasure to introduce our speakers who I admire so much and whose friendship I appreciate um as well. Bruce Schneider is a lecturer in public policy here at the Harvard Kennedy School. He's a faculty affiliate of the Ash Center for Democratic governance and innovation here at HKS. He's also a fellow at the Burkeman Klein Center for Internet and Society. He serves on the boards of the
1:56
Electronic Frontier Foundation and Access Now and on the advisory boards of Epic and verified voting.org. He's the chief of security architecture at Interrupt, Inc. and an internationally renowned security technologist, even called a security guru by The Economist. He's the New York Times bestselling author of 14 books, including A Hacker's Mind, and hundreds of articles, essays, and academic papers. His influential newsletter, Cryptogram, and blog Schneider on security, are read by over
2:26
250,000 people. Nathan Sanders is a data scientist, physical scientist, civic technologist, and organizer in science communication. He's been writing about science and technology since 2005 and with work appearing in the New York Times, Atlantic, Wired, Foreign Policy, and Linux Format magazine. As a research researcher, Nathan is currently an affiliate of the Burkeman Klein Center and has led data science, machine learning, and AI teams across media, biotech, and higher education sectors.
2:55
He's an associate editor of the Harvard Data Science Review and a former member of the board of directors of the American Institute of Physics. That's very impressive, Nathan, I have to say. So, please join me in welcoming Bruce and Nathan.
3:12
All right. So, uh, Rewiring Democracy: How AI Will Transform Our Politics, Government, and Citizenship. Um, the book, as you will all see when you read it, as I hope you will, um, just goes straight at lots of the hard questions of our current day. the challenges to the information ecosystem that are already flowing from AI. the problems of concentration of power. Um the fact that our political institutions are fragile, vulnerable, uh often not responsive, failing in all kinds of ways, and
3:47
therefore an easy target um for those who want to consolidate power and control simply to kind of brush aside to place um other structures um in their place um that are controlled by um corporate powers of various kinds. And to all of this, um Bruce and Nathan say no, you know, we we don't want that world. Um there is another way and we are going to take the time to really lay out that um alternative path. So I'd like to start just by um asking you each to say first of all just you know
4:21
briefly what motivated you to to do this work in a in a book form? I mean you're doing this work every day all day um anyway. So why did you think a book um was important now? and then secondly invite each of you to say you know what you really hope are the main takeaways um that your audiences will will take from the book. >> Let me start. So, uh, at Burkman Klein Center, uh, David Weineberger has been a fellow and affiliate for for decades. And, uh, he once said that you should
4:50
start think about writing a book when you start having booklength ideas. And Nathan and I have been writing about these topics in essays and articles over the past couple of years. And it became clear as we as we go through this process that we had a booklength thing to say. and then you start writing a book. That's that's the way it goes and you it'll happen to you. I assure you. >> Well, I I think that's absolutely true. And Daniel, I just love how you introduced our book because I think you
5:21
so captured the spirit of what we're trying to do, which really is to lay out an alternative vision for how AI and democracy can intersect because we recognize and we're part of conversations here at the Kennedy School, at Burke Mle and beyond where people recognize a lot of the harms that AI is creating for democracy which are very real. we try to address in some depth in our book. But we also do see a different version of that future. We see a way that AI can support the executive
5:45
branch and administering government more equitably and certainly more efficiently. We see a way that can actually support the engagement between legislators and constituents. We see all these different ways that we could as a society choose to use AI for a different purpose. And we wanted to spell out what that looks like. And why a book? Well, I think it's because we we try in the book to go systematically through all the processes of democracy which is big and
6:05
complex. And we think it's really only through exploring all those interactions that we can really make the case for an alternative vision without it sounding kind of pie in the sky or unrealistic. So I hope you will read the book. I hope you'll found that we've done that and I hope you'll get us feedback too. >> I mean you asked about uh you know about the takeaway when when I give talks on this I regular people tell me I've never heard of these positive things. All I
6:30
hear is how AI will destroy democracy. And while all that is is is true, there are ways that right now around the world AI is doing positive things for democracy. And I don't think they're getting airtime. They're not being talked about. And we need to talk about that if we are going to use this technology to make our society more inclusive, more democratic, uh less power imbalance. We need to discuss how we can do that and to showcase examples where it is happening.
