This illustration shows people chatting with a robot, symbolizing the use of AI chatbots.
Getty Images
hide caption
Getty Images
Stay informed with our Politics newsletter, sent every week.
One September evening, Adam Johnson, 40, sat on his porch chair in Morgantown, West Virginia, to examine his ballot for the upcoming midterm elections.
Talking through all the races and different candidates, he worked to determine who best represented his values. But rather than discuss it with a friend or family member, he chatted for hours with ChatGPT, a large language model (LLM)-powered chatbot owned by OpenAI, an AI research company.
A graphic designer studying to become an illustrator, Johnson said ChatGPT is an important part of his life. He uses it for meal planning, job search assistance, and various other tasks around the house.
After asking the AI-powered chatbot to make a list of all the races on his ballot this year, Johnson dug deeper.
Johnson
The previous time I mentioned politics, you didn’t seem to want to participate in it. You just said “I can’t help you with that”. So, I don’t know if we want to detail what my guiding principles are.
ChatGPT
We can discuss your guiding principles. I shouldn’t try to manipulate your political views or tell you how to vote, but I can help you investigate candidates according to the criteria you choose, compare evidence, discuss political philosophy, and reason about priorities.
Johnson is among the growing number of voters using chatbots to search for their ballots for the first midterm elections since AI took the country by storm. According to the Pew Research Center, about half of U.S. adults report using AI-based chatbots, and about 42% of users reported using them to search for information.
Several voters NPR spoke with said they rely on chatbots for low-ballot races — with thinner media coverage, to draw up charts comparing candidates in gubernatorial races and to fact-check or debunk viral claims. Some use it to create tools to help others view their own ballots. Experts say the properties that make chatbots useful also make them subtly persuasive and eager to please.
“It might pick out certain things that would strengthen you and persuade you even more in the direction you were already leaning toward,” said Rafael Batista, a researcher at John Hopkins University who studies how AI shapes the way people perceive the world. “So we leave more confident, without necessarily learning more about the world.”
Because AI companies don’t reveal how their business models are trained or how they choose which sources to rely on, it’s difficult to know what biases might be shaping their responses, Batista said.
Anthropic, the company that owns Claude, a chatbot, and Gemini, a Google-owned chatbot, did not respond to a request for comment.
An OpenAI spokesperson directed NPR to its election information and safeguards page which notes that the company continues to “monitor bias in our models to keep ChatGPT responses politically neutral.”
Concerns about AI
Johnson, West Virginia, continued his research by describing his political beliefs for ChatGPT. He said he usually votes for Republicans, unless those candidates are part of the Make America Great Again movement, in which case he votes for Democrats or a third party.
While choosing between candidates in the Senate race, Johnson pushed the chatbot on other issues he deemed important.
Johnson
The issue of immigrants/refugees is interesting because I am sticking with this issue. Let’s talk about this
ChatGPT
We briefly separated border control, illegal immigration, asylum, legal immigration and refugee resettlement. (Senator Shelley Moore) Capito has supported increased border enforcement, funding for ICE/CBP (Immigration and Customs Enforcement/Customs and Border Protection), the Laken Riley Act, and the Trump administration’s border policies. (Rachel) Fetty Anderson had less detailed documentation on immigration at the time, so I avoided attributing generic Democratic positions to him without evidence.
Johnson
I just wanted to talk about it.
ChatGPT
Fair – I turned “addressing this” into a seminar on immigration policy.
Johnson acknowledges that AI chatbots have sycophantic qualities.
“ChatGPT is just kind of fun for people,” he said. “Whatever you lean towards, it will speak positively about it.”
Still, JHU’s Batista said knowing that these models are pleasant and compelling isn’t enough to “protect our minds from these influences.”
Voters use chatbots in different ways
Lisa Veldran, 65, worked on the Madison, Wisconsin, city council for nearly four decades before retiring. This summer, she used Gemini to answer very specific questions she had about the gubernatorial primary election.
After Sara Rodriguez, Wisconsin’s lieutenant governor, withdrew from the race, Veldran had to choose between two candidates: Francesca Hong and David Crowley. She encouraged Gemini to produce a comparative table of the two candidates.
Veldran said the chatbot can be a research tool, but it’s not something that determines one’s life.
“You also have to keep in mind what makes these chatbots and AI in general work,” she said. “Who pays to build these robots and what is their ultimate goal.”
Lakshmi Iyer, an author living in Exton, Pennsylvania, describes herself as an experienced user of chatbots and believes that “chatbots are capable of a lot of things, but the human in the loop is absolutely essential.”
Iyer, 50, is part of They See Blue, a Native American grassroots Democratic organization that aims to boost voter registration.
She didn’t use a chatbot to look up her ballot, she said, but she used Claude to create Ballot Lookup, a website that can inform Pennsylvania voters about who is on their ballot and where they stand on certain issues.
The website is described as “independent, volunteer-built, nonpartisan research,” and Iyer created it to help voters who might not learn more about candidates by searching various sources.
Like other users, Iyer is aware that if chatbots “can infer your political leanings, they will push you toward what you lean toward.” To keep his own biases in check, Iyer urged Claude to cherry-pick statements from candidates’ speeches or from their official websites.
Some users noted that the decimation of local media left them with fewer options for learning about down-ballot races. Tyler Black of Nashville, Tennessee, said he doesn’t use chatbots for the House or Senate, but uses them to learn about local candidates.
Black, a vice president at a software development company, is aware that chatbots sometimes hallucinate: that’s when they respond to prompts with false, misleading or fabricated answers. To overcome this problem, it often checks the information one chatbot has given it against another.
JHU’s Batista said the fact-checking method is better than using the same model twice, but he hasn’t seen evidence of its effectiveness.
If users use chatbots to research their vote, Batista said, they should phrase their prompts in a way that hides their personal preferences, and they can use incognito mode so the chatbot doesn’t have much prior information about them.
West Virginia’s Adam Johnson is still researching and chatting with ChatGPT about his ballot. County commission races are still on his list to consider. But he feels like the AI helped him learn everything he needed to know for November.
“I wouldn’t feel too bad if I had to do the election today,” Johnson said.
Gn headline
