AI chatbots: the new voter guides?

You talk to labor experts, economists, even just normal people, and there's a consistent fear. Artificial intelligence will replace jobs. But can it replace voter guides and informing people about issues in upcoming elections? Today on the pod, we look at the rise of voters turning to chatbots before going to the polls. It's the NPR Politics Podcast. I'm Tamara Keith. I cover politics. I'm Maham Javid. I cover politics. And I'm Domenico Montanaro, senior political editor and correspondent. And Maham, you've been doing some reporting on this. Why are people turning to chatbots for political advice? So I think it's for a few different reasons. One is that people are already using chatbots for so many other things that it seems like a natural move for them. I spoke to one voter in West Virginia. His name is Adam Johnson, and he has a paid subscription to ChatGPT, and he uses it to plan his meals. He uses it to search for jobs. So for him, it's a natural move to also use this to decide who to vote for. For other people, you know, there could be different reasons, such as the dearth of local news in some regions. They say that they don't really need to use chatbots for House and Senate races. National media is helpful for those. But for down-ballot races, they just can't find what they need from local media. One person I spoke to in Tennessee, Tyler Black, he is the vice president of a software development company. And he says that he uses chatbots for down-ballot candidates because he told me that he doesn't know much about the down-ballot candidates beyond their names in his district before he used the chatbot. Domenico, do you have any sense of just how widespread this is? Like, how many people are using chatbots at this point to get election information? Yeah, there's not a whole lot of information out there. But we know, of course, AI is the future. And we don't know exactly how it's going to be used. But it's increasingly a daily part of people's lives and seems to be heading more and more that direction. And then we're also seeing this split where in polls, people are becoming more skeptical of AI, in particular, younger voters saying that they are really frustrated or irritated or don't like AI or have negative opinions of AI. And yet they are kind of. Super users, as we know, of social media and all of that. So how they wind up interacting with AI is going to be a big question in years to come. And there was a poll from the American Association of Political Consultants that showed that six in 10 people are using AI a few times a month. About a quarter of people use it daily. And what I thought was interesting was when they asked specifically about election information, only 4% of people said that they use AI tools like chatbots for that. But 26% said that they use search. And of course, a lot of search engines are using those summaries that are generated by AI. And when people in this poll were asked whether or not they knew that those summaries were AI driven, only about half said they did. Yeah. And Maham, you mentioned the idea that there aren't good local news sources in a lot of places. But aren't chatbots fed? By local news? Isn't this a little bit of a cycle where, you know, like it's eating its tail? No, that's so true. I think for places where there's nothing else, there's always Ballotpedia. That's a nonpartisan website that provides voters insight about a number of different elections. You know, that's where some of the chatbots pick up their information from. And some of the information on these chatbots is also coming from candidates' official websites, especially when nothing in local media, then you really see the chatbot is taking a lot from the official website of the candidate. Right. The problem is, though, you know, they're scraping everything, right? And I don't know that they're distinguishing all the time between what the best nonpartisan sources are versus, you know, maybe good guides that are done by partisan outlets. You know, I just before we did the podcast, I just decided to test it out and see with a couple of the chatbots. And, you know, I asked who I should vote for in the Ohio Senate race, for example. And it, you know, said, well, you've got to decide based on things that you like, et cetera. And then it said, but if you want to know more information about the candidates, you know, there's a couple of resources. And the first thing that they linked to was the NAACP's voter guide, which, you know, may be very good, but that is not considered, you know, a news source or a completely nonpartisan outlet like Ballotpedia would be. Right. Maham, you talked to voters, people who are actually using these chatbots. How are they using these services? In so many ways. One of them is kind of exactly how you would imagine someone to be using it. You know, they kind of log in and then they prompt the chatbot to tell them, you know, who's running from my district and what can you tell me about each candidate? And they kind of go down that list, candidate by candidate. That's what Johnson from West Virginia did. He told ChatGPT his political opinion and he told ChatGPT his political preferences. And then once ChatGPT understood what Johnson is looking for, they went candidate by candidate and helped him choose. But there's people who use them, who use these chatbots in a very different way. There's a voter in Exton, Pennsylvania, and she said that she didn't need AI to decide who she wanted to vote for, but she did use an AI-powered chatbot to build a tool that can help other voters decide who they want to vote for. So, you know, it kind of highlighted how people use these chatbots in different ways and for very different purposes. Yeah. The other complication here is if you're asking a chatbot who you should vote for in a, you know, down-ballot race, a state representative or your county executive or school board, they're even in, you know, what's the opposite of a news desert, a news jungle or news rainforest? I don't know. But even if you have lots of good, you know, newspaper, or websites that are nonpartisan and deliver the news, it can be really, really difficult to find, you know, good information about the stances and political positions of each of those candidates on the ballot. And if that information isn't even out there for a lot of people, then how is ChatGPT or Claude or Gemini going to tell you, you know, what those people's positions are? And we know that there have been some dangers where some of these bots have hallucinated and gotten things wrong. I mean, I've tested it out here and there and found things, you know, when I ask about polling or where polls are, and I'm like, that one's not quite with that poll. And then you kind of click through the link to find the information, but not everybody's doing that. People just want to get the information quickly. Now, I will say on the other side of this, you know, there are some, you know, good places where you can find information on Secretary of State's websites about poll closing times and things like that. But sometimes it can be really difficult to find the information. And I think that's one of the things that's really, really difficult to find when you go to that website. And there are times when those chatbots are pretty good at being able to find the information very quickly and tell you the right information. The issue, I think, is that we're all kind of being experimented on as the bicycle is built. That's a new one on me. Usually it's the airplane that's being built. Let's keep it closer to the ground. All right. It's time for a quick break and we'll have more in a moment. And we're back. And we