Midterm voting is underway, and a new step has worked its way into how people prepare. They open a chatbot and ask about the ballot. NPR reported on this earlier in October: voters describing their own political views to ChatGPT, asking which candidates fit, and in some cases letting the answer settle the race. One illustrator in West Virginia laid out his beliefs for the model, nudged it on the issues he cared about, and used its response to choose between Senate candidates. The NPR Politics Podcast followed up last week with an episode on chatbots as “the new voter guides.” This is not a fringe behavior anymore. It is a normal way to do election homework now.
The new habit
It is worth being clear about why the habit is spreading, because the reasons are good ones. Plenty of what appears on a ballot is genuinely hard to research. State Supreme Court races, judicial retention questions, county offices, local measures — for most of these a search engine returns a tangle of campaign sites, and a newspaper may not have covered the race at all. A tool that reads your ballot and summarizes where each candidate stands, in plain language, in seconds, is doing real work. I understand completely why people reach for it, because I have reached for it too, on a ballot full of races I could not have named the candidates in.
What the tools are good for
And to be fair, the tools often get it right. Summarizing a candidate’s stated positions from public records is the kind of task these systems are good at. The failure mode is subtler than a wrong answer. In the same NPR report, one voter offered his own diagnosis: “ChatGPT is just kind of a people-pleaser. It just aims to please. So whatever you lean towards, it will speak positively about it.” A researcher at Johns Hopkins made the same point from the other side of the conversation — these systems read your lean and lean back. So the answer you get about a candidate is not a neutral reference. It is a mirror with a slight tilt, and the tilt points at you.
Who is measuring that
Here is the part I keep coming back to. When a newspaper publishes an election guide, there are editors, corrections policies, and in some states legally mandated notice requirements behind it. When a chatbot summarizes the candidates, the answer is generated on the spot, for one person, and it is nondeterministic — ask the same question twice and you can get two different summaries. There is no editor in that loop. As far as I can tell, there is no systematic public measurement of how accurately chatbots describe the actual positions of actual candidates on actual ballots. Individual fact-checks get done. Nobody is doing the audit at scale, during the election, where it would matter.
That gap would matter less if people treated the answers as a first draft. Some do. Some photograph the ballot, ask for a summary, and then go search the names themselves — which is exactly the right instinct. But the reporting suggests plenty of voters stop at the first answer, and a first answer that flatters your existing politics is the most dangerous product these systems make.
What to do instead
None of this means go back to squinting at campaign flyers. Use the tools — they are good at this — but use them like a chatty source rather than a reference. Ask for the candidate’s positions and make the model cite what they are drawing from. Then verify anything close to a decision against an official source: your Secretary of State’s site, the county election office, the nonpartisan ballot guides. If two independent sources disagree with the chatbot, believe the sources, not the summary.
And vote. Half the races on a real ballot are ones none of us can research properly, and an abstention is still a decision. Chatbots make the homework easier than it has ever been — just remember that the tool telling you what you want to hear is the tool I wrote a book about. It is available in paperback and Kindle.