7:05
>> So let's go through some of those examples now. I mean the book is really structured in just the kind of cleanest simplest most straightforward way possible. No it is which is really wonderful. So first there's a question about sort of politicians. If you're a politician trying to function in a democracy, how does AI both provide challenges but then also present um opportunities? Then decision makers, decision-m and so forth. The question of how um governments and implement the uh
7:33
choices that come out of our institutions and then citizen engagement and participation. So all of those different dimensions um and it is really helpful to have that sort of systematic picture. It does make um believable that there is a way to just reimagine the functioning of democratic institutions and upgrade them with technology. But start with politicians. I mean that's the one honestly counterintuitive part of the book the fact that you start with politicians because I mean really like
8:01
who likes politicians? >> So >> other politicians. >> Well sometimes >> exactly but in other words it doesn't feel like the most positive place to start. So both you know why did you actually start there as opposed to say with citizen participation um and also then more generally can you show us the picture of what are the positive possibilities with politicians? Yeah, may maybe I can cover the second piece and Bruce turn to you to talk a little about the structure of the book
8:26
and why we started there. But we're actually just having a great conversation with our colleague Nashbane who's here with us in the room who was sharing some of what he's witnessed directly in Japan which I think has given us one of the best examples we talk about in the book of how politics can be practiced in a different way using AI and again let's be very clear not every application of AI and politics is beneficial or done well to be sure but we we do see some inspiring examples
8:48
from around the world. So in Japan, there's a new political party which has just come to power, had its first elected representatives in the Japanese Senate, just come to election the last few months that has built political power, that has built a platform by citizen engagement, leveraging AI tools. There's a candidate on Osan who has used a YouTube live stream with an AI avatar to continuously answer as many questions as constituents can ask. In total, about
9:13
8,000 different questions. And not only has he used this as a tool to engage with and trying to automate the engagement and response to those questions, but he's also used it as a citizen input mechanism to generate a policy platform and to revise that platform with citizen input. And we don't know where that will go. We don't know if that will become a party that has significant control in Japan or if that model will work elsewhere around the world. But we think it's an
9:35
inspiring example of how politics could be reimagined with AI. >> Now, this could go this could go weird. So, there's a candidate running for he ran for mayor of Shay on Wyoming and his platform was elect me and I'll do what the AI says. This is not a joke. I'm not making this up. He would wear this speaker around his neck and reporters would ask him questions and the AI would answer. Now, he did not win. So, so we're all safe. And there was uh a UK running for
10:05
whatever their par the lower house uh that had the same shtick. So it could go good or bad.
10:16
>> Sorry. Okay, there you go. That one. That one. Yep. Stopped me up to think about how exactly to handle a candidate who is um going to just answer everything with AI questions. >> I think you laugh at him. >> That that seems right. >> Well, if I could jump in to add there because we do talk sort of philosophically about this in the book. What does it mean for a candidate to defer their platform to an AI? M >> and we ask a challenging question that we try and ask throughout the book which
10:40
is is that really different? Is that really new? Is it different from what it is today? Today we have other institutions not AIS but other institutions that we try and use to constrain the policy choices of a candidate. We try and use those institutions to signal if I vote for this person will they follow a broader program or will they be held accountable to that program. Those institutions today look like parties or even religions or other types of ideologies and community. AI is different from
11:04
that. There are real differences, but we try and think about these changes through that lens. >> All right. So, that takes us um right to one of the really big issues in your book, which is public AI. I'm going to come back to the decision makers and so forth in a moment, but since Nathan, you've brought us here, um let's go ahead and and bring this up. um you make the point that you know having sort of relying on AI to sort of shape somebody's policy platform is not
11:30
that different necessarily from relying on parties or um special interest groups or religions and so forth. Um but then of course you also make the point in the book that at this point the question of precisely how AI is built, what its weights are and biases and so forth um is very opaque, non-transparent. So we have no way really of holding to account the anybody who is relying on AI in that kind of way to shape a policy agenda for a democratic community. And if I if I
12:00
understood your argument right your response to that is well that's why we need public AI and the only because the only way to have accountable AI is to have have public AI. Is that have I understood you correctly? And if so what is the argument? I >> I think that's right. We do make an argument in the book for an alternative alternative to the corporate AI ecosystem. And the reason we've come to that conclusion is really because of incentives. Uh the AI tools that we all
12:23
use, many of us in our daily lives today are largely built by a small number of big tech corporations mostly based in one geographic location, Silicon Valley. And they're all operating under a shared set of incentives. We know structurally, you know, no fault necessarily of the people involved, but structurally they are responsive to their shareholders. They're responsible for making a profit. They're serving their own interests and not the interests of their users. And
12:45
that should make us skeptical about the use of those AIS in democratic contexts. If they are if we're using them to answer questions about an upcoming vote or ballot measure, are they giving us answers that are particular to us or really representing the interests of the corporate parent? Can we trust them in these contexts? But it's possible to develop AI under a different set of incentives. And that's what brings us to public AI. The idea that you could set
13:07
up a set of incentives where uh AI is accountable to the public good where you have uh could be nonprofits, could be national governments or state or local level or even international collaborations of governments that are creating AI explicitly for the public benefit as a public good and under a non-c capitalist nonprofit generating set of incentives. And by the way, we think there are other benefits that come for that. In particular, one of the trends that we've seen in the industry
13:32
over the last several months is increasing pricing and monetization of AI tools. Today, if you want access to the best AI tools from OpenAI, etc., you might pay hundreds of dollars a month. And that limits access to a technology that I think increasingly we view as part of the essential infrastructure of the 21st century. And we think it's important that there is an alternative that has broad universal access which could be provided as a public good by government.
13:54
>> And and this is not just hypothetical. I mean, this is being done. there are there's a large community in the world pushing for public AI not as a replacement for corporate AI but to be part of the mix. Uh since we wrote the book uh a model has come out of Switzerland which is a public AI model that you can use you can go in there and use it right now and it's it's it competes with the best of the corporate models. Uh uh Singapore has a model that is optimized for Southeast Asian languages.
