The Synthetic Governor Poll
Democrats are favoured in 21 of 36 governor’s races
Every number on this page starts from synthetic survey answers: three language models answering as 3,600 real respondents to the Cooperative Election Study. No real voter was asked anything. We combine those answers with published polls and past results, then simulate the election 10,000 times.
Updated 1 November 2026, 6:08 PM Toronto time. The models were asked again each day until 1 November.
Democratic governors not up this year Republican governors not up this year
27 Democratic governors after the election, median of 10,000 simulations (80% range 21 to 32)
61% chance Democrats hold 26 or more of the 50 governorships
28% chance Republicans hold 26 or more
Every race, closest first
Each row represents one of the 36 governor races ordered from the closest to the safest. A dot on the left favours the Democrat while a dot on the right favours the Republican. The farther the dot is from the centre, the safer the race is for that party. The percentage shows the favoured candidate’s chance of winning not their share of the vote. Ratings are classified as Toss-up, Lean, Likely or Solid based on the candidate’s chance of winning. You can filter races by rating, search for a state or select the plus sign to view candidates and estimates from each source.
The current forecast has 12 Solid Democratic races, 4 Likely Democratic races, 3 Lean Democratic races, 3 Toss-ups, 5 Lean Republican races, 1 Likely Republican race and 8 Solid Republican races. Georgia, Nevada and Ohio are the three Toss-ups.
The layout of this strip is adapted from the Decision Desk HQ governor ratings page.
Fourteen governors are not up for election this year, including six Democrats and eight Republicans. These are shown as empty squares in the bar at the top of the page. Democrats need to win 20 of the 36 contested races to reach 26 governorships, while Republicans need 18. The forecast gives Democrats a median of 27 governorships, with an 80% range of 21 to 32. Democrats have a 61% chance of holding a majority, compared with 28% for Republicans. In about 10% of simulations, neither party reaches a majority. The results vary because a national swing can affect many states at the same time.
| State | Democrat | Republican | Independent | Rating | Democratic win | Republican win | Independent win | Margin | 80% range | Polls |
|---|---|---|---|---|---|---|---|---|---|---|
| Georgia | Keisha Lance Bottoms | Rick Jackson | Toss‑up | 51% | 49% | D+0.1 | R+8.1 to D+8.5 | 10 | ||
| Nevada | Aaron Ford | Joe Lombardo | Toss‑up | 47% | 53% | R+0.4 | R+8.6 to D+7.8 | 3 | ||
| Ohio | Amy Acton | Vivek Ramaswamy | Toss‑up | 59% | 41% | D+1.5 | R+6.7 to D+9.7 | 16 | ||
| Vermont | Amanda Janoo | Phil Scott | Lean D | 62% | 38% | D+2.3 | R+7.1 to D+11.8 | 3 | ||
| Florida | David Jolly | Byron Donalds | Lean R | 36% | 64% | R+2.3 | R+10.6 to D+5.9 | 17 | ||
| Alaska | Jonathan Kreiss-Tomkins | Dave Bronson / Treg Taylor / Bernadette Wilson | Lean R | 33% | 67% | R+2.9 | R+11.4 to D+5.6 | 7 | ||
| Texas | Gina Hinojosa | Greg Abbott | Lean R | 31% | 69% | R+3.1 | R+11.4 to D+5.1 | 35 | ||
| New Hampshire | Cinde Warmington | Kelly Ayotte | Lean R | 29% | 71% | R+3.5 | R+11.7 to D+4.8 | 9 | ||
| Iowa | Rob Sand | Zach Lahn | Lean D | 71% | 29% | D+3.6 | R+4.7 to D+11.9 | 17 | ||
| Kansas | Cindy Holscher | Ty Masterson | Lean R | 29% | 71% | R+3.7 | R+11.9 to D+4.5 | 7 | ||
| Wisconsin | David Crowley | Tom Tiffany | Lean D | 72% | 28% | D+3.6 | R+4.4 to D+11.7 | 7 | ||
| Michigan | Jocelyn Benson | John James | Likely D | 80% | 20% | D+5.4 | R+2.8 to D+13.7 | 21 | ||
| Arizona | Katie Hobbs | Andy Biggs | Likely D | 80% | 20% | D+5.4 | R+2.9 to D+13.7 | 13 | ||
| Rhode Island | Helena Foulkes | Aaron Guckian | Ken Block | Likely D | 82% | <1% | 18% | D+35.7 | D+27.2 to D+44.0 | 6 |
| Nebraska | Lynne Walz | Jim Pillen | Rick Beard / Brett Lindstrom | Likely R | 18% | 82% | <1% | R+5.9 | R+14.1 to D+2.4 | 9 |
| Oregon | Tina Kotek | Christine Drazan | Likely D | 86% | 14% | D+7.0 | R+1.4 to D+15.3 | 4 | ||
| South Carolina | Jermaine Johnson | Alan Wilson | Solid R | 8% | 92% | R+9.1 | R+17.4 to R+0.9 | 4 | ||
| Minnesota | Amy Klobuchar | Lisa Demuth | Solid D | 94% | 6% | D+9.9 | D+1.8 to D+18.2 | 7 | ||
| Oklahoma | Cyndi Munson | Mike Mazzei | Solid R | 5% | 95% | R+10.3 | R+18.5 to R+2.1 | 3 | ||
| New Mexico | Deb Haaland | Gregg Hull | Solid D | 96% | 4% | D+11.5 | D+3.2 to D+19.8 | 4 | ||