talked about this a little bit before, but I want to dig in deeper. One of the big concerns that people have about AI is that it hallucinates information. It gets things wrong quite confidently, in fact. So how does that factor into how people are researching topics and candidates for elections where, you know, misinformation is already a big problem? Right. Yeah. No. So hallucination is a big problem. And, you know, just to remind listeners, it's when chatbots are just making stuff up. Some of the voters told me that sometimes if you push back to hallucinations, the chatbot will confess that it made something up. But in other cases, you know, it will double down and it'll like really hold its ground to what it just made up. But apart from hallucinations, some other problems are that chatbots are known to be subtly persuasive and they are known to be agreeable to a fault. And these are two, quote unquote, qualities that can be very problematic, for when you are doing election research. And the reason it's problematic is because basically they're feeding back to the user what the user just told them. So you can kind of leave a conversation feeling that you've learned so much new information, but actually you may not have learned anything new. You're just kind of leaving more firm in your belief. That AI confidence is really quite something. And, you know, it really, makes the old political axiom of strong and wrong beats weak and right every time really come to life or at least artificially alive. Right. So I also think what's interesting about these problems is that the users and the voters, they it's not like they don't know. They know about these problems. They are the ones who brought up the hallucination to me. They are the ones who brought up the agreeableness to me. They are the ones who brought up the persuasiveness to me. And they've kind of, you know, they're kind of trying to find their own solutions to these problems. So Tyler Black, who I mentioned earlier, he said that to check for hallucinations, what he does is whatever one chatbot tells him, he'll take that answer and he'll feed it to the next chatbot and so on and so forth to kind of make them check each other. I asked an expert, Rafael Batista, he's a postdoctoral fellow at Johns Hopkins University, and Batista said that while this isn't a foolproof method, he does see some merit in this in the sense of that it's better than making the same chatbot check itself. Great. It does feel like we're all feeling around in the dark a little bit. And in a way, this reminds me of a different time in a different election. I'm thinking back to 2016 when everyone discovered that, oh, wow, people are getting different things in their Facebook feeds or their Twitter feeds. And the algorithm is feeding people information that puts them in a bubble and sort of steers them in a certain direction. And, you know, there was this question of like, oh, how do we get people out of their bubbles? Well, now it's even more of a black box because there there isn't that public data. This is a private chat between a voter and a chatbot. I'm wondering, how campaigns or people who are concerned about, you know, the flow of information, how they deal with the sort of black box nature of this? I spoke to some candidates who have been thinking about this from from the campaign side of things. Right. And so, you know, how Google relies on SEO, which is search engine optimization. So chatbots rely on something called AEO, which is answer engine optimization. And so some candidates in local races specifically, I could not find any candidates in federal races who who who were kind of looking for this as a solution. So candidates in local races, some of them are are looking into this. I spoke to one candidate. His name is Winscholus. He is a Democrat running for election to the Iowa House of Representatives. And he thinks that kind of looking at AEO has been worth his time and money. So he basically got the job done. And he's been got someone who understands AEO to tweak his website in a way that the chatbot will be able to glean the maximum amount of information that it can about him and put it on to whichever chatbot that is. I will add that Collis is a software engineer and he has helped run digital campaigns for Democrats in the past. So, you know, make of that what you will. Yeah, this is so fascinating. And it's going to be a huge, huge issue in the upcoming 2028 presidential campaign, not just about, you know, whether or not people want things like data centers or regulation, but also how campaigns are using AI, how AI plays into the campaign, because you better bet that that 4% number of people who are using chatbots to figure out the issues, you know, when they start to really get tuned into the election, that's going to rise pretty substantially. And certainly we've seen that in this age of social media, you know, with those algorithms that you talked about, you know, really kind of feeding people's confirmation biases and wanting to see what they believe. And it's going to be really interesting to see how all of that plays out on these new chatbots where people are going to increasingly start to use them. Well, and the other thing is these chatbots can be very verbose. They return paragraphs and paragraphs and paragraphs of results, which makes them look very smart. But then the average person is just going to skim it. You know, when I was looking at some of the chats that people were having with chatbots, they were just that. That's one of my problems with it, which I was like, these are so long. Like, if I'm reading this, why not just go and read the real deal? Like, if I have if I have the motivation to read this. Yeah, it's just so verbose. And, you know, we haven't even talked about deepfakes or all of these videos or potential audio and all of these other things that have been swirling around the campaigns this season. But what are you guys watching for as this election gets closer? Well, you know, first of all, in talking about deepfakes, we've seen an increased use of AI voices, you know, deepfakes of different candidates. And sometimes people recognize it. And sometimes people don't. People seem to be turned off by that kind of thing. That was one of the things that the that the political consultants poll really kind of harped on that people don't necessarily really like these things and don't necessarily have a strongly favorable opinion on on campaigns using them. But the problem is that two thirds of people said in this survey, that they worry about other people being fooled by deepfakes, but only a third of them worry about it happening to themselves. And if you're not conscious of the potential problems or biases of your own, it's going to be really hard for you to, you know, recognize something if you're not already dialed up with a healthy dose of skepticism. It speaks so clearly to the reporting that I did. Every voter I spoke to was one of the best voters in the country. And he was well aware of the problems. And they all seemed confident that because they knew what the issue is, they could kind of jump over it. Right. And this is where I'm going to bring up Batista again, because he was like, just because you know, the issue doesn't mean that you can kind of jump over it. Because that's why the chatbots they are they are built like that, right? Like they're built to persuade you they're built to make you believe that you're okay. And that's why you should only worry about others. Okay, let's leave it there for today. Come back tomorrow for more. I'm Tamara Keith, I cover politics. I'm Maham Javid, I also cover politics. And thank you for listening to the NPR politics podcast.