About 16 million Americans will use a chatbot to help decide at least one contest in November. The dominant way they do it is photographing the ballot, being declined, and rephrasing the question until the model answers.
Estimated The figures come from a Civly model. It re-weights Pew’s February 2026 age crosstabs from adults onto the midterm electorate, then applies the conversion rates measured after Scotland’s May 2026 election. See the five steps in the method section below.
In 2024 few voters used AI for election information, and some platforms blocked their AI from answering election questions. Two years later, using AI for election information is a mass behavior with a monitoring industry built around it.
Voter prompts below are quoted from published reporting. Model responses are described rather than reproduced.
Rikki Powers, in Baltimore, photographed her ballot and asked Claude for bullet points on each candidate, then checked the links it returned before she voted. She estimated that it cut her research time from 20 hours to one.
Mia Taylor, in Los Angeles County, photographed her ballot and asked Claude outright. It declined, citing political bias. She reframed, and got what she wanted anyway.
Chris Johnson, a registered Republican in Atlanta with a 40-year voting history who describes himself as libertarian, hit the same refusal before Georgia’s May primary. He got an answer by telling the model to reason from the candidates’ voting records.
Both voters got an answer on the second ask. Each rephrased the question, and the answer came back looking like sourced research.
Robert Siebelink, in Corona, California, faced 61 candidates for governor plus a slate of lower-profile races. He uploaded the ballot, asked Claude to suggest candidates aligned with his values, narrowed the governor’s race to two, then asked it to reason about strategy. He finished the ballot in under 30 minutes.
Jeremiah Hain, in Los Angeles, used ChatGPT for mayor and other offices (“I don’t have the time, nor did I want to do the same kind of research”), then posted a TikTok showing others how to repeat it. The method spreads through social media.
60% of US adults read AI-generated summaries at the top of search results. A voter who searches sample ballot [county] or a judge’s name receives a synthesized answer whether or not they wanted one. Pew found that 10% were unsure whether they had ever read an AI summary, which means self-reported AI use will always undercount the real figure.
AI’s share of the decision is near zero at the top of the ticket and largest at the bottom. Senate and governor races come with a party cue, saturation advertising and news coverage. House races get some news coverage in competitive seats. A nonpartisan county judgeship comes with none of those. In a race like that, a voter has no other source of information, and the model’s answer can be the only input. Analysis
| Contest type | Party cue | Other information available | AI’s likely role |
|---|---|---|---|
| US Senate / Governor | Strong | Saturation ads, news, debates | Negligible |
| US House | Strong | Some coverage | Minor |
| State legislature | Present | Thin to none | Meaningful |
| Ballot measures | None | Legalese text, partisan mailers | Substantial |
| Judicial / county / school board | Often none | Effectively none | Can be decisive |
Cornell’s experiments found that chatbot conversations moved likely Trump voters 3.9 points toward Harris on a 100-point scale, roughly four times the effect of a political ad. The same experiments found that a persuasion-optimized model shifted opposition voters 25 points. If a tenth of the down-ballot consultations involve an open choice, that is on the order of one to two million votes shaped by model output, concentrated in races decided by hundreds of votes. Colorado’s 3rd district was decided by 546 votes in 2022. Analysis
A monitoring-and-optimization industry formed around this behavior in 2026, which is itself evidence that the behavior is real at scale.
| Operation | Who runs it | What it does |
|---|---|---|
| CampSight | Run for Something Action Fund | Runs AI chats in real browser sessions to mimic voters, across ChatGPT, Claude and Gemini. 300,000+ campaign-focused conversations analyzed, with a waitlist of 60+ campaigns and roughly 400 candidates. |
| Caucus AI | Two chief analytics officers from the 2024 Harris campaign | Tracks what Gemini, Grok and ChatGPT say about candidates state by state, stores the responses, and exposes which sources the models drew on. |
| Research Books | American Bridge 21st Century | Gives chatbots direct access to opposition research files on 2026 House, Senate and Governor Republicans. |
| AI Narratives | CivlyOurs | Probes all seven engines (ChatGPT, Gemini, Claude, Perplexity, Grok, Google’s AI Overviews and Google’s AI Mode) on the questions voters type, grades every answer helps / neutral / hurts for the race, and returns the sources each engine used plus the pages to publish to change them. |
Tom Steyer spent over $200 million in California’s gubernatorial primary and finished third. Run for Something’s analysis found that chatbots ranked him sixth on affordability, cost-of-living and education-funding questions. Causation is unproven, but the industry now makes this argument to every campaign. Estimated
A model that states only true things but leaves out half the candidates shows the voter only half the field.