| Alabama | Doug Jones | Tommy Tuberville | Solid R | 3% | 97% | R+12.0 | R+20.1 to R+3.7 | 3 | ||
| Colorado | Phil Weiser | Victor Marx | Solid D | 97% | 3% | D+12.5 | D+4.1 to D+20.7 | 0 | ||
| Maine | Hannah Pingree | Robert B. Charles | Rick Bennett | Solid D | 97% | 3% | <1% | D+12.3 | D+4.0 to D+20.7 | 13 |
| Arkansas | Fredrick Love | Sarah Huckabee Sanders | Colt Shelby | Solid R | 1% | 99% | <1% | R+14.9 | R+23.2 to R+6.5 | 2 |
| Tennessee | Jerri Green | Marsha Blackburn | Solid R | 1% | 99% | R+16.1 | R+24.2 to R+8.0 | 3 | ||
| New York | Kathy Hochul | Bruce Blakeman | Solid D | 99% | 1% | D+16.4 | D+8.1 to D+24.8 | 12 | ||
| Illinois | JB Pritzker | Darren Bailey | Solid D | >99% | <1% | D+17.7 | D+9.4 to D+25.8 | 2 | ||
| South Dakota | Dan Ahlers | Larry Rhoden | Solid R | <1% | >99% | R+18.6 | R+26.7 to R+10.4 | 2 | ||
| Pennsylvania | Josh Shapiro | Stacy Garrity | Solid D | >99% | <1% | D+18.7 | D+10.5 to D+26.9 | 11 | ||
| Connecticut | Ned Lamont | Ryan Fazio | Solid D | >99% | <1% | D+19.8 | D+11.6 to D+28.1 | 9 | ||
| Wyoming | Kenneth Casner | Eric Barlow | Solid R | <1% | >99% | R+19.4 | R+27.9 to R+11.1 | 0 | ||
| California | Xavier Becerra | Steve Hilton | Solid D | >99% | <1% | D+20.9 | D+12.6 to D+29.2 | 11 | ||
| Idaho | Terri Pickens | Brad Little | John Stegner | Solid R | <1% | >99% | <1% | R+22.6 | R+30.9 to R+14.3 | 2 |
| Maryland | Wes Moore | Dan Cox | Solid D | >99% | <1% | D+31.2 | D+22.9 to D+39.4 | 2 | ||
| Hawaii | Josh Green | Gary Cordery | Solid D | >99% | <1% | D+28.9 | D+18.8 to D+39.3 | 0 | ||
| Massachusetts | Maura Healey | Mike Minogue | Solid D | >99% | <1% | D+26.3 | D+18.1 to D+34.6 | 6 |
Where the AI surveys agree with the polls
AI surveys and polls generally agree on which states favour each party but they often differ on the size of the lead. Across 33 states with polling data their estimates have a correlation of 0.85 and agree on the leading party in 26 states. Overall the AI estimates are about 7.2 percentage points more Republican than the polls.
The differences are particularly large in Oklahoma, Idaho, Tennessee and South Dakota where the AI is much more Republican than the polls. New Hampshire shows the opposite pattern with the AI favouring the Democrat while the polls favour the Republican. We correct these differences before using the AI estimates in the forecast and give more weight to polls where they are available.
Asking the model for a forecast instead
We also asked the same three models to function as forecasters (Method C) and compared their answers to the polls in the 33 states that have them (Table 2). Method C favoured the same candidate as the polls in 30 of the 33 states, compared to 26 for the raw AI surveys. It was 10.3 points from the poll margin on average (vs 12.4 for the raw AI surveys and 6.2 after their correction). Like the AI surveys, it leaned Republican (9.3 points on average). Its worst performances were in Vermont (R+27.7 vs R+2.5 in the polls), Nebraska (R+27.2 vs R+3.2), Oklahoma (R+26.3 vs R+4.8) and Rhode Island, where it had the Democrat only 19 points ahead of the Republican (vs 48 in the polls). We report Method C purely for comparison and avoid including it in the forecast. This is because a model asked to forecast is likely to be drawing on the polls, the ratings and past results it has already seen, which are sources the forecast already uses.
| State | Past results | AI surveys (uncalibrated) | Method C | Polls | Forecast | |
|---|---|---|---|---|---|---|
| Method A | Method B | |||||
| Georgia | D+1.0 | R+5.1 | R+5.0 | R+6.2 | R+1.8 | D+0.1 |
| Nevada | R+10.4 | R+6.2 | R+5.8 | R+3.3 | R+0.8 | R+0.4 |
| Ohio | R+2.5 | R+12.5 | R+12.4 | R+11.0 | D+3.9 | D+1.5 |
| Vermont | D+5.3 | D+3.3 | D+5.7 | R+27.7 | R+2.5 | D+2.3 |
| Florida | R+2.5 | R+6.9 | R+10.9 | R+12.7 | R+4.0 | R+2.3 |
| Alaska | R+3.4 | R+12.7 | R+26.1 | R+18.3 | R+1.4 | R+2.9 |
| Texas | R+14.8 | R+8.7 | R+14.8 | R+13.3 | R+2.5 | R+3.1 |
| New Hampshire | R+7.8 | D+5.5 | D+4.2 | R+4.9 | R+10.6 | R+3.5 |
| Iowa | R+3.4 | R+13.3 | R+12.1 | R+6.5 | D+7.8 | D+3.6 |
| Kansas | R+5.1 | R+16.3 | R+18.6 | R+12.6 | R+3.1 | R+3.7 |
| Wisconsin | D+1.8 | R+4.8 | R+6.7 | D+0.2 | D+4.5 | D+3.6 |
| Michigan | D+1.8 | R+4.0 | R+3.8 | D+0.8 | D+7.0 | D+5.4 |
| Arizona | D+11.5 | R+8.1 | R+11.7 | R+0.4 | D+7.8 | D+5.4 |
| Rhode Island | D+8.8 | D+37.1 | D+28.6 | D+19.1 | D+47.7 | D+35.7 |
| Nebraska | R+18.3 | R+22.5 | R+21.9 | R+27.2 | R+3.2 | R+5.9 |
| Oregon | D+20.2 | D+11.6 | D+9.8 | D+6.6 | D+1.3 | D+7.0 |
| South Carolina | R+5.1 | R+11.2 | R+18.0 | R+18.5 | R+14.5 | R+9.1 |
