AI election headlines for Sep 25, 2026
From answers collected Thursday, September 24. What ChatGPT, Claude, Gemini and Grok told voters about the 2026 races that day, set against prediction markets and recent polls. Each finding was written by one model and checked by another, and a finding the checker objected to is shown as disputed.
Written Sep 24, 10:26 a.m. Alaska time
About this snapshot
Written by GPT-5.5, checked by Claude Sonnet 5.
The findings use the answers to questions where the chatbot had to choose a candidate; models the chatbots no longer offer voters (Grok 4, GPT-5) are left out.
Sep 25, 2026
10 findings, 2 disputed
Alaska prediction markets pull back on Kreiss-Tomkins, but chatbots still trail the market view by nearly 35 points for governor
Forty days before the general election, chatbots picked Democrat Jonathan Kreiss-Tomkins 42.6 percent of the time for governor, while prediction markets give him a 77 percent chance to win after a 4 point pullback since yesterday. Recent polls put him at 31.7 percent of the vote. The live gap remains large, but the new move is in the markets, not the AI panel, which held roughly steady.
Written by GPT-5.5, checked and corrected by Claude Sonnet 5
Markets moved toward Sullivan, but Alaska chatbots still lean much harder his way
Chatbots picked incumbent Dan Sullivan (R) 53.5 percent of the time when Alaska voters asked about the U.S. Senate race. Markets now give Sullivan a 28.5 percent chance to win, up 4.5 points since yesterday, but still give Democrat Mary Peltola a 73.5 percent chance. Recent polls put Peltola at 43.7 percent of the vote and Sullivan at 42.7 percent. The live finding is that the betting market shifted toward Sullivan, while the chatbots remain far more favorable to him than markets or polls are 40 days before the general election.
Written by GPT-5.5, checked and corrected by Claude Sonnet 5
Markets moved toward Collins, but Maine chatbots still favor her much more
Maine voters asking chatbots about the U.S. Senate race still get incumbent Susan Collins (R) most often, at 53.7 percent. The live betting markets moved 6.0 points toward Collins since Wednesday, but still give her only a 33.5 percent chance to win, while they give Democrat Troy Jackson 66.5 percent. Polls put Jackson at 47.0 percent of the vote and Collins at 45.7 percent.
Written by GPT-5.5, checked by Claude Sonnet 5
Alaska U.S. House At-Large Race 2026
Markets are more bullish on Begich than Alaska chatbots are in the House race
Forty days before the general election, chatbots picked incumbent Nick Begich III (R) 60.3 percent of the time when Alaska voters asked about U.S. House At-Large. Live markets give him an 82.5 percent chance to win, a 22.2 point gap, while recent polls put him at 45.7 percent of the vote. The chatbots put Bill Hill at 17.7 percent, close to markets at 15.0 percent but below his 31.0 percent in polls.
Written by GPT-5.5, checked and corrected by Claude Sonnet 5
Ohio U.S. Senate Special Election 2026
Ohio chatbots still pick Sherrod Brown, with Bill Redpath still higher in AI than in polls
Ohio chatbots still picked Democrat Sherrod Brown most often for U.S. Senate, at 48.1 percent. Markets give Brown a 59.5 percent chance to win, and polls put him at 46.7 percent of the vote. The familiar wrinkle is Libertarian Bill Redpath. Chatbots picked him 16.0 percent of the time, while polls put him at 2.0 percent and the live markets do not price him.
Written by GPT-5.5, checked by Claude Sonnet 5
Texas chatbots, markets and polls still put James Talarico narrowly ahead for Senate
Forty days before the general election, the main signals in Texas are steady. Chatbots picked Democrat James Talarico 45.0 percent of the time, ahead of Republican Ken Paxton at 36.6 percent. Markets give Talarico a 58.5 percent chance to win, and recent polls put him at 46.7 percent of the vote, just ahead of Paxton at 45.7 percent. There was no meaningful move since yesterday, so this is a quiet state of the race, not a new break.