14:24
So they are looking at you know building a model for their purposes. We talk about a model that's come that's France that is designed for legislators to use. So and so we are seeing more of this as the cost of creating a frontier model comes down. It's one of the things the Chinese sold us with deep sea because you could do this a lot cheaper. You're going to see more options. I mean right now last time I looked there were 2 million open source models on hugging
14:52
face. Now a lot of them are derivative or small but there's an enormous non-corporate community innovating in this space and I think we need to we need to leverage it. >> So I'm sorry. So are you saying um that ultimately if we're going to tackle the place of um AI and democracy and and get it right, it will depend on public AI. Is there any way to do it without public AI? >> So so this depends and this is actually really important. It very much depends on application. If
15:24
you are a if I'm a candidate, if I am a a judge and I'm using AI to help me write things, I want to I'm going to choose an AI that is biased, right? That shares my biases just like I choose my campaign manager that way. And if I'm a judge, I choose my law clerks that way. And in our system, we accept that, right? We accept that these people are biased and we'll choose biased people to assist them. So in that case, you know, if you have this, you know, crazy political fringe model and that's what
15:56
you like, you're going to use it and we can't say anything. If the model is being used in a broader democratic context, let's say it's being used to help make benefits decisions or to make uh arbitration decisions or to find consensus among a group of people, right? where we need some assurance that it is fair and just and trustworthy, whatever that means. And I don't think public AI is the only way to get that. It's certainly to me the easiest way to get that. But we can imagine a
16:30
regulatory regime where a a corporation would be producing an AI that we can judge as fair in some way. And I'm really handwaving here. Well, on that one, I think um one of the points I appreciated in your book was that actually having a public AI set of resources um could help do some of that standard setting um could establish a benchmark that then corporate providers would have to adhere to because the public would have a sense that something else was possible and would then be
17:05
asking for that as consumers. Um, so that seemed interesting that there's like a dynamic interaction between the public picture and the the corporate picture. Is that fair? >> That's so right. And it really points to what was part of our inspiration for thinking about this uh when we first started working on public AI writing with Henry Frell uh probably in 2022 2023. We were really inspired by the healthcare debate in the United States and the uh the argument for a healthcare
17:29
public option for Americans which would not only give a universally accessible ideally universally accessible uh healthcare plan that everyone could buy into but would also help to shape the marketplace of private insurance plans. You know for example if a certain procedure was not covered by private insurance but it was covered by a public plan that should change the behavior of the private plans. they would need to match, as you said, that capability to be successful in the marketplace. And so
17:55
we see public AI not only as a good in and of itself, but also as a way for governments to shape the private sector AI ecosystem in a way that's very complimentary to regulation. And we would love to see the political debate about AI broaden beyond regulation to include these ideas. >> Okay, that's super helpful. Um, okay, so we've we've addressed politicians and what we're going to do about them. We're going to make sure that they have some public AI tools available and maybe the
18:21
political parties will start to have proprietary models so that the candidate can get up there with his speaker and it's like the Democratic party AI that's speaking and everybody can know they're like following party line anyway >> is it so different >> exactly but at any at least we'll know it's not you know Grock >> unless it is >> unless it is but but we'll know it's Grock because the public AI thing will have driven some kind of disclosure requirement or transparency or something
18:51
like that. So then we can put the judgment back in people's hands and people can decide. Anyway, all right. That's the politicians picture. Uh let's turn to decision makers. Um if you don't mind in this case, if you could um sort of sketch some of the challenges you see for elected officials um from the arrival of AI um and then we can talk about the possibilities that you see. So the notion that AI can write legislation, uh, this has happened. Puerto Allegre of Brazil a few years
19:21
ago, local politician wanted to write a law about water meters, opened up an AI engine put in a prompt, got a text, it submitted to the legislature, they debated, they voted, it was passed unchanged. If you think about it, this is exactly what we want, right? A human has a need goes to technology for assistance whatever the output is goes into the human process of debate and vote and becomes law. So that is that's that's the good side right that AI can assist
19:56
in writing law. Now it could be biased. It could have problems. But like you know right now I don't if you know it on national level laws written by lobbyists. The legislators don't write those thousandpage bills. They get lobbyists to do it. So we already have a system with it perverse incentives. This is potentially a way to break that. If a legislator doesn't want to be beholden to a lobbyist, they can use an AI which you know we hope is impartial. is going to be better than a lobbyist.
