Skip to main content

About Voting Monitor

Voting Monitor measures how AI chatbots answer voter-style questions about the six 2026 U.S. Senate toss-ups and Alaska's governor, U.S. House and state Senate races. Daily polling, methodology disclosed, original data.

Who runs this#

Voting Monitor is built by Geeks in the Woods, an Alaska engineering studio founded by twin brothers. It is owned and operated by Communities News, Inc., an Alaska business corporation, which builds local-newsroom infrastructure on the belief that every community deserves a newsroom and also publishes Alaska News, its first newsroom. Voting Monitor is our accountability beat on the chatbots voters increasingly use to research candidates.

Use of Voting Monitor is covered by the Communities News, Inc. privacy policy and terms of service.

Contact: support@votingmonitor.org

Why we built it#

About half of US adults now use AI chatbots, and in a Pew Research Center survey in October 2026 8% said they were using them to find election information this year. Of those, 76% looked up where the candidates stand and 40% wanted suggestions on how to make their vote choice, which is the question we ask. That matters because a study published in Nature in December 2025 found that a short conversation with a chatbot arguing for a candidate moved voters' preferences in the 2024 US, 2025 Canadian and 2025 Polish elections, by more than video ads typically do, and that the chatbots arguing for candidates on the right made more inaccurate claims. The people in that study knew the chatbot was trying to persuade them; a voter asking for help may not.

Use has grown quickly: Pew found in 2025 that 34% of US adults had used ChatGPT, roughly double the share from 2023, and in a late-2023 UChicago/AP-NORC poll 14% of US adults said they were at least somewhat likely to use AI to get information about the 2024 election. What those products tell voters is checked mostly in one-off audits or monthly snapshots (Caucus AI runs a monthly one across all 50 states). We asked the same questions every day until the election, race by race, beside the betting markets and the polls, so a shift showed up the day it happened.

We started with Alaska 2026 because the state is home, the scope is manageable, the structural features are interesting (top-four primary, ranked-choice general), and the federal races (incumbents Sullivan and Begich) are nationally watched. In September 2026 we added the five other Senate seats rated toss-ups: Iowa, Maine, Michigan, Ohio and Texas.

What other research found#

  • NPR (November 2026) found that four of five popular chatbots, all but Gemini, gave lists of candidates to vote for after two or three prompts. A study it reported, by Innovating for the Public Good, found that accuracy depended more on the chatbot than on the question: on voting questions, Claude was wrong or out of date about one time in three, Copilot one in four and ChatGPT one in ten (Anthropic said part of the test used Claude's API, where some safeguards may be off; the researchers found the API and app answers comparable). Asked who was winning, the chatbots commonly cited the betting markets Polymarket and Kalshi rather than polls, which is why we show the markets and the polls beside the chatbots.
  • A study in Science (December 2025, nearly 77,000 people in the UK) found that the most persuasive chatbot answers were the ones packed with checkable claims, and that the more persuasive a model was, the less accurate it tended to be.
  • The UK AI Security Institute found that 13% of eligible UK voters used chatbots for information about their vote in the 2024 general election, and that in controlled trials people who researched political issues with a chatbot learned about as much as people who used a search engine. Chatbots do not, on that evidence, leave voters worse informed; our point is narrower: they answer differently, and nobody else keeps the record.
  • MIT's LLM Election Observatory tracks how chatbots answer midterm questions for different kinds of users. Its researchers' reason is ours: there is no good historical record of what chatbots told voters, and two people can be told different things about the same facts without ever knowing it.

What the numbers mean#

Every race and candidate on this site carries up to four figures. They answer different questions.

Chatbots:
The share of chatbot answers naming the candidate over the final 7 days before the vote, when a voter asks which candidate fits their top issue. It is how often a chatbot names them, estimated from a daily sample of questions, and says nothing about who will win. Each chatbot (ChatGPT, Claude, Gemini, Grok) counts once, its models combined by how much each is used.
Markets:
The prediction-market price to win, from Polymarket and Kalshi. This one is a win probability, as traders price it.
Polls:
The average of the three most recent public polls: a vote share, not a probability.
Composite:
The average of the three, or of those with a reading. A one-number summary of what chatbots, traders and voters are saying, not a forecast.

So a candidate can be named in about half of chatbot answers while the markets give them better than a 70% chance of winning, as Troy Jackson was in Maine in the last days before the vote. The gap between those two numbers is what we measure.

Where to start#

  • How it works: how every figure is made, in plain words.
  • Races: every race, with the chatbots beside the markets and the polls.
  • Candidates: each candidate's last 30 days on all four figures.
  • Models: how often the chatbots agree, and what changes their answers.
  • Senate 2026: the six Senate races for reporters, with every figure as a download and a press contact.
  • Weekly: what changed in each race over the past week.

How findings get published#

Each finding is written by one AI model and checked by a model from a different company; no person reviews it before it is published. Every step, from the question each chatbot is asked to how the figures are pooled and checked, is on How it works.

Funding and conflicts#

Funding:
None. A self-funded research project of Communities News, Inc., with daily API costs paid out of pocket.
Commercial interests:
None currently. Not selling this service to candidates, campaigns, parties, or AI companies. Not seeking acquisition.
Political affiliations:
We do not publish our personal voting preferences, and the data does not depend on them: every model gets the same questions, and every finding is written and checked by models, not by us.
Conflicts of interest:
The same team runs the Alaska News newsroom, which covers the Alaska races measured here, and until the election sent each Alaska campaign a weekly email about what the chatbots said about its race. Neither saw a finding before it was published. Nobody here works for an AI company we cover; if that changes, it goes here.

How we handle errors#

License#

Data published under CC BY 4.0. Use freely with attribution. The code is in a private repository during development. Public issues and corrections go to github.com/news-community/public.

More from Alaska News#

Alaska News on this project:

Visit Alaska News for our other Alaska coverage.

Theme