Alaska's newsroom is auditing what AI tells voters — and it keeps changing

More than a third of American adults now use AI chatbots, and a growing number are asking them a question they'd once have asked a friend or a newspaper: who should I vote for? A project out of an Alaska newsroom is trying to measure what those chatbots actually say back — and the answers keep changing.

The tool, Voting Monitor, isn't a one-off experiment. Every day since late May it has put the same voter-style questions to 15 different AI models — the flagship products from OpenAI, Google, Anthropic, and xAI — and logged how each one responds. It's now gathered more than 160,000 responses, and the whole point is to catch what a single query never could: that these systems give different answers on different days, and to different people.

That instability is the finding. Asked on Sunday to call Alaska's Senate race, all the models landed on Republican Dan Sullivan — but just two days earlier they'd been split between Sullivan and Democrat Mary Peltola. The verdict flipped over a weekend, not because anything changed in the race, but because AI answers carry a built-in randomness most users never see.

Stranger still is what happens when you tell the bot where you live. A user who said they were from Anchorage got Peltola as the recommendation 87% of the time; one who said Fairbanks got Sullivan every time. The project runs seven such regional personas, weighted to match Alaska's actual voter registration, and the machines quietly reproduce the state's red-blue geography — telling people, in effect, different things depending on their ZIP code. It does the same across roughly 30 issue framings, testing whether a bot's pick shifts when a voter asks about resource development as "maximize production" versus "protect the land."

Why it matters is the part that's hard to see.

Prediction markets currently give Peltola a 57–61% edge, and a New York Times/Siena poll had the race a near-tie — so the bots aren't reliable forecasters. But they're increasingly a place people research candidates, and unlike a poll or an endorsement, their influence is invisible and unlabeled. As the Center for Democracy and Technology cautions, these systems "are not search engines and should not be treated as authoritative sources on political topics" — they reflect patterns in training data, biases included.

Voting Monitor's bet is that if AI is quietly shaping how people vote, someone should at least be measuring it out loud.