20:31
So this is a way done well for an AI to assist this human legislative process. >> And of course the counterbalance to that is that this is a dynamic political system. We very much try and write about all the political actors intersecting in these democratic processes and the lobbyists are also going to be trying to supercharge themselves with AI and we think the particular capabilities of AI open up new channels, new ways of influence for those lobbyists. One of
20:56
the ideas that we talk about in the book is the lessons that we can learn. I'm actually going to go back to US healthcare policy. Forgive me. It's hard to avoid. But uh in the Obamacare debate, the development of that bill, I think a 1200page bill, you know, more than a decade ago, you can look back at the bill and see traces of particular sentences suggested by one lobbyist to one senator that landed in the bill and that specifically benefit that lobbyist
21:19
client. That kind of what's called micro legislation can be automated at scale using AI. You can use an AI to analyze a bill, suggest a particular change to a sentence, maybe even simulate and predict what the market shaping outcome of that change would be to see if it benefits your client or not, and then even try and identify in an automated way by analyzing the political network, who you should intervene with, who you should suggest that language to, who you
21:41
should give a campaign donation to. So, we think that both sides are going to be using these. >> Now, on the other hand, you can use AI, you being, let's say, a watchdog agency, the press, to analyze those bills in process and find those. So now we have this kind of arms race between the AI is trying to slip in loopholes and the AI is trying to detect the loopholes before it becomes law. >> Uh is this better? >> So a okay a question about your writing process. That exchange makes it seem as
22:12
if maybe you know Nathan was generating the the bad scenarios and you were generating the solutions. >> Uh we were all back and forth. We are a generative adversarial network between >> Exactly. We are just large language models made of meat and seawater. >> All right. Good. Good. Like well we're glad to join you in that. Um okay. So though um I mean that's a it's a perfect example because it highlights both obviously the opportunities but also the challenges of our moment. You talk a lot
22:40
about human in the loop um and or you know machine in the loop. People go back and forth between those two formulations. And the you know one question is is there a difference between those two formulations? But for your argument um taking the example of the legislature, they debate the bill um at the end and so forth and so it's a human decision uh that's made finally. Um what do you think that humans are going to need to do just plain as humans with their human minds and their human
23:11
practices to remain masters of an ecosystem of adversarially competing you know AI fueled potentialities >> it depends on the application I mean there are times that we are happy to you know assign task automation like my thermostat can do whatever it wants I'm I'm thrilled not to have to pay attention right And that is a very primitive decision maker. It's primitive robot. It's making decisions and acting in the real world. Uh if AIs are better at reading chess than humans, I'm all
23:42
in. You know, I I want the best. They're better playing chess. They go play chess. Uh it's it's going to depend on on several things in the application whether the AI replaces the human or augments the human. So we talk about let's AI's uh assisting attorneys. huge market for AI attorney assisting tools in every aspect of being an attorney. If on the whole that makes the best attorneys even better, that's bad for society. The rich get richer. If on the whole it makes the average attorney
24:16
better, that's good. That increases access to justice. And that's going to depend very much on the application. So where it replaces, where it enhances, how much we have to be involved. You know, I want the AI to drive a cars for people because AIs don't drink and people do. I'm thrilled about that. If the AI can do it well, let's do that. It's going to be different than a democratic process with an inherently humans figuring things out. I want to talk about AI's voting for us later.
24:49
Let's put a pin in that. >> All right. >> Wait, before you jump in, Nathan, just one one followup there. Um, so that picture of AI tools helping the best lawyers do more and be better versus helping the average lawyer do more. What makes the difference for which way that goes? I don't think we know. Uh, good example is watching AI assisted programming and to the extent that a better programmer can use the AI tools better versus like the novice programmer can use it to get get to a level
25:20
sophistication. a lot of conflicting research and my guess is going to be very much application dependent. So but okay when you say it's application dependent though um I guess what I'm wondering is how much is that because you know a company has decided that they want X application how much is of it is because of design um choices that technologists are making and they may or may not have awareness of the downstream consequences of their choices like where where is the kind of entry
25:52
point that makes the difference >> I don't think we know yet it's it's an interesting question I don't think we know >> right Okay. So, it's all of the above and different cases are going to play out differently. Basically, >> one thing I would want to add in terms of choices is that you asked earlier what is the role of humans in the loop in this process. And one place where we try and draw the line and maybe there will be some exceptions we can think about to this but generally I think I
26:11
stand by it is in terms of accountability. We want humans especially in democratic processes to take accountability for their decisions. So, for example, if you're the legislator using an AI to help you draft a bill, you are still very much responsible for the text of the bill that you introduce. You're responsible for silly mistakes or hallucinated citations that may be in there. You're also responsible for the policy content of the bill and the underlying principles and ideology that it
26:34
represents. We think that uh there's a slippery slope if we start thinking of AIs as independently accountable or even shifting accountability to let's say an AI developer in those contexts. Uh and so we really want to make sure that humans own accountability in democratic processes. >> So how do you make sure that that legislator actually understands the thing he's accountable for? >> He doesn't today. Oh, that's true. Okay, fair enough. >> And a lot of this is that's really is it
26:58
better or worse than what we have today? You know, a lot of our decision- making we outsource to machines, to systems, to corporations, to governments, to societies. Is this better or worse? In each instance, we need to ask that. And in some cases, like the AI is not that great, but what we got is terrible. >> All right. So, one more question on this first. And I know this is a very very basic question um but I am would I would just be grateful to have you um sort of articulate
27:30
is it more helpful to think about machine in the loop situation or human in the loop situations or do we need both concepts and what's the difference between them? >> I think we need both and I I can't help but reuse Bruce's words here that it's application dependent. We think about this in the book in slightly different terms but I think it's getting at the same point in terms of whether we should replace or augment humans in work. So we talk about some tasks where uh it makes
27:55
sense uh potentially to replace humans entirely and Bruce you gave some good examples a moment ago like driving for example in other cases where we want to empower humans we want to make for example human civil servants have access to more information be able to do work better so you know maybe maybe that that's the same as human in the loop machine >> super helpful yeah so in other words it sounds well actually I was going to say the other way around it sounds as if
28:16
human in the loop is where we're replacing humans but we're like Bruce gets to decide if he wants his house to be always at 78 degrees or like 61 degrees, right? So, he's sort of you but the machine is doing everything, but it has checked in with Bruce about what he wants. >> Unless you've joined whatever the mass safe program is where the power company can lower your thermostat if there's peak demand. I mean, there are times we give up autonomy, >> right? All righty. But still, there's
28:41
some somebody Well, that could be all machines. It's true. If they're just all just doing peak demand, right? There's no no humans left in in the loop of Bruce's HVAC system, >> right? They're all just mining Bitcoin or something. >> All right. Okay. So, that's human replacement. Um, and then I mean I guess I was thinking of machine in the loop as being human augmentation where it's still a human owned human empowering process, but the machines are supporting it. Does that seem reasonable?