| Minnesota | D+4.5 | D+13.9 | D+8.9 | D+13.4 | D+9.0 | D+9.9 |
| Oklahoma | R+13.0 | R+37.8 | R+35.1 | R+26.3 | R+4.8 | R+10.3 |
| New Mexico | D+5.3 | D+22.1 | D+15.8 | D+9.1 | D+8.7 | D+11.5 |
| Alabama | R+11.2 | R+29.7 | R+26.0 | R+24.1 | R+13.4 | R+12.0 |
| Colorado | D+7.1 | D+15.8 | D+11.0 | D+9.5 | D+12.5 | |
| Maine | D+5.3 | D+16.2 | D+11.2 | D+6.8 | D+12.4 | D+12.3 |
| Arkansas | R+22.6 | R+24.2 | R+28.8 | R+26.5 | R+18.0 | R+14.9 |
| Tennessee | R+10.4 | R+43.6 | R+36.1 | R+29.0 | R+16.1 | R+16.1 |
| New York | D+20.2 | D+29.8 | D+20.4 | D+11.1 | D+13.2 | D+16.4 |
| Illinois | D+18.5 | D+18.8 | D+11.9 | D+11.5 | D+20.1 | D+17.7 |
| South Dakota | R+22.6 | R+44.6 | R+39.4 | R+30.1 | R+17.8 | R+18.6 |
| Pennsylvania | D+12.4 | D+4.3 | D+2.9 | D+8.4 | D+27.0 | D+18.7 |
| Connecticut | D+20.2 | D+30.2 | D+19.8 | D+12.2 | D+19.5 | D+19.8 |
| Wyoming | R+18.2 | R+41.6 | R+46.4 | R+43.2 | R+19.4 | |
| California | D+12.3 | D+29.2 | D+25.0 | D+17.6 | D+21.8 | D+20.9 |
| Idaho | R+25.2 | R+46.8 | R+52.3 | R+36.9 | R+21.9 | R+22.6 |
| Maryland | D+26.3 | D+43.3 | D+34.4 | D+24.0 | D+34.6 | D+31.2 |
| Hawaii | D+24.6 | D+50.1 | D+32.3 | D+29.9 | D+28.9 | |
| Massachusetts | D+25.4 | D+40.8 | D+35.2 | D+21.4 | D+25.4 | D+26.3 |
How the forecast is made
We create 3,600 voter personas based on Cooperative Election Study respondents with 100 personas sampled for each state. Three language models answer survey questions as these personas or as groups of them. We adjust their responses to better represent each state’s voters and correct the estimates using available polls. We then combine the corrected AI estimates with polls and historical election results. Finally we simulate the election 10,000 times to explore possible outcomes. The sections below explain each step.
Data
The personas and the weighting frame come from the 2024 and 2025 Cooperative Election Study (Schaffner et al. 2025, 2026), which 01a downloads from Harvard Dataverse and checks against stored checksums. Candidates, race ratings and past results for 2014, 2018 and 2022 come from Wikipedia (Wikipedia contributors 2026) through 01c, and 01b scrapes each state’s table of polls. Of the 443 polls scraped, the 280 with fieldwork ending on or after 1 August 2026 are used in the forecast (Table 3). The poll average weights each poll by its sample size and its age, so a poll a month older counts half as much. Colorado, Hawaii and Wyoming had no polls.
| State | Polls | Likely voters | Latest fieldwork | Poll average |
|---|---|---|---|---|
| Texas | 35 | 89% | 21 Oct | R+2.5 |
| Michigan | 21 | 86% | 18 Oct | D+7.0 |
| Florida | 17 | 94% | 20 Oct | R+4.0 |
| Iowa | 17 | 94% | 19 Oct | D+7.8 |
| Ohio | 16 | 88% | 20 Oct | D+3.9 |
| Arizona | 13 | 85% | 20 Oct | D+7.8 |
| Maine | 13 | 85% | 20 Oct | D+12.4 |
| New York | 12 | 83% | 23 Oct | D+13.2 |
| California | 11 | 100% | 20 Oct | D+21.8 |
| Pennsylvania | 11 | 55% | 21 Oct | D+27.0 |
| Georgia | 10 | 100% | 20 Oct | R+1.8 |
| Connecticut | 9 | 78% | 12 Oct | D+19.5 |
| New Hampshire | 9 | 100% | 20 Oct | R+10.6 |
| Nebraska | 9 | 67% | 18 Oct | R+3.2 |
| Minnesota | 7 | 86% | 12 Oct | D+9.0 |
| Alaska | 7 | 100% | 18 Oct | R+1.4 |
| Wisconsin | 7 | 86% | 9 Oct | D+4.5 |
| Kansas | 7 | 100% | 21 Oct | R+3.1 |
| Massachusetts | 6 | 67% | 18 Oct | D+25.4 |
| Rhode Island | 6 | 100% | 11 Oct | D+47.7 |
| New Mexico | 4 | 75% | 9 Oct | D+8.7 |
| Oregon | 4 | 100% | 4 Oct | D+1.3 |
| South Carolina | 4 | 50% | 18 Oct | R+14.5 |
| Nevada | 3 | 100% | 13 Oct | R+0.8 |
| Oklahoma | 3 | 100% | 13 Oct | R+4.8 |
| Tennessee | 3 | 100% | 15 Oct | R+16.1 |
| Vermont | 3 | 100% | 4 Oct | R+2.5 |
| Alabama | 3 | 67% | 17 Oct | R+13.4 |
| Maryland | 2 | 100% | 27 Sep | D+34.6 |
| Illinois | 2 | 100% | 22 Oct | D+20.1 |
| Idaho | 2 | 50% | 22 Oct | R+21.9 |
| South Dakota | 2 | 100% | 14 Oct | R+17.8 |
| Arkansas | 2 | 100% | 13 Oct | R+18.0 |
Personas and the weighting frame
For each of the 36 states, 100 registered voters were selected from the CES in proportion to their survey weight (without replacement, fixed seed). Respondents from 2025 were prioritized, and in the 10 states where the 2025 survey included fewer than 100 individuals, additional respondents were drawn from 2024. As a result, 3,068 personas are from 2025 and 532 are from 2024. Each persona consists of a short description based on age, gender, state, race and ethnicity, education, family income, party affiliation, and ideology. Additionally, 10 personas per state were used for validation checks.
