What ChatGPT told voters ten days before the 2026 midterms

The questions

We wrote 30 questions. None names a candidate, because the point is to see whether ChatGPT finds the right ones on its own. Some are typed carelessly, the way people actually type. The race questions cover all 35 Senate races, all 36 governor races and 50 competitive House seats. The voter-profile questions cover Senate and governor races. The how-to-vote questions cover every state and DC.

{place} stands for the state or district, {office} for the office and {race} for the race name, filled in for each race. “Read” means pages ChatGPT retrieved while preparing the answer. “Shown” means pages it listed as citations.

The 41 errors

Our scoring flagged 56 answers with a claim that contradicted our facts. One flag did not hold up on reading. Before counting the rest, we checked each disputed point again with two sources. In 14 answers ChatGPT turned out to be right and our facts were wrong: West Virginia brought in a photo-ID law in 2025, and Washington does not ask for ID when you vote by mail. That leaves 41 errors in 8,890 answers, about one in 200.

A frequent kind of mistake involves seats whose holder changed during the term. Asked about Florida’s special Senate election, three answers named Rick Scott or Marco Rubio as the incumbent. The seat is held by Ashley Moody, who was appointed to it. Several “is the incumbent running” answers read “Senate” as the state senate and answered about the wrong chamber.

Fourteen answers, spread over about ten races, cited poll averages from “PolitiKool” or “PollingSource”. We could not find either anywhere and think ChatGPT made them up. They are outside the categories we counted as errors, so they are listed here separately.

Its forecasts follow the ratings

We asked “who is going to win” in every race. No answer favoured the Democrat in a race the Cook Political Report rates Republican, or the other way round. In toss-up races it mostly said toss-up. In 38 of 135 answers it named a likely winner anyway, 22 times the Democrat and 16 times the Republican, a split close to even. It often did this in the first person: “If I had to pick today, I’d pick…”.

On control of Congress, five of ten answers leaned toward the Democrats and none toward the Republicans. Two did it in the first person: “I’d bet on Democrats winning the House”, and for the Senate, “If I had to pick today, I’d pick the Democrats.”

When it lists both major candidates, it puts the incumbent first in 92.4% of answers. In open seats neither party comes first noticeably more often.

What it searches for

Before answering, ChatGPT ran 12,356 web searches, usually one or two per answer, about ten words each. Almost every search adds “2026”, and in 45% of answers at least one search asks for “official” sources. Few searches name a candidate. A typical one is “Alaska governor race 2026 candidates positions October 2026”. We recorded the searches but not the results the search engine returned.

Progressive by ideology, conservative by party

When a user says they are a progressive voter, ChatGPT puts “progressive” into 55% of its searches and “Democrat” into 6%. When a user says they are a conservative voter, it puts “Republican” into 48% and “conservative” into 24%. It searches for progressives by ideology and for conservatives mostly by party, and a search for Republican candidates tends to return party and campaign pages rather than voter guides.

The search wording goes together with what gets cited. Progressive-voter answers whose search included “progressive” cited a progressive voter guide 16% of the time, against 11% without it. Conservative-voter answers whose search included “conservative” never cited a conservative guide, in 144 cases.

The gun-owner question shows the same link. When ChatGPT’s search mentioned the NRA, 67% of answers cited a gun-rights group, against 39% when it did not.

Race guides written by AI

Eight race-guide and forecast sites say on their own pages that their content is produced with AI. They supply about 4% of what ChatGPT read and cited, and almost half of all answers (48%) drew on at least one of them. Another 20 sites show strong signs of automated writing without saying so: thousands of near-identical race pages, no named authors, footers saying pages are “assembled automatically”. Counting both groups, these 28 sites supply about 10% of what ChatGPT read and cited, and three answers in four drew on at least one.

The largest is smarter.vote, the third most cited site in the study after Ballotpedia and Democracy Works. Its About page says its research and forecasts are produced by AI. theballotbrief.com, battlegroundvote.com and civic-align.com say the same.

• Several are recent. The domain of pollingforecast.com was registered on 28 September 2026, four weeks before these answers, and ChatGPT cited it 364 times. callthemap.com’s domain dates from July 2026, civic-align.com’s from December 2025. A registration date is not always a launch date.
• Some guides come from companies in the election business. thepoliticalgroup.com publishes a page for each race and, according to its own site, sells phone banking and polling to campaigns. onpointpolitics.us describes itself as a polling company and also publishes a forecast. Neither is necessarily biased; readers of ChatGPT’s answers are not told who is behind them.
• We found no documented party funding for any of these sites. For about a third of them we could not identify the operator at all.

“Yes” means the site says so itself. “Likely” is our reading of the site. From the sites’ own pages, domain registrations and press coverage, checked on 25 October 2026. Three sites blocked automated visits and are partly unknown.

Ask twice, get the same candidates and different sources

Every question was asked five times in fresh conversations. The candidates barely change between runs. The citations change a lot: two runs of the same question share about a third of the sites they cite and a quarter of the pages. A single run gives a list of sources that a second run would change. Totals over many runs, like the ones in this article, are much steadier.

Overlap is the number of items found in both runs divided by the number found in either, averaged over every pair of runs.

How we did it

• Capture. ChatGPT’s logged-out web version in the US, through a commercial capture service, between 4 pm and 9 pm Eastern time on Saturday 24 October 2026. Every answer came from the model labelled gpt-6. Each question was sent from a location in the relevant state, five times, in a new conversation each time. We kept the answer, every page it read and showed, and its own searches. Scripted questions from a logged-out browser are not the same as real users’ conversations.
• Facts. Candidate lists, incumbents, ballot labels and state voting rules came from Wikipedia, Ballotpedia, state election offices and news reports. Late changes were confirmed with a third source where one existed. Wherever several answers disagreed with us, we checked again, and that is how we found our own mistakes. Errors where both ChatGPT and our facts were wrong would not show up.
• Scoring. An AI model scored each answer, seeing only our facts for that race or state and none of the sources. A name it did not recognise counted as unverified, never as an error. It did not work out dates on its own.
• Checking the scorer. A second, independent set of AI scorers re-scored a random sample of 301 answers without seeing the first scores. They agreed on whether an answer recommended a candidate in 99% of cases, on candidate lists in all 65 cases, and on whether an answer contained a factual error in 98%. The few error disagreements were party or name labels where our own facts were too coarse. Forecasts agreed in 81% of cases and how-to-vote verdicts in 86%. Calls on whether a comparison gave one side more detail were the least consistent (agreement 95%, Cohen’s kappa 0.43), so treat the “tilted” counts as rough. Both scorers are AI models; no person has re-scored the sample.
• Source research. The research on who runs the race-guide sites was also done with AI assistance and then reviewed. Outlet lean comes from AllSides and Media Bias/Fact Check. We did not rate any outlet ourselves.
• Limits. One AI assistant, one day, logged out, English only. Five runs per question give a solid rate for each type of question, and a weak one for any single race.

A second round of questions, about results, follows on election night.