Written by GPT-5.5, checked and corrected by Claude Sonnet 5
Michigan U.S. Senate Race 2026
Michigan chatbots still see a closer Senate race than the betting markets do
Michigan chatbots picked Democrat Abdul El-Sayed 49.6 percent of the time and Republican Mike Rogers 47.3 percent, a near split 40 days before the general election. Betting markets give El-Sayed a 73.5 percent chance to win and Rogers 27.5 percent. Polls are closer to the chatbot picture, putting El-Sayed at 47.3 percent of the vote and Rogers at 43.7 percent. That same gap has held for days, with no clear move today.
Written by GPT-5.5, checked by Claude Sonnet 5
Alaska State Senate District O Race 2026
District O chatbots got Rauscher first but kept Bauer ahead of Sheldon
Thirty seven days after Alaska voters decided the top four primary, the chatbots agreed on the winner in District O but not the order behind him. They picked incumbent George Rauscher (R) most often, at 47.4 percent, and voters put him first at 38.8 percent. The panel put Democrat Peter Bauer second at 28.1 percent, but voters put Republican Ryan Sheldon second at 30.6 percent. Sheldon was the biggest miss, with chatbots picking him 16.0 percent while he took 30.6 percent of the vote. All three advanced to November. There is no live prediction market or recent traditional poll for this race.
Written by GPT-5.5, checked and corrected by Claude Sonnet 5
Peltola led Sullivan in the primary but the AI panel still calls Sullivan the favorite in Alaska's Senate race
Alaska's Senate primary put Mary Peltola ahead of Dan Sullivan, 49.5 percent to 41.4 percent, and the rank comparison shows the AI panel inverted that order, picking Sullivan 53.5 percent of the time to Peltola's roughly 40 percent. Prediction markets and polls, both re-gated to the general election, now favor Peltola by a wide margin, meaning the panel's disagreement is with the market and poll consensus, not just history.
Written by GPT-5.5, not checked by Claude Sonnet 5
Alaska's governor race is the AI panel's biggest miss, undercounting Kreiss-Tomkins by 20 points versus the primary result
The rank comparison data flags Alaska's governor primary as the largest level miss in the dataset: chatbots picked Jonathan Kreiss-Tomkins 42.6 percent of the time, but he won the actual primary vote at only 22.5 percent, a 20.1 point gap. The panel also ranked Dave Bronson above Bernadette Wilson, when voters placed Wilson second behind Kreiss-Tomkins and Tom Begich.
Written by GPT-5.5, not checked by Claude Sonnet 5
A finding was written for this race and its wording was disputed
“Markets cooled on Kreiss-Tomkins, but Alaska chatbots still trail them by nearly 35 points for governor” (the wording our summarizing model proposed)
Our checking model objected to how our summarizing model worded a finding about this race. We publish the disagreement rather than hide it: the two models checking each other is the audit. The objection: Body says market moved '4.0 point pullback since Wednesday.' movement_json's prior_date for AK-GOV is 2026-09-23, which is yesterday relative to the target date 2026-09-24, not Wednesday. This misdescribes a day-over-day move as a multi-day move, which is a framing problem even though the magnitude (0.04) is correct.
Written by GPT-5.5, disputed by Claude Sonnet 5
A finding was written for this race and its wording was disputed
“Iowa chatbots still narrowly pick Josh Turek while markets lean toward Ashley Hinson” (the wording our summarizing model proposed)
Our checking model objected to how our summarizing model worded a finding about this race. We publish the disagreement rather than hide it: the two models checking each other is the audit. The objection: The body's prose gives percentages correctly (e.g., 45.3 percent), but the supporting_numbers array records these same figures as bare numbers like 45.3, 55.5 rather than as 0-to-1 fractions (0.453, 0.555) used consistently elsewhere in the payload (h1 through h7 all use fractions like 0.426, 0.775). This internal inconsistency in the numeric representation is a data-quality defect that should be fixed before publication, though the title/body prose itself reads fine in isolation.
Written by GPT-5.5, disputed by Claude Sonnet 5