The personas illustrate how each voter type would respond, and the weighting frame indicates the number of each type present in the state. The frame divides each state into 24 groups by party (three categories), age (four bands), and whether the individual holds a four-year degree. It is constructed from all 54,764 eligible CES respondents, with weights from each year rescaled to average one prior to pooling the two years. The 2024 survey contributes 76–88% of the weight for each state. In Method A, persona responses are weighted by their stated likelihood of voting (definitely 1, probably 0.7, unlikely 0.2, will not 0), averaged within each group, and the groups are then combined according to their share of the state’s frame. Method B’s group answers are combined the same way. In Method A, a group with no personas borrows the rate of its party group within that state. The frame matters as much as the models. CES 2025 included only 113 Arkansan respondents, who leaned Democratic; consequently, half of the Arkansas personas are Democrats (compared to 34% in the frame; Table 5). Before the frame was built from all eligible respondents, the Method A estimate for Arkansas was D+12; with the pooled frame, it is R+24. Effective sample sizes range from 47 in Alaska to 2,708 in California, indicating that frames for smaller states are also subject to greater uncertainty.
| Personas | Weighted frame | Difference | |
|---|---|---|---|
| Party | |||
| Democrats (incl. leaners) | 44% | 43% | +1 pts |
| Independents | 15% | 12% | +3 pts |
| Republicans (incl. leaners) | 40% | 45% | −4 pts |
| Age | |||
| 18-29 | 14% | 13% | +1 pts |
| 30-44 | 23% | 23% | −1 pts |
| 45-64 | 33% | 34% | −1 pts |
| 65+ | 30% | 30% | 0 pts |
| Education | |||
| Four-year degree | 40% | 40% | 0 pts |
| No four-year degree | 60% | 60% | 0 pts |
| State | CES respondents | Effective sample | Weighted share of registered voters | By survey year | |||
|---|---|---|---|---|---|---|---|
| Democrats | Independents | Republicans | Republican, 2024 survey | Republican, 2025 survey | |||
| Alaska | 141 | 47 | 31% | 21% | 49% | 48% | 53% |
| Wyoming | 101 | 51 | 26% | 6% | 68% | 71% | 56% |
| Vermont | 139 | 84 | 58% | 16% | 25% | 25% | 25% |
| Idaho | 348 | 112 | 22% | 14% | 64% | 64% | 65% |
| Hawaii | 221 | 123 | 50% | 20% | 29% | 29% | 30% |
| South Dakota | 201 | 135 | 26% | 13% | 61% | 57% | 74% |
| Rhode Island | 233 | 154 | 51% | 22% | 27% | 28% | 20% |
| Nebraska | 397 | 187 | 37% | 10% | 54% | 55% | 50% |
| New Hampshire | 324 | 224 | 50% | 14% | 37% | 34% | 51% |
| Maine | 337 | 226 | 46% | 14% | 39% | 40% | 33% |
| New Mexico | 490 | 240 | 52% | 13% | 35% | 37% | 26% |
| Arkansas | 554 | 243 | 34% | 11% | 55% | 60% | 32% |
| Nevada | 848 | 266 | 44% | 14% | 42% | 43% | 38% |
| Kansas | 554 | 289 | 35% | 16% | 49% | 52% | 38% |
| Iowa | 714 | 317 | 38% | 12% | 50% | 50% | 47% |
| Connecticut | 689 | 385 | 53% | 14% | 33% | 34% | 29% |
| Oklahoma | 777 | 386 | 29% | 9% | 62% | 63% | 58% |
| Alabama | 1,053 | 490 | 34% | 11% | 55% | 56% | 52% |
| South Carolina | 1,147 | 524 | 40% | 9% | 51% | 50% | 53% |
| Colorado | 1,259 | 569 | 47% | 14% | 39% | 41% | 32% |
| Tennessee | 1,528 | 589 | 29% | 10% | 61% | 63% | 54% |
| Arizona | 1,739 | 599 | 38% | 11% | 50% | 51% | 48% |
| Oregon | 1,225 | 636 | 50% | 12% | 38% | 38% | 36% |
| Minnesota | 1,203 | 637 | 48% | 11% | 41% | 40% | 41% |
| Massachusetts | 1,273 | 659 | 59% | 13% | 27% | 28% | 26% |
| Maryland | 1,255 | 679 | 59% | 13% | 28% | 28% | 27% |
| Wisconsin | 1,499 | 726 | 42% | 10% | 47% | 47% | 50% |
| Georgia | 2,423 | 939 | 44% | 10% | 46% | 48% | 39% |
| Michigan | 2,332 | 1,145 | 44% | 12% | 45% | 45% | 44% |
| Ohio | 2,841 | 1,358 | 40% | 9% | 51% | 52% | 49% |
| Illinois | 2,677 | 1,423 | 51% | 11% | 38% | 39% | 34% |
| Florida | 5,194 | 1,722 | 42% | 10% | 49% | 49% | 49% |
| Pennsylvania | 3,507 | 1,862 | 47% | 7% | 46% | 46% | 47% |
| New York | 4,132 | 1,927 | 57% | 11% | 32% | 32% | 33% |
| Texas | 5,454 | 2,532 | 40% | 11% | 49% | 50% | 48% |
| California | 5,955 | 2,708 | 57% | 11% | 32% | 32% | 31% |
Asking the models
We sent all prompts through OpenRouter to three models (Table 6): OpenAI’s GPT-6 Luna, DeepSeek V4.1 Flash, and Mistral Small 24B. Every model was locked to a single host provider so that our daily runs always hit the same deployment. For DeepSeek and Mistral, we set the temperature to 1.0. GPT-6 Luna blocks custom temperature settings, so we left it on its default temperature with reasoning set to minimal. Pulling from three separate model families keeps our forecast from leaning entirely on one lab’s training decisions, even if biases common to all three models still carry over.
We ran three different prompt formats:
Method A (individual personas): we fed individual voter personas into each model and asked three things: how likely they were to vote in the 2026 midterms, who they would pick for governor, and who they would back for the U.S. House. Each persona went to every model three times.
Method B (groups): rather than prompting one voter at a time, we broke each state’s electorate into 24 buckets by party (3), age (4), and college degree (2). We then asked the models to give the turnout and candidate vote split for each bucket, adding up to 864 calls per model a day.
