Voters Use AI Chatbots to Study Ballots as Public Unease With AI Grows

NPR reported that Adam Johnson, 40, a graphic designer in Morgantown, West Virginia, spent an hours-long September evening discussing his ballot with ChatGPT. Johnson, who uses the chatbot for meal planning and job hunting, asked it to list every race and then outlined his political beliefs. 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. In a Senate race, ChatGPT told him that Senator Shelley Moore Capito had supported stronger border enforcement, ICE and CBP funding, the Laken Riley Act and Trump administration border policies, while Rachel Fetty Anderson had less detailed immigration material publicly available; the chatbot said it avoided assigning her generic Democratic positions without evidence. Johnson said ChatGPT is “just kind of a people pleaser” and that “whatever you lean towards, it will speak positively about it.”

Pew Research Center data cited by NPR found that about half of American adults report using AI-powered chatbots and about 42 percent of users said they use them to search for information. Several voters told NPR they rely on chatbots for down-ballot races with thinner news coverage, to make tables comparing candidates for governors’ races and to verify or debunk viral claims; some have used the tools to help others look up their own ballots. Rafael Batista, a Johns Hopkins University fellow who studies how AI shapes the way people experience the world, said a chatbot “might select some things that would reinforce and persuade you even more towards the way that you were leaning already,” so users “leave more confident, without necessarily learning more about the world.” Because AI companies do not disclose how their commercial models are trained or how they choose sources, Batista said, it is hard to know what biases might shape answers. An OpenAI spokesperson directed NPR to its election information and safeguards page, which says the company continues to “monitor bias in our models to keep ChatGPT’s responses politically neutral.” Anthropic, which owns Claude, and Google’s Gemini did not respond to a request for comment, NPR reported.

NPR also reported that Lisa Veldran, 65, a retired Madison, Wisconsin, City Council staffer of almost four decades, used Gemini this summer to answer specific questions about the primary election for governor. After Lieutenant Governor Sara Rodriguez dropped out of the race, Veldran needed to choose between Francesca Hong and David Crowley and prompted Gemini to make a comparative table of the two candidates. Veldran said the chatbot can be a research tool.

The voter trend is unfolding alongside a broader paradox described by MIT Technology Review: many people say they dislike AI while use keeps climbing. Pew found that more US adults think AI will have a negative impact on them personally and on society than expect a positive one, with pessimism strongest among the young. A Stanford University report found that more than half of people worldwide say AI products and services make them nervous. In a May Gallup poll, 71 percent of US adults said they would oppose construction of a new AI data center in their area, compared with 53 percent who would oppose a new nuclear power plant. A March NBC poll found AI was less popular than ICE. Yet Sensor Tower said ChatGPT hit a billion monthly users in May, and Google DeepMind’s Gemini was close behind with 950 million users in July. Pew found half of US adults now say they use a chatbot, more than twice the number in 2023, and one in four say they do so every day. More than a third of adults across all 38 OECD countries reported using generative AI tools in the last three months.

MIT Technology Review argued that when people say they hate AI, they may dislike the relentless drive by companies behind it to push the technology into as many parts of life as possible, and warnings of the biggest social and economic upheaval in generations. It noted that social media saw a similar dynamic over the past 20 years, with billions using Facebook and Twitter despite growing techlash. With AI, the article said, there is more political appetite for regulation: all 50 US states have passed or proposed laws governing AI development and deployment, creating a patchwork of more than 2,100 bills nationwide, a tenfold increase in three years. Open-source alternatives to Google, OpenAI and Anthropic also offer, at least for now, the potential for more consumer choice and market pressure. The CEO of Springboards, a startup building an LLM, told MIT Technology Review, “We often say that we’re a self-loathing AI company. We don’t know if we really like what we’re doing.” He said there was no walking back from LLMs, but they could still be made to do something different; the article said it hoped for AI that is clear about what it can and cannot do and is not posturing as if it is about to take over the world.

A separate essay in The Verge, titled “Our minds aren’t equipped to handle AI,” argued that AI is “junk food for the mind; easy, tempting and ultimately very bad for you.” It cited Norbert Wiener, the godfather of cybernetics, saying, “The thought of every age is reflected in its technique.” The essay noted that Google’s Demis Hassabis calls the brain “a biological approximation to a Turing machine” and that Elon Musk says “people should just think of the brain as a biological computer,” but argued humans are more complex than that comparison allows. It described the computational model that treats thought as input, computation and output, and noted that John von Neumann doubted the model could capture the “exceptional complexity of the human nervous system.” The essay favored the approach of Paul Cisek, a neuroscientist at the University of Montreal, who contends that brains are better understood as feedback-control systems rather than information processors. It quoted philosopher John Dewey describing the mind as a circuit, “more truly termed organic than reflex, because the motor response determines the stimulus, just as truly as sensory stimulus determines movement.” A baseball example illustrated the difference: catching a fly ball could require calculating velocity and gravity under the computational model, while feedback control suggests keeping the ball in the same position in one’s visual field and moving to maintain that situation. The essay said this maps to the biological evolution of nervous systems from ancient fish to amphibians, mammals and primates, a history Cisek describes as the “continuous extension of control further and further into the world.”