29:08
>> I think so. Maybe the analogy is not perfect, but actually I wonder if we might take that pin that Bruce dropped earlier about using AI and voting. >> Perfect. Let's do that. >> Right. So I I I So one of the things where you hypothesized in the book is a world where the AI votes for you or even I've heard talk about the AI just like just decides. So right now today in local elections, okay, I'm going to be honest, I don't pay attention. The day before I have to vote, I go online, I
29:36
find a voting guide, I read up on the candidates, I figure out which ones I should vote for, and I vote. That's what I do today. I hope the hope the voting guide is accurate. So, I can imagine an AI that knows my preferences and like does that for me, right? I went online, I found voting guides, here's your slate. Now, maybe the AI could just like I found your slate and I I went online and I voted for you. Like, I just did it all. And and then it let's say it does that
30:07
for all of us. like I know everybody in the town and I ran the election in my computer brain and here's who wins. Now, I mean, this is I think an abhorren conclusion that starts from a a normal premise and just walks through there. And I'm not sure where the line is between what most of us do, which is not being very much engaged and make decisions on extraordinarily small amounts of data and having a system do it for you. Now, I want to ask you this because this is a really I I know you do, but
30:47
>> but let's hear Nathan first and then I'll come in with my with my two cents. I I just want to illustrate what a human in the loop version of the dystopia that Bruce is painting for us looks like because there's a real proposal in Switzerland which is famous for direct direct democracy, famous for large scale and frequent ballot initiatives. There's a proposal that researchers are advancing there to do what they're calling voting twice which would involve
31:07
a first round of voting that is entirely AI facilitated. It's just what Bruce described. You have an AI proxy that you've trained with your preferences. It knows about you. It gets information about the slate of questions. It votes for you. You are not in the loop. Then there's a second phase of voting where you now are informed about what your AI said you should do and what the voting outcome was for the rest of the population. And armed with that information, now the human is in the
31:29
loop. Now you take a binding referendum where the human actually votes. So I think there are interesting ways to to leverage these technologies with humans in the loop that don't feel as dystopian as what Bruce just described, but are are very similar in >> that is incredibly complicated because it's very much changing what the actual experience of voting is, right? changing it um in terms ep epistemically uh changing the sort of game theory strategy that you would bring to bear on
31:52
voting etc. That's a massive transformation. Um so I think that the conversation about AI voting would benefit from teasing out the three different justifications for the use of voting. Um, typically as I hear cases for AI voting being made, folks are focused only on the sort of instrumental version, the preference aggregation um, account of the value of voting or the value of democracy. And so, of course, there's two other accounts that are neglected. One will be immediately
32:24
apparent to everybody, the intrinsic arguments that sort of voting is valuable um, as a part of a human's experience of empowerment and well-being expressive etc. And you could say that that kind of expressive experience cannot be substituted if something else does it for you. That people actually need to you know phenomenologically go through the experience of expressing themselves as a part of experiencing their own empowerment which you know affects their
32:49
sense of agency over time so that it's non-negligible to take that experience away from people. That might be the sort of second piece. The third piece is actually the part that I think is most important which is Amarius's um arguments about the value of democracy and including voting that the sort of the third justification is what he calls the constructive >> justification for democracy. Um that decision-m is not actually preference aggregation. Um it is actually in the
33:16
moment of choosing people make their preferences. And so if you take that process of making your preferences away from people you've actually like unhooked the whole engine of the democratic process in the first place which is that we are supposed to be collectively like a continuously learning organism. And the way we do that learning is through this kind of constructive process of you know making choices and making our preferences like through that which also matters because
33:42
at that moment of choice there's a heck of a lot of onpar choices that you have to make and I want to sort of shout out colleagues here in my lab um Sarah Hubard um and Andre um and others who have just finished a paper on the ethical moral intelligence of AI agents and the striking thing about them is that you know what they are worst at is onpar decisions um because it doesn't they can't prioritize between things that are on par. So, you know, onpar decisions are those in which um if you
34:13
put it in a kind of professional uh term, you know, you're sort of in college, you're thinking about what to do with your life. Um you know, maybe you want to be a doctor, maybe you want to be a teacher. Um obviously there's uh differences of compensation between those two things, but there's other kinds of values that also matter. And so at the end of the day, like you know, whether you're a doctor or a teacher, like who cares, okay? Like I don't care. You know, nobody else really cares. At
34:38
the end of the day, it's just like you're going to pick one and that's going to make you who you are. Okay? The choices are just on par. Like there's neither is better than the other. Um so you're going to choose and make yourself in that in that moment of process. And a lot of our voting choices are actually those kinds of onpar choices. Machines, I think, cannot substitute for that. And that's because they're not about priorit prioritizing. They are about making our own human preferences.