Method C (direct forecasting): we asked the models to act as election forecasters directly. For every race, they gave the Democratic two-party vote share, the uncertainty around that number, and each candidate’s chance of winning. This ran three times per state every day.
In every method, we shuffled the candidate order using a fixed seed to keep the prompts reproducible, and required the models to reply in JSON using only a preset list of valid values. Whenever a reply came back blank or broken, we ran that prompt one more time and saved all raw outputs. Taking all days and methods together, 100% of DeepSeek’s replies were usable, while GPT-6 Luna and Mistral both came in at 99.8%.
| Model | OpenRouter identifier | Host | Reasoning | Temperature | Answers | Usable |
|---|---|---|---|---|---|---|
| GPT-6 Luna (OpenAI) | openai/gpt-6-luna | openai | minimal | provider default | 22,660 | 99.8% |
| DeepSeek V4.1 Flash | deepseek/deepseek-v4.1-flash | deepinfra/fp8 | none | 1 | 22,620 | 100.0% |
| Mistral Small 24B | mistralai/mistral-small-24b-instruct-2501 | deepinfra/fp8 | not offered | 1 | 22,654 | 99.8% |
System message
You are taking part in a public opinion survey. You will be given a description of a person. Answer every question as that person would, based on their background, values and circumstances. Do not answer as an AI assistant, do not add commentary, and do not refuse. Reply only with valid JSON in the exact format requested.
Question
You are a 63-year-old man who lives in Michigan. You describe your race/ethnicity as White. Your highest level of education is some college, and your family income is $10,000 to $19,999. In politics, you consider yourself a strong Democrat and your views are very liberal. You are registered to vote.
The 2026 general election for Governor of Michigan is on Tuesday,
November 3. The candidates are:
- Jocelyn Benson (Democratic Party)
- John James (Republican Party)
Question 1. How likely are you to vote in this election?
Options: "Definitely will vote", "Probably will vote",
"Unlikely to vote", "Will not vote"
Question 2. If the election were held today, who would you vote for
for Governor?
Options: "Jocelyn Benson", "John James", "Other", "Don't know"
Question 3. In the election for the U.S. House of Representatives in your
district, which party's candidate would you vote for?
Options: "The Democratic Party candidate", "The Republican Party
candidate", "Other", "Don't know"
Reply in this JSON format:
{"turnout": "<one option>", "governor_vote": "<one option>",
"house_vote": "<one option>"}GPT-6 Luna (OpenAI) replied
{"turnout":"Definitely will vote","governor_vote":"Jocelyn Benson","house_vote":"The Democratic Party candidate"}
| Democrat | Republican | Independent or third party | Other | Don't know | Unusable | Total | |
|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | 5,248 | 4,902 | 57 | 0 | 592 | 1 | 10,800 |
| GPT-6 Luna (OpenAI) | 5,282 | 4,713 | 57 | 0 | 748 | 0 | 10,800 |
| Mistral Small 24B | 5,242 | 4,701 | 56 | 26 | 769 | 6 | 10,800 |
| All models | 15,772 | 14,316 | 170 | 26 | 2,109 | 7 | 32,400 |
| Model | Answers | Split ticket |
|---|---|---|
| DeepSeek V4.1 Flash | 10,102 | 0.9% |
| Mistral Small 24B | 9,777 | 1.4% |
| GPT-6 Luna (OpenAI) | 9,891 | 1.1% |
Combining the sources and simulating the election
We combined historical election results, polling data and AI survey estimates to produce the final forecast.
First, we used a regression model to estimate the Democratic vote margin based on the state’s political lean, whether an incumbent was running and which party holds the presidency. The results suggested that an incumbent running for re-election benefits their party by around 11 percentage points. The president’s party faces a disadvantage of approximately 3.5 points during a midterm election. The coefficient for state political lean was 0.87.
| Term | Estimate | Standard error |
|---|---|---|
| Intercept | −1.64 | 1.35 |
| State lean (Cook PVI, per point) | 0.87 | 0.13 |
| Incumbent running (+1 D, -1 R) | 11.39 | 1.69 |
| President's party (+1 D, -1 R) | −3.49 | 1.32 |
Next we compared the AI survey results with polling averages. The raw AI estimates were around 7.2 percentage points more Republican than the polls.
Using data from the 33 states with available polls, we fitted the following relationship:
Poll estimate = 5.9 + 0.58 × AI survey estimate
The slope of 0.58 suggests that the AI models produced more extreme results than the polls. Their estimates tended to exaggerate the political lean of strongly Democratic and strongly Republican states.
We then combined the three sources based on their estimated precision. On average the polls received 55% of the weight, AI surveys received 33%, and historical results received 12%. These weights varied between states depending on the available data and the level of disagreement between estimates.
For example, in Vermont, the three models and two methods disagreed widely. As a result, the AI surveys received only 7% of the weight, while the three polls received 80%.
Finally, we simulated the election 10,000 times. Each simulation included a national swing shared across states and additional state-level uncertainty. The candidate with the most votes was considered the winner, subject to the rules of each state’s election.
Georgia was treated differently because a candidate needs more than 50% of the vote to avoid a runoff. In our simulations when no candidate received more than 50% of the vote we counted the race as going to a runoff. The winner we show for Georgia is the candidate leading the first round.
Our simulations estimated a 36% chance of a runoff. We also kept independent candidates separate from the major parties. In Rhode Island, Ken Block received approximately 34% support in the polls and won around 18% of the simulated elections.
Validation, costs and timing
Do the answers react to the right things?