| # | Step | Result |
|---|---|---|
| 1 | Take Pew’s February 2026 chatbot use by age (18–29: 66% · 30–49: 61% · 50–64: 42% · 65+: 23%) and re-weight it from US adults onto the 2026 midterm electorate, using the 2022 CPS shape in which voters 65 and over are 30.4% of the electorate. | 49% → 44.1% |
| 2 | Correct back up for education and income. Voters are about 8pp more likely to hold a BA. Quinnipiac finds a 14-point education gradient (60% against 46%) and a steep income one (72% above $200k against 42% below $50k). | +2pp |
| 3 | Age the February data forward to November. Pew’s series runs 23% → 33% → 49%. Quinnipiac’s research-use figure ran 37% → 51% in eleven months. The projection blends the two and assumes growth slows as use saturates. | +6pp → 52% |
| 4 | Apply the conversion rates measured after the fact in Scotland (Survation / Strathclyde, n=1,042): 29% of adults used AI for election information and 7% used it to help decide. That is a 62% query rate among AI users and a 24% decision rate among those who queried. The rates are adjusted upward for US ballot length and vote-by-mail. | ×66% ×38% |
| 5 | Multiply by turnout: a voting-eligible population of roughly 247M at 47–52% turnout, central case 49%. | 121M ballots |
52% of voters using AI at all × 66% = the 34% who query about the election. The 34% × 38% = the 13% who use it to decide. The third tier applies Pew’s 60% AI-summary readership to the same electorate.
Scotland is the only place a real electorate has been measured after the fact. A Scottish voter casts two votes for parties they already know. A US midterm voter faces 15 to 40 contests. Some of them, such as nonpartisan judges and legalese ballot measures, have no party cue and no substitute source. Roughly a third of US ballots are also filled out at home with a phone in hand. Lower midterm salience and heavy straight-ticket voting push the rate back down.
Falsifiable predictions. If a US pollster finally asks before November, we expect “used AI to help decide” at 8–12%, “used AI for election information” at 25–35%, and under-40 decision use at 20–28%. Expect the first figure to come in below our 13%: fewer than 1 in 10 Americans think it is a good idea for voters to use chatbots to pick candidates, and 76% trust AI output “hardly ever” or only sometimes, so admitting to it is low-status. We would attribute that gap to social-desirability bias. A result above 15% or below 5% would mean the ballot-complexity adjustment in step 4 is wrong.
No survey has measured this directly. The three headline figures are modeled, not observed. The main assumption is that the rate at which Scottish voters, choosing between parties they already know, turned to AI says something about an American facing 40 contests and a nonpartisan judge. Step 4 adjusts that rate upward, and the size of the adjustment is Civly’s judgment.
The two 16 million figures measure different things. Caucus AI’s figure counts voters who received election information from AI during the primaries, which is closer to our 42M tier. Ours counts voters who will use AI to help decide a race in November. The two were produced independently and do not confirm each other.
The four patterns come from five voters in one newspaper account. They describe how voters use AI. They do not measure how many voters do it, and the projection does not use them.
Every voter in the 42M tier is asking an engine about a candidate, and the engine answers with whatever it can find. Find out what it says about yours and what to publish to change it.
Civly Politics Research · prepared August 20, 2026. No US survey has yet asked voters whether they used AI to decide their vote. Pew, AP-NORC, Quinnipiac, Gallup, Ipsos, Harvard IOP, Verasight and States United have not asked it. The November projection is modeled from adjacent measurements.
Voter prompts are quoted from published reporting, principally the Philadelphia Inquirer’s July 2026 account. Model responses are described rather than reproduced. Estimated figures carry the stated method and its ranges; anything labeled Analysis is Civly’s interpretation and is not a measurement.
Findings describe measured and modeled behavior and do not show that any chatbot answer caused any vote. Not legal, compliance, or investment advice.