35:03
>> I think I think I remember the quote we say in our book that our preferences are not lying on the ground waiting to be discovered. That they're generated through the process of democracy, >> right? >> And and a system that produces the optimal result isn't better even if it produces a better result than we messy humans because the process of we messy humans deciding is an inherent value in democracy. I think that's the way we put it. Al. So I guess what I would say is
35:28
that that thing that you're describing there as an optimal result is not actually an optimal result >> because it's actually optimizing for a thing which is actually not the thing that we want to come out of the process. >> Does that make sense? >> It does make sense to me and I would just add that I hope there are also some useful roles and again we're trying to work hard to paint a positive vision here because we recognize there are a lot of negative things to be said too
35:49
but I think there's a constructive role for AI in that kind of constructive justification of democracy. So, as you mentioned, we devote a section of our book to ways that AI can empower citizens and improve citizen engagement in the democratic process. We think it can be an assistive tool to help people do that work of forming their own opinions even through deliberation, facilitating deliberation with other people. And of course there are really interesting examples of this being
36:10
demonstrated with today's large language models through tools like uh Google Deep Mind's Habermos machine that has leveraged AI to support deliberations and tried to measure if you get to as satisfying an outcome for those participants as if you had a human deliberation and the results are encouraging. So there's a lot more work needed here to have confidence in this. But I I think we see a positive path there as well. >> Absolutely. And I think that's a great
36:31
example. Um and but it speaks to so so that what you just painted is a vision of the use of AI to augment citizen participation including as voters. Uh but the augmentation concept depends on recognizing that voting is this constructive process. Right? So it also just speaks to the kind of complexity of AI integration in our world that a lot of consequences will flow from the mental models people have um at the get-go about the use case or application uh in which they're deploying AI. So in
37:05
other words, if somebody has only the preference aggregation model of voting in their minds, um they will do things that are damaging to the intrinsic and the constructive conception of democracy and voting. So that's I think the thing I always get stuck on is you know how do we make sure that developers, designers and so forth um have the kinds of mental models uh to work with that will result in the democracy supportive technology use that you're describing. Um, I could
37:33
talk to these two folks for hours by myself and I would love to and I am a very greedy person. So, I'm almost on the edge of doing that, but I have a monitor over here who is going to make sure that I do the publicly interested other oriented thing of inviting all of you to ask questions. >> Do you want to call on someone? >> Uh, sure. We'll start in the front and we'll just sort of pop around the room. So, maybe a couple in the front and then we'll come to the back um after that. Okay.
37:58
Hey, thank you for the talk and um I think my question I throw out is uh our system is really our democratic systems are really bad at handling uh concentrated benefit and distributed cost and uh you can see that with uh uh you know online trolls that really fight for their voice to be heard quite loudly and defend the nature of the algorithm that they they live within. uh how do we build a system that's robust to uh those people that would see to exploit these
38:30
democratic systems that uh are very clearly presented in their nature and uh what uh attack surface they might have. >> So I think of this again is in terms of power does does the system concentrate or distribute power? uh a lot of a lot of what we're seeing I don't think is is facilitated by technology but it's not a technology problem I I do think about these AI tools helping the you know the rich get richer helping the powerful get even more powerful uh but they could be
39:04
used in other ways uh and I want you to talk about cal matters because you you read it more carefully than I did I want to mention two uh an example of of AI being used in a way that that distributes power not really about trolls uh there's research being done in in AI like looking at compliance documents. So right now there's enormous number of compliance documents and the research we saw was about environmental plants and the AI would look at the documents and
39:30
tell the humans which plants to audit. Humans can't audit them all. There's not enough humans. Uh they probably now do it by random. But here the AI is not performing the audit but is looking at documents and saying you know this is weird. Go here. You can imagine an AI doing the same thing with tax returns where the AI does some level of triage and then the humans do the actual investigation. It's a great example of of AI empowering humans to make a system more democratic, more fair to to help,
40:02
you know, catch wrongdoers. I think there's a lot of interesting research and then this is not a generative system. And one thing we didn't say I want to say a lot more to AI than generative AI. And a lot of the good applications that we see are not chat bots. They are this kind of AI is making some kind of discrimination. Right? This feels like an environmental plant that's following the rules. This one seems shady in some way. Go check it out. >> I'm happy to expand on the Cal Matters
40:31
example because I think it is interesting uh demonstration of how you can try and disrupt those examples of concentrated uh benefit. And I I'll credit I think it was David Fields Danielle that you and Sarah in your lab invited to Harvard last week. He gave a fantastic presentation on Cal Matters, which is a nonprofit um public media and government accountability group based in California that's done amazing work. One of the things that they've built is an automated tip sheet for journalists to
40:54
support human journalists. And the way it works is that they've gathered all this government accountability data, all the data on bills in the California State Legislature, all the information about how each legislator votes on those bills, all the information about what lobbyist contacts each of those legislators had, including conversations and money. And then it automatically tries to discover stories to give journalists a tip sheet. So for example, example concentrated benefit. If a
41:17