To see what the answers depend on, we took 10 personas per state with a 2026 governor’s race (360 personas in total) and asked them again with one detail of the prompt changed. To establish a baseline, we also asked the identical question a second time, which changed 3.9% to 4.7% of governor votes (Table 10). The most impactful change came from removing the persona’s party, which changed 16% to 20% of votes (four to five times the baseline), and party-line voting fell from 97% to 80%. Removing party labels from the candidates changed 5% to 12% of votes, with the greatest movement in Mistral’s responses. Reversing the candidate order or hiding the incumbent label altered 3% to 7% of votes, similar to the baseline, and every persona was able to identify their own state when asked. Our experiments show that the models barely react to ballot position, and their responses mostly follow the persona’s party. Argyle et al. (2023) found that models conditioned on a person’s background reproduce how groups vote, and Wang et al. (2025) found that they also make members of a group more alike than they are. Our party-drop result shows both, and it fits the AI surveys spreading the states too far apart: in our AI surveys, a persona’s party appears to decide its vote more often than it does for real voters.
| Votes that changed | |||
|---|---|---|---|
| DeepSeek V4.1 Flash | Mistral Small 24B | GPT-6 Luna (OpenAI) | |
| Asked again, nothing changed | 3.9% | 4.7% | 3.9% |
| Asked which state they live in | 4.4% | 4.7% | 4.7% |
| House question removed | 3.1% | 5.3% | 4.2% |
| Incumbent label removed | 4.2% | 5.6% | 3.3% |
| Candidate order reversed | 5.6% | 7.2% | 4.2% |
| Party labels on candidates removed | 6.4% | 11.9% | 5.0% |
| Persona's own party removed | 16.1% | 19.7% | 19.7% |
Do the answers hold from one day to the next?
Synthetic answers can drift with small changes to wording or timing, as Bisbee et al. (2024) found. We asked the models the same Method B, Method C and validation questions over six days: the full run on 25 October, followed by daily GitHub Actions runs from 28 October to 1 November (Table 11). On all measures, Mistral was the least stable model. Between days, a race’s margin in Method B shifted by 3.8 points for Mistral, 2.1 for DeepSeek and 1.4 for GPT-6 Luna. Method C was more consistent, moving 0.7 to 1.5 points between days, and 4% to 8% of the validation personas’ votes differed from the first run. The Method B shifts are slightly smaller than the spread between repeats on the same day, so most of the movement comes from the models’ own randomness rather than a change over time, and all of it is small next to the 7.2-point gap between the AI surveys and the polls.
| Days | Spread within a day | Spread between days | |
|---|---|---|---|
| B: group split, margin | |||
| DeepSeek V4.1 Flash | 6 | 2.4 | 2.1 |
| Mistral Small 24B | 6 | 4.8 | 3.8 |
| GPT-6 Luna (OpenAI) | 6 | 1.5 | 1.4 |
| C: forecaster, Democratic share | |||
| DeepSeek V4.1 Flash | 6 | 1.4 | 0.9 |
| Mistral Small 24B | 6 | 2.3 | 1.5 |
| GPT-6 Luna (OpenAI) | 6 | 1.0 | 0.7 |
Costs and timing
The 67,934 model calls over the six-day period cost US$3.55 in total (Table 12, Table 13). Of that, $2.39 was for the full run on 25 October, when we asked every method. Method A on its own made 32,440 calls and took 107 minutes to complete, with eight calls running at a time. On the other days we re-ran Methods B and C with the validation checks, and each of those runs took about 14 minutes and cost about $0.23. By model, GPT-6 Luna cost $1.95, because its reasoning tokens are billed as output, DeepSeek $1.18 and Mistral $0.43. By method, A cost $1.61, B $1.27, the validation checks $0.58 and C $0.10. The GitHub schedule was set for 10:23, but the job usually started about five hours late, which made no difference to a daily forecast.
| Calls | Input tokens | Output tokens | Seconds per call | Cost (US$) | |
|---|---|---|---|---|---|
| A: each persona votes | |||||
| DeepSeek V4.1 Flash | 10,800 | 3,841,797 | 370,981 | 0.8 | $0.60 |
| Mistral Small 24B | 10,800 | 4,038,170 | 391,266 | 0.8 | $0.23 |
| GPT-6 Luna (OpenAI) | 10,840 | 3,829,116 | 783,080 | 1.9 | $0.77 |
| B: each group's split | |||||
| DeepSeek V4.1 Flash | 6,912 | 1,430,400 | 318,459 | 0.6 | $0.33 |
| Mistral Small 24B | 6,920 | 1,558,088 | 375,374 | 2.0 | $0.11 |
| GPT-6 Luna (OpenAI) | 6,912 | 1,449,600 | 1,357,467 | 3.4 | $0.82 |
| C: model as forecaster | |||||
| DeepSeek V4.1 Flash | 648 | 78,354 | 25,402 | 0.6 | $0.02 |
| Mistral Small 24B | 670 | 85,586 | 30,591 | 4.1 | $0.01 |
| GPT-6 Luna (OpenAI) | 648 | 78,966 | 131,352 | 3.8 | $0.07 |
| Validation checks | |||||
| DeepSeek V4.1 Flash | 4,260 | 1,383,624 | 128,004 | 0.5 | $0.22 |
| Mistral Small 24B | 4,264 | 1,451,909 | 134,447 | 1.2 | $0.08 |
| GPT-6 Luna (OpenAI) | 4,260 | 1,385,279 | 275,782 | 1.8 | $0.28 |
| All methods and models | |||||
| Total | 67,934 | 20,610,889 | 4,322,205 | $3.55 | |
| Day | Run | Calls | Cost (US$) |
|---|---|---|---|
| 25 October | Full run on a laptop | 47,020 | $2.39 |
| 28 October | Daily GitHub Actions run | 4,196 | $0.23 |
| 29 October | Daily GitHub Actions run | 4,179 | $0.23 |
| 30 October | Daily GitHub Actions run | 4,181 | $0.23 |
| 31 October | Daily GitHub Actions run | 4,180 | $0.24 |
| 1 November | Daily GitHub Actions run | 4,178 | $0.23 |
What the ranges include, and what they leave out
Understanding the forecast ranges
We ran 10,000 election simulations to understand how much the results could vary. The 80% range covers the outcomes from 80% of the simulations, while the 95% range covers 95% of them.
A 60% chance of winning does not mean that a result is guaranteed. It means the candidate could still lose in about four out of ten simulations.
Our forecast gave Democrats a median of 27 governorships with an 80% range of 21 to 32. The range is quite wide because many races are competitive and a national shift in voter preferences can affect several states at once.
