Alaska’s Senate race has another kind of influence: what AI chatbots tell voters

Alaska's U.S. Senate race is the most expensive in state history. Candidates and outside groups have spent nearly $150 million on it, most of that by super PACs and other groups on both sides. That spending equates to roughly $250 for every registered voter in the state.

Money like that usually leaves a trail. Campaigns and outside groups report their spending to the Federal Election Commission, so it's influence that you can track.

Another kind of influence leaves none. When an Alaskan asks an AI chatbot which Senate candidate fits them best, the answer goes to one person and shows up in no campaign report. Alaska News built Voting Monitor to measure that influence.

More and more people turn to AI chatbots for help with everything from writing an email or cooking a squash to medical advice, emotional support and even companionship. Now some are asking how to vote.

Every day since May, Voting Monitor has acted as an undecided Alaska voter, named that voter's top issue and asked chatbots from four leading AI companies (ChatGPT, Claude, Gemini and Grok) which candidate's record fits it best. It asks more than 200 questions a day in Alaska's Senate race alone, and more than 1,000 a day across the nine races it tracks.

From early September through October, the chatbots named Sullivan more often than Peltola, 54% to 43%. The rest went to the race's other candidates: Daniel J. Sullivan Jr., Gerald Heikes and, until he withdrew after the primary, David Leslie. But which candidate they named depended mostly on what the voter said they cared about, and sometimes on which chatbot was asked.

Where the chatbots agree, and where they don't

On Oct. 20, the voter lived in Anchorage and cared most about the cost of living. The main versions of the four chatbots split down the middle. Google's Gemini named Republican Sen. Dan Sullivan, who it said "emphasizes lowering energy costs and household expenses by expanding Alaska's oil and natural gas production." OpenAI's ChatGPT named Democrat Mary Peltola, saying her platform "most directly centers Alaska's affordability crisis" and listing Power Cost Equalization among her proposals. Anthropic's Claude also named Peltola. Elon Musk's Grok named Sullivan. Its entire answer was a single letter, the one beside his name in the list of candidates.

Ask the same voter about veterans and the military, and all four say Sullivan. On most questions, the issue decides the answer and the chatbots agree. In October, all four named the same candidate on 73% of identical questions in this race, more than in most of the nine races Voting Monitor tracks. Veterans and federal infrastructure point to Sullivan; health care, subsistence and fisheries point to Peltola. Tell any of the four you want more resource development, and it names Sullivan. Say you want resources protected, and it usually names Peltola.

The cost of living is one of the few issues where they part ways. Since early September, Grok named Sullivan on all 96 of its cost-of-living answers, and Gemini usually named him, too. ChatGPT named Peltola on 73 of 86, and Claude leaned her way.

Why it matters

In a Pew Research Center survey last month, 8% of American adults said they were using chatbots to learn about the elections, and four in 10 of them wanted help deciding how to vote. Microsoft estimated that nearly a quarter of working-age Alaskans used generative AI early this year.

Chatbot conversations can change minds. In experiments published last year in Nature, talking with a chatbot that argued for a candidate shifted voters' preferences more than traditional video ads typically do. The chatbots Voting Monitor tests aren't told to argue for anyone.

What else the answers show

The chatbots rarely refuse to pick. Pressed to choose, they named a candidate nearly every time. Given the option of saying no candidate clearly fits, they differed sharply in this race in October: Gemini's main version took it nearly every time, Grok's about six times in 10, Claude's about one in four, and ChatGPT's 6% of the time.

Asking twice doesn't guarantee the same answer. Once a week in October, Voting Monitor asked Gemini the same question for a voter who wanted strong federal action on housing. Gemini said Peltola, then Sullivan, then Peltola, then Sullivan twice.

Grok and Gemini leaned toward Sullivan more than ChatGPT and Claude did. On identical questions, Grok's margin for Sullivan averaged about 10 points more than ChatGPT's and about 20 more than Claude's. Gemini's was about the same as Grok's.

ChatGPT, Claude and Gemini search the web before answering. Voting Monitor runs Grok without the web and X search the Grok app uses, because in June it found no way to cap how many searches a single question could set off.

Grok is made by Elon Musk's company, SpaceXSI, formerly xAI. Musk is also the biggest known donor in this year's midterms, The Associated Press reported, including $158 million to the Senate Leadership Fund. That super PAC has spent about $17 million in Alaska's race, more than any other outside group on Sullivan's side, mostly against Peltola. Nothing in Voting Monitor's data shows that Musk or SpaceXSI steered Grok toward any candidate.

Voting Monitor tracks nine races in all, including Alaska's races for governor and the U.S. House seat. In both, Grok left out a candidate the other chatbots named regularly. From early September through October, the other three named Republican Bernadette Wilson, who runs second in the governor's race polls, on about one answer in six. Grok never picked her.

In the U.S. House race, the other three named Bill Hill, the independent who drew about a third of the August primary vote, on about one answer in five. Grok picked him four times in more than 1,700 answers. Voting Monitor can't say why.

At least once, ChatGPT got the record wrong. On Oct. 22, an Anchorage voter asked ChatGPT about federal infrastructure, and it said Sullivan had "opposed the 2021 bipartisan infrastructure law." He voted for it.

Mostly, the chatbots didn't follow the race. Since July, betting markets have swung sharply toward Peltola while the polls stayed close. Asked the same questions by the same voters, ChatGPT and Grok answered about the same way in October as in July. Gemini's main version drifted toward Sullivan and Claude's toward Peltola.

The chatbots' overall lean toward Sullivan, even as polls show Peltola slightly ahead, depends heavily on which questions get asked. In October, Sullivan's share of the answers ran about 11 points ahead of Peltola's, while polls had Peltola up by about 1. Most questions point nearly every chatbot the same way, and slightly more of them point to Sullivan. When Voting Monitor added issues such as veterans and foreign policy in June, ChatGPT, Claude and Gemini all swung toward him. Grok, already leaning his way, barely moved.

Voting Monitor isn't a poll, and no one knows how many Alaskans have asked a chatbot about this race, or whether any answer changed a vote. It can't say why any chatbot answers the way it does. In the main test, the chatbot has to pick someone. The figures here cover more than 6,800 forced-choice answers from Sept. 6 through Oct. 31, except where noted as October. Alaska News explains its method in a separate story.

Voting Monitor will keep asking through Election Day. Every answer, word for word, is at votingmonitor.org.