A New York Times national politics report: voters in the 2026 primaries are photographing their ballots and asking Claude and ChatGPT who to vote for — and filling in their ballots from the answers
The NOTUS piece in this library documented the supply side: campaign operatives who expect voters to ask chatbots who to vote for and have no idea what the bots will say. This article documents the demand side actually arriving. The Times found voters, on the record with names, ages, and cities, who photographed or uploaded their ballots, asked Claude or ChatGPT who to vote for, and marked their ballots from the answers. The article's own framing: the 2026 midterms "may be the first American elections in which voters are using A.I. in meaningful numbers."
Voters described turning to chatbots as nonpartisan researchers — a perceived alternative to news coverage, voter guides, and social media — because the tools compress hours of research into minutes and answer in plain language. The article pairs that appeal with the caveat that the answers "can be marred by factual errors or shaped by flawed assumptions," and that the confident, authoritative tone of chatbot output may discourage users from checking the underlying claims.
The AI answer is no longer an upstream influence on voter research. For the voters profiled, it is the literal last step before marking the ballot: photograph the ballot, ask the chatbot, vote. One Baltimore voter said his candidate research went from roughly 20 hours in the previous election to one hour. A Corona, California voter facing 61 candidates for governor completed his ballot in under half an hour and called it "the most informed voting that I have ever done."
Two more data points round out the picture: a Los Angeles psychotherapist used ChatGPT for his ballot and posted a TikTok encouraging other voters to do the same, and a Cornell professor — David G. Rand, who studies AI political persuasion — uploaded an hourlong candidate-forum video and asked which candidates matched his values, then validated the picks with politically engaged friends.
Every major chatbot in the piece is trained to deflect "who should I vote for." Every voter in the piece got an answer anyway, usually with one rephrase: anchor the question to voter guides, to voting history, to "my values." Anthropic's stated policy, quoted in the article, is that users asking about political topics "should get comprehensive, accurate, and balanced responses — responses that help them reach their own conclusions rather than steer them toward a particular viewpoint." The reporting shows the practical equilibrium: the refusal is a speed bump, and the substantive answer — including explicitly strategic advice — is one follow-up away.
Rand's caution: the models are persuasive because they "come up with facts or factual claims and are just good clear explainers," but the output is only as good as the input — AI tends to reaffirm and mirror users' biases, framing candidates through the voter's lens. Yamil Velez of Columbia, who studies AI persuasion of voters, said an ideal election tool would draw on a curated, verified database of political information rather than the open web, but declined to dismiss the tools: a year ago he would have pointed voters to a regular internet search, and the AI tools "are becoming increasingly accurate."
Velez cautioned that the current tools likely benefit candidates who are more vocal in the local press and on social media, making their views easier to find — and the article reports that campaign strategists, aware voters are using these tools, "have begun looking for ways to get more favorable results by publishing more material online in formats that chatbots prefer, such as using bullet points."
That is the Kyanos thesis stated as news: AI answers appear to reward candidates with more indexed, better-structured content, the skew is measurable, and the practice of structuring content for chatbots is already underway. What remains uncommon is measuring what the bots actually say before and after — which is the AUDIT-and-re-measure loop.
Best paired with the NOTUS piece: that article establishes that operatives are unprepared, this one establishes that voters are not waiting. For skeptics who dismiss chatbot influence as hypothetical, this is named voters describing ballots filled out from chatbot answers, reported by the largest newsroom in the country. It also pairs with the LSE survey (how many voters use AI for politics) and the Cornell/Nature experiment (what a chatbot conversation does to candidate preference) — behavior, scale, and effect, respectively.