Sources of uncertainty
One important source of uncertainty is the national swing. We assumed a standard deviation of 4 percentage points. This accounts for the possibility that polls may overestimate or underestimate support across several states in the same direction. Such errors can make the final number of seats vary considerably between simulations.
We also included uncertainty for each data source. After calibration the AI surveys had an estimated standard deviation of around 6.4 points. Polling estimates had an error of at least five points, even when several polls were available. Historical results had a standard deviation of approximately 12.9 points making them the least precise source.
We tested the historical model on 107 previous elections that were not used to fit it. The model identified the correct winner in 87% of these races but its average error was around 13 points. Its Brier score was 0.11 compared with 0.25 for a simple 50–50 prediction. This suggests that historical results can help identify likely winners, although they are less reliable for predicting the exact vote margin.
Limitations of uncertainty estimates
There are several limitations to consider when interpreting these forecasts.
The three AI models may share similar biases because their training data could contain overlapping information. Using different models does not necessarily remove these biases or make their estimates fully independent.
The AI survey estimates were calibrated using this year’s published polls. If those polls contain systematic errors, the calibration process may carry those errors into the final forecasts.
The synthetic voters were created using respondents from the 2024 and 2025 Cooperative Election Study (CES). Public opinion may have changed since those surveys were conducted, meaning these synthetic voters might not fully represent the views of the current electorate.
Colorado, Wyoming and Hawaii had no polling data in our dataset. Their forecasts therefore rely more heavily on AI survey estimates and historical election results. This makes their actual uncertainty potentially greater than the reported estimates suggest.
The simulations are based on information available up to 1 November. They cannot account for later developments, such as campaign controversies, unexpected events or changes in voter turnout before election day.
Turnout was estimated using the models’ predictions of whether respondents would definitely or probably vote. This approach is not equivalent to using a dedicated likely-voter model, which could produce different estimates of the electorate.
Uncertainty also affects the overall number of governorships each party may win. Across the 10,000 simulated elections, the middle 80% of outcomes gives a range of 21 to 32 Democratic governorships, rather than a fixed result. This range reflects uncertainty across individual races as well as the shared national swing applied in the simulations. It should therefore be interpreted as a range of possible outcomes, not a guarantee of the final result.
The models relied heavily on each persona’s party. Removing the persona’s own party changed 16% to 20% of governor votes, while removing the party labels next to the candidates’ names changed 5% to 12%. Party may therefore shape the answers more than other information about the voter or the candidates.
Discussion
Our silicon sampling forecast puts the median number of Democratic governors at 27, with a 61% chance that Democrats hold 26 or more seats and a 28% chance for Republicans. Concerning the overall result, our forecast aligns with that of the professional forecasters. Assuming that each race goes to whichever candidate the forecasters favour in their October ratings, as assembled on Wikipedia (Wikipedia contributors 2026), the ratings of Cook Political Report would result in 24 to 28 governors for Democrats, and those of Sabato’s Crystal Ball and Silver Bulletin would result in 26 to 28. Nevertheless, there are two states where we forecast differently from professional forecasters. This section attempts to answer the question whether such a difference indicates that we discovered something the experts overlooked or that we overstepped the boundaries of our approach.
The total agrees with the experts, two races do not
In every race where the majority of the seven rating agencies in our dataset favour one party, we favour that party, except for two races. Georgia (even odds), Nevada (Republican 53%), and Ohio (Democrat 59%) are toss-ups in our assessment and also rated so by the agencies. In Alaska and Vermont, however, we see different preferences from the agencies.
In Alaska we give the Republican side a 67% chance. There are five out of seven agencies who rate this race as favouring Democrats while Silver Bulletin rates this race as Solidly Democratic. At least some part of the difference is due to the way we model the race. Alaska uses ranked choice voting and we treat the final round of voting by adding up the percentage share of three Republican candidates. This implies that every Republican voter ranks another Republican second and that none of them rank the Democratic candidate above another Republican. The polls, when summed in the same way, show an advantage of Republicans by 1.4 points.
The probability that the Democrat beats the incumbent, Phil Scott, is 62% in Vermont. All seven agencies predict a win for Phil Scott. The polls predict a lead for him by 2.5 points, while the corrected AI surveys predict a lead for the Democrat by 8.5. Vermont is one of the most Democratic states in America. If we have a persona that leans Democratic, they will vote for the Democrat no matter what the record of the Republican running is. The personas do not capture Scott’s appeal to Democrats, and the reason is shown below.
If both races went the way the agencies expect, Democrats would take Alaska and lose Vermont. There are 21 races in which we favour Democrats. Adding the 6 Democratic governors whose seats are not up this year, we have 27. The two discrepancies offset each other. A forecast can agree with the experts on the top line while disagreeing with them on individual races, making the seats tally an unreliable measure of a technique’s validity.
The synthetic voters know party, not candidates
Argyle et al. (2023) found that GPT-3, given the background of real survey respondents, reproduced how groups of Americans voted. They called this “algorithmic fidelity”. Our results support the coarse version of that claim. Across the 33 states with polls, the raw AI-survey margin correlates 0.85 with the poll average (Figure 3), so the models rank the states much as the polls do.
The finer version fails to apply. When we regress the poll margin on the AI-survey margin, the slope is 0.58: the AI surveys spread the states too far apart. Massachusetts is D+38 in the AI surveys and D+25 in the polls; Idaho is R+50 against R+22. Between 96.5% and 98.4% of the Democratic and Republican validation personas vote for their own party’s candidate, depending on the model. Removing party from the persona changes 16% to 20% of the votes, while removing any reference to incumbency affects 3% to 6%, about the same as asking the identical question twice (4% to 5%). The models choose a candidate by party affiliation and barely react to incumbency. Bisbee et al. (2024) showed the same phenomenon to occur with ChatGPT: while its synthetic answers corresponded to the average opinion, there was not enough variance. Wang et al. (2025) call this phenomenon flattening, where the model treats all members of a group as its archetype. A state full of typical partisans is a safer state than a real one, which is what our slope shows.