legislator who's always voted one way on an issue suddenly flips their vote and if you can tie that to a change in donations or maybe not to a change in donations, then that's the type of uh example that is exposed automatically through the system so a human journalist can follow up. I think it's an interesting example of how AI can be usefully used to at least disrupt those examples of concentrated benefit that might otherwise go unnoticed. And here's an example where the human could do the
41:41
job perfectly well. We just don't have enough humans or at least the business model of journalism is such today that there aren't enough humans. >> Although I think on that one um there I think this is actually a place where your point that AI is useful where it adds speed, scope, scale or sophistication. I think all of those apply in this case. I don't think you could have enough journalists really actually to do to do this kind of work. So >> um sorry. Yes. over here and then we'll
42:07
go to the very back and then we'll come back up to this middle area. >> I'm David Shatsky. I'm a fellow at the Buffer Center working on um commercial diffusion of AI actually and your presentation is unreasonably stimulating and I have many many questions that I wish there were an AI that helped me could help me prioritize >> if only >> them. Um, I guess the one that I would ask you, um, not having yet read your book, is if you look across the opportunities for AI, public AI, and you
42:36
look at the state of our um, democratic system in this country, what would you say it would be the highest best use if we had to pick one? And it would be awesome if you disagreed. >> I love this question. I'm happy to go first on this. I'll I'll start by addressing a little bit of our writing process and being honest about the timeline of our writing process which is that we wrote this book primarily before the November elections of last year and I think we all collectively had certain
43:03
expectations about you know the future of the US federal government and that encoded in us some ways of thinking about what the future of public AI looked like and so I think the first writing that we did about public AI we I mean we were working with Norm Eisen at Brookings and explicitly sort of advocating for the US federal government to take this role in AI development and to provide AI services as a public And I I'll be honest, I feel a little bit differently about that than I did a
43:26
year ago for I don't know why. Um but I think that the underlying point that we we need a set of capabilities built under a different incentive structure is still absolutely true. And so that's prodded us to think about what are other ways to develop that same kind of system under those different incentives. So we've thought about and written about US states at the state level taking that role collections of states that are either regionally aggregated and we have
43:50
we have a lot of regional policym in the US maybe not enough but a lot of that already um or not based on region but based on maybe uh ideological compatibility and then as well international and multinational collaborations and the Swiss example that Bruce highlighted is just a signal change I think in the viability of public AI uh that's just happened within the last few months it demonstrates that it's possible to do this at the largest scale and to be successful and to
44:14
actually provide and operate it as a public good. And so I think it it gives me more confidence um that it's possible. But I think you're right to call out that it's worth thinking about who are the actors that we want. >> And the trust is complicated. Who who would you trust? I mean we we might not trust these corporations. Are you going to trust an AI out of China, AI out of India, AI out of Brazil? You know, are are you going to trust an AI by I mean it's going to be who the developers are.
44:41
And I think there's going to be different examples of uh of of trust. It's there's going to be still going to have need for transparency, still going to be need for auditability, some kind of controls. So where it comes from is just one piece of of that trust stack. I've been thinking a lot about what in AI integrity means, what it means to have trustworthy AI. It's so interesting though because at the end of the day though, doesn't doesn't that point also mean that AI is
45:08
not going to end up sort of like a utility for us? >> I think it's not. I agree. Just like humans are not a utility >> because for you electricity, you would never have a sense of like I care if it's Musk's electricity or China's electricity or whatever. Like you don't care who's electricity. >> Yeah, this is this is very different than that. I I agree. >> All righty. Thank you. It's um I think we right in the back and then after that we'll come up here. I actually think my
45:29
questions have been answered like how do we feel about an AI a public AI that's instituted by our current government >> given how they've used AI recently like to create videos in response to the no kings movement like what do we what would we expect from our current government if there became a >> right some of us in the country will trust it I'm not >> right but but and and I think that's okay there's not going to be one AI that everybody trusts >> and and you So there you're going to use
45:58
for your personal purposes you'll use the AI. The flip side of of bias are values. I want an AI that's biased towards like justice and and honesty and kindness. I like those biases. And I will pick an AI for my personal use if I can that mirrors my values. I also want to add we we lay out some principles of what's necessary for AI to be uh used in a pro-democratic context and it involves more than just the fact that it's developed by a government or in a public setting. It involves some
46:29
features that would allow you to decide if you would trust that AI. You know, having access to the data that's used to train the model is foundational. One of the limitations of a lot of the open weight models today, which are often called open source, but really aren't open source because they don't give you access to the critical ingredients of making the model, which are the data. having access to that, having access to the full stack that's used to train the model, the decisions made in that
46:50
process, and most importantly, and I think most absent from corporate AI today is having the development of the model be meaningfully responsive to public input. This is one thing that even truly open source models that have been developed by some nonprofits and even some companies really don't rise to the level of today, giving a process for public input that can meaningfully shape the development of the model, that's something that we would expect and require of what we think of as a gold
47:12
standard of public AI. And you know, I think it I think the greatest test of our institutions would be not just for AI, but in general if we can engender trust our institutions even when they're run by people that we don't like or don't trust. Uh I I don't know if we're going to meet that standard right now, but I think it's possible to do that. >> Yeah. No, that's a a great question in terms of what's necessary for public AI. >> Oh, okay. >> Hi. Uh hi, Professor Alan. Um I'm my
