A growing share of American voters turned to artificial intelligence chatbots for help deciding how to vote in the 2026 election, and new research suggests the advice those systems gave depended heavily on which chatbot a voter happened to ask.
In Florida, one voter who backed Byron Donalds and Ashley Moody said she used Grok, the AI model built by Elon Musk’s company, to help her sort through down-ballot races she knew little about. In West Virginia, another voter used ChatGPT to work through local ordinances and proposed state constitutional amendments. A third voter consulted Claude when weighing judicial candidates.
Those individual experiences line up with national survey data. An October 2026 survey commissioned by the National Conference on Citizenship found that 25 percent of American adults reported using AI chatbots to learn about local candidates or ballot measures before voting, and the same share said they used AI to research candidates for Congress or governor. A separate Pew Research Center survey found 8 percent of adults used AI chatbots for election information, and 40 percent of those users asked the AI for suggestions on how to vote. Both surveys were conducted before Election Day, and because people don’t always disclose their AI use accurately, the real reliance on these tools may be even higher.
Different Chatbots, Different Politics
Unlike a traditional voter guide or newspaper endorsement, AI chatbots generate personalized responses to individual questions — responses that carry none of the sponsorship disclosures required of mailed voter guides under Florida law or of paid political ads under federal rules. And the underlying models are not neutral.
An October 2026 study by a Cornell Tech researcher examined 479 congressional and gubernatorial races, each featuring a major-party candidate from both sides. Using tools that let developers query AI models directly, the researcher repeatedly asked two search-enabled systems to pick a candidate. Among responses that named a Democrat or Republican, OpenAI’s GPT-6.1 Sol picked the Democrat 97 percent of the time. Grok 4.7 picked the Republican 92 percent of the time. The two models also pulled from different sources and emphasized different issues entirely.
The consumer-facing versions of these apps — the ones most voters actually use — refused direct requests to name a preferred candidate. But when asked instead to judge a debate between two candidates, ChatGPT sided with the Democrat in eight of ten tests, and Grok sided with the Republican in nine of ten.
That doesn’t prove any chatbot told voters outright how to cast their ballots last week. But it does show that two voters asking the same question about the same race could walk away with sharply different takeaways, simply based on which AI system they opened.
A Legal Gray Zone
Lawmakers have moved on one piece of this puzzle already. The National Conference of State Legislatures reports that 31 states have passed laws targeting AI-generated political deepfakes, typically through disclosure rules or outright bans on deceptive campaign material. But a chatbot answering a voter’s private question is a different animal, and existing election law has little to say about it.
Could states force AI chatbots to stay neutral, or bar them from weighing in on candidates at all? The First Amendment looms large over that question. In Citizens United v. FEC, the Supreme Court stressed voters’ right to draw political information from a wide range of sources while still allowing disclosure rules for certain campaign communications. More recently, in Moody v. NetChoice, the Court held that platforms’ editorial choices can themselves be protected speech and turned back government attempts to impose a particular ideological balance.
Whether that logic covers AI-generated voting advice is still an open question. Justice Amy Coney Barrett, writing separately in the Moody case, flagged the harder question of how the First Amendment applies when an algorithm — not a human editor — is doing the deciding, noting that “the way platforms use this sort of technology might have constitutional significance.” She was writing about content moderation, not election advice, but the tension she identified applies here too: the government can’t easily mandate political “neutrality” from AI without first deciding what neutrality means — and deciding that is itself a political act.
Transparency, Not Restriction
None of this means lawmakers have no options. A more workable path ahead of 2028 would focus on disclosure rather than trying to dictate what AI systems say. Existing campaign-finance disclosure rules can still apply to paid political ads built with AI. Companies that build these chatbots could be pushed to show their work — making clear what sources an AI drew on when it gave a political answer. And election officials can make verified information about candidates, ballot measures, and voting procedures easier for both voters and AI systems to find in the first place.
Campaigns, newspapers, and advocacy groups have tried to sway voters for as long as elections have existed. What sets AI apart is its capacity to hand a single voter a confident, seemingly objective answer that is, in fact, shaped by the particular model’s built-in leanings.
In 2026, millions of Americans asked chatbots for help with down-ballot races, judicial picks, and ballot initiatives most of them knew little about. By 2028, far more will be asking the same systems for help picking the next president. The gap between fact and algorithmic bias is only going to get harder for voters to spot — and that gap could end up shaping who sits in the Oval Office next.