Moreover, these misses go against the narrative of bias. Santurkar et al. (2023) noted that fine-tuned language models are more likely to be biased towards liberals with high education levels in America. While our surveys are 7.2 points more Republican than the polls on average, in New Hampshire our surveys show an advantage of 4.8 points to the Democrat while polls favour the incumbent Republican, Kelly Ayotte, by 10.6 percentage points, and in Vermont, our surveys are again tilted towards the Democrat. Pennsylvania misses the other way: D+3.6 in the AI survey against D+27 in the polls for the Democratic incumbent, Josh Shapiro. A single lean cannot explain all three misses. But ignoring the incumbency effect can – in each state, our surveys are close to where just considering the party predicts while polls have a governor taking voters from the other party. Alaska and Vermont, where we disagree with the professional forecasters, will be the first test of this argument when the results come in on 3 November.
Reproducing this page
Every number, table and chart on this page is made by the scripts in the repository, and its README lists the order to run them. The analysis uses Python (Van Rossum and Drake 2009) with pandas (McKinney 2010), NumPy (Harris et al. 2020), SciPy (Virtanen et al. 2020), Matplotlib (Hunter 2007), pandera (Bantilan 2020), great_tables (Iannone and Chow 2026), PyArrow (Apache Arrow developers 2016), Requests (Reitz 2011), HTTPX (Encode 2019), Beautiful Soup (Richardson 2004), pyreadstat (Fajardo 2018) and PyYAML (The YAML Project 2006), with packages managed by uv (Astral 2024); the page is built with Quarto (Allaire et al. 2022). The models were called through OpenRouter (OpenRouter 2026). Polls and candidates come from Wikipedia (Wikipedia contributors 2026), and the map uses US Census state outlines (U.S. Census Bureau 2018). The fonts Karla and Lekton are used under the SIL Open Font License.
| Setting | Value |
|---|---|
| Personas per state | 100, drawn in proportion to CES weight |
| Validation personas per state | 10 |
| Groups per state (Method B) | 24: party x age x degree |
| Repeats in the full run | 3 |
| Temperature | 1 (provider default for GPT-6 Luna) |
| Simulated elections | 10,000 |
| Random seed (forecast) | 853 |
| Shared national swing, SD | 4 points |
| Extra polling error, SD | 5 points |
| Polls used | 280, fieldwork ending on or after 2026-08-01 |
References
Allaire, J. J., Charles Teague, Carlos Scheidegger, Yihui Xie, and Christophe Dervieux. 2022. Quarto. Released. https://doi.org/10.5281/zenodo.5960048.
Apache Arrow developers. 2016. PyArrow: Python Bindings for Apache Arrow. Released. https://arrow.apache.org.
Argyle, Lisa P., Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. 2023. “Out of One, Many: Using Language Models to Simulate Human Samples.” Political Analysis 31 (3): 337–51. https://doi.org/10.1017/pan.2023.2.
Astral. 2024. Uv: An Extremely Fast Python Package and Project Manager. Released. https://github.com/astral-sh/uv.
Bantilan, Niels. 2020. “pandera: Statistical Data Validation of Pandas Dataframes.” Proceedings of the 19th Python in Science Conference, 116–24. https://doi.org/10.25080/Majora-342d178e-010.
Bisbee, James, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson. 2024. “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models.” Political Analysis 32 (4): 401–16. https://doi.org/10.1017/pan.2024.5.
Encode. 2019. HTTPX. Released. https://www.python-httpx.org.
Fajardo, Otto. 2018. Pyreadstat. Released. https://github.com/Roche/pyreadstat.
Harris, Charles R., K. Jarrod Millman, Stéfan J. van der Walt, et al. 2020. “Array Programming with NumPy.” Nature 585: 357–62. https://doi.org/10.1038/s41586-020-2649-2.
Hunter, J. D. 2007. “Matplotlib: A 2D Graphics Environment.” Computing in Science & Engineering 9 (3): 90–95. https://doi.org/10.1109/MCSE.2007.55.
Iannone, Richard, and Michael Chow. 2026. great_tables: Make Awesome Display Tables Using Python. V. 1.1.0. Released. https://posit-dev.github.io/great-tables/.
McKinney, Wes. 2010. “Data Structures for Statistical Computing in Python.” Proceedings of the 9th Python in Science Conference, 56–61. https://doi.org/10.25080/Majora-92bf1922-00a.
OpenRouter. 2026. OpenRouter: A Unified Interface for Large Language Models. https://openrouter.ai.
Reitz, Kenneth. 2011. Requests: HTTP for Humans. Released. https://requests.readthedocs.io.
Richardson, Leonard. 2004. Beautiful Soup. Released. https://www.crummy.com/software/BeautifulSoup/.
Santurkar, Shibani, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. 2023. “Whose Opinions Do Language Models Reflect?” Proceedings of the 40th International Conference on Machine Learning, Proceedings of machine learning research, vol. 202: 29971–30004.
Schaffner, Brian, Stephen Ansolabehere, and Marissa Shih. 2025. Cooperative Election Study Common Content, 2024. Harvard Dataverse. https://doi.org/10.7910/DVN/X11EP6.
Schaffner, Brian, Stephen Ansolabehere, and Marissa Shih. 2026. Cooperative Election Study Common Content, 2025. Harvard Dataverse. https://doi.org/10.7910/DVN/DFUGJR.
The YAML Project. 2006. PyYAML. Released. https://pyyaml.org.
U.S. Census Bureau. 2018. Cartographic Boundary Files: States. https://github.com/PublicaMundi/MappingAPI.
Van Rossum, Guido, and Fred L. Drake. 2009. Python 3 Reference Manual. CreateSpace.
Virtanen, Pauli, Ralf Gommers, Travis E. Oliphant, et al. 2020. “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python.” Nature Methods 17: 261–72. https://doi.org/10.1038/s41592-019-0686-2.
Wang, Angelina, Jamie Morgenstern, and John P. Dickerson. 2025. “Large Language Models That Replace Human Participants Can Harmfully Misportray and Flatten Identity Groups.” Nature Machine Intelligence 7: 400–411. https://doi.org/10.1038/s42256-025-00986-z.
Wikipedia contributors. 2026. 2026 United States Gubernatorial Elections. https://en.wikipedia.org/wiki/2026_United_States_gubernatorial_elections.