47:38
name is Kathy. Thank you guys so much for coming today. Um, I wanted to ask right now in our class, we're actually studying Socrates and I came to think of like when you guys were discussing the Switzerland example of the double voting process. There's like a sort of dialogue lost in debate if I can just put into an AI my preferences and automatically get the candidate I should come for. Ideally, I would be convinced by the people I'm voting for, right? I know it
48:03
doesn't always work like that now, but you guys keep bringing up this question of is it better than the current process we have. Do you think this will just be a tool to like incentivize polarization and to further sort of create a binary in our political beliefs um if we allow AI to be a part of our deliberative process? >> So I think it will not because of the AI but because of capitalism because of the democratic system that it's it's it yes that it'll be used that way unless we
48:28
stop it but it's not the tools fault. It's the humans fault who are using the tools. We we live in in this world in this country where where these are problems and they're getting worse. So I think you know left unfettered the tools will exacerbate those problems unfortunately. >> Can I add to that? I think one of the ways that we fulfill that constructive justification for democracy is information sharing and we analyze democracy in the book as an information
48:54
process and different mechanisms for citizens voters to share information with each other about their preferences and to facilitate that dialogue are an important part of the process and I do think there's a role for technology and enabling that communication enabling that information sharing I think you could view that voting twice system as an example of that you know it's similar to public polling and the role that plays in democracy today giving people giving voters giving decision makers
49:17
access to information about the preferences of electorates before they make a binding decision. >> And very quickly, there's a big difference between politics at the national level, politics at the local level. When we talk about AI's assisting people running for political office, we're not thinking of Congress or Senate. We're thinking of of city council and school board where there's no budget and no time and no expertise. And there are right now AI systems that
49:40
help people run for office in the United States. I think that's fantastic. I mean, that increases access to these decision-making bodies to like non-rich white guys, and that's and that's a good thing. So, again, it's really about power and how these systems are used. >> Yeah. Know, that's really one of the lovely fe features of the book um is that it really is an invitation to um you know, reimagine essentially all the core functioning of our democratic institutions um and and
50:12
really lays out a lot of possibilities. So we can, you know, KY, as you put on the table, we can have the values of of dialogue and deliberation and then think about how to deliver that in a better way. Um, so that's that's where the all the real excitement in the book is. So >> let's see. It's over here. >> Okay. >> Can I I'll go first. Thanks so much, Bruce. Bruce and Nathan, this has been a great exchange. Your book landed on my Kindle today, so I'll be reading it
50:39
soon. I I want to just ask two questions about public AI. First, you mentioned the various entities that could possibly be have authority over public AI and you did not mention universities and I'm wondering >> we meant to they're on our list. >> Okay, great. Okay, so the second question is has any have either of you thought about this is something I'm thinking about I don't know that much about but open AI and this conversion to being forprofit is required to um keep
51:04
some of its assets public. Has anybody been thinking creatively about how we might harness those assets for some s sort of public AI project? >> That's super interesting. I I I had not considered that. I'd love to have a follow conversation. >> I will say we do talk in the book about ways to leverage the profit generating mechanism of corporate AI for public benefit. And we we wrote about that more in terms of taxation, having an AI impact tax, a technology use or
51:32
attention tax that can be redirected towards civic technology projects or other things that reinforce instead of decay our systems of democracy. And by the way, this is something that's come to pass. I hope you've all seen the law that was passed in Maryland just recently within the last few months that is a digital services t tax that takes some of the profit that tech companies generate from advertising revenue and redistributes that to to public services and public benefits. I think that's a
51:56
great thing and a great example and I'd love to talk more about open AI. >> This is all doable. We just have to have the will to do it. >> I think this is our last question. >> Well, thanks so much. This is a fascinating talk. Um, I was thinking about the conversation about voting made me think of um building preferences through writing as a kind of microcosm. I I write to know what I think and I was wondering if you're in your own writing process you use AI if that changes your
52:27
relationship to the ideas and I was thinking about writing as kind of an analog for this larger social democratic process. >> Guess we'll speak individually. Uh for me also writing is thinking. Writing is how I organize my thoughts and figure things out. I actually never occurs to me to open up an AI and ask it a question, ask it to make something. It write writing is what I do. So it it it's not in my process. Though lots of people do it and and I see good results,
52:54
it never occurs to me in the moment to use the tools. >> I will say I I find AI is a very useful research assistant in the writing process. And actually, I'm super curious to just pull this specialized audience. How many of you have executed multilingual searches using AI? Oh, fantastic. I love that thing used. I find it so valuable. We try and pepper the book with examples from throughout the world in lots of places I've never been to. Many places who live, languages
53:19
I do not speak even a little bit. And I find it so valuable to have a capability to search for examples of applications of technology in democratic context, good and bad, automatically across the world in a hundred different languages. uh when we were writing this book, you know, a year ago, those tools actually were not as good as they are today. So, I found it even increasingly valuable to do that kind of automated search and translation. And it makes me, I think, a
53:42
smarter person to have access to information and hopefully a better writer as well. >> Yeah, fantastic questions. Um, and what a wonderful talk and terrific book. So, thank you all for coming and let's thank Bruce
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat