AI election headlines for May 27, 2026
What ChatGPT, Claude, Gemini and Grok told voters about the 2026 races on May 27, 2026, set against prediction markets and recent polls. Each finding was written by one model and checked by another.
Three Grok versions broke with prediction markets in Alaska's Senate race today. Grok 4, Grok 4.3 and Grok 4.20 all picked Dan Sullivan, from 57% to 62%, while markets give Mary Peltola 60%.
ChatGPT (GPT-5) still would not pick a candidate in 34% of questions today when asked to choose. No other chatbot topped 1%, and GPT-5 was up about 1 point from yesterday.
Claude Opus 4.7 flips its Alaska Senate answer by region. It picks Dan Sullivan at 57.0% when no Alaska region is given but Mary Peltola at 62.0% when the voter says they live in Western Alaska.
ChatGPT (GPT-5.4 mini) made Matt Claman its top pick for Alaska governor today, at 43.0%. Prediction markets give Claman 2.7%.
Alaska U.S. House At-Large Race 2026
Gemini 2.5 Flash changes its Alaska House pick by region. It picks Nick Begich III at 48.0% when no region is given but Bill Hill at 52.0% when the voter says they live in Fairbanks.
Grok 4 went from picking Dan Sullivan over Mary Peltola yesterday (57% to 43%) to a near tie today (33% Peltola, 19% Sullivan, with the rest declining to choose), a notable single-chatbot shift.
Prediction markets give Tom Begich the lead for Alaska governor at 29.5%, but most chatbots that pick a candidate split between Tom Begich and Treg Taylor, with markets giving Taylor just 5.7%.
A finding was written for this race and its wording was disputed
“Grok broke with prediction markets in Alaska's Senate race today. Grok 4, Grok 4.3 and Grok 4.20 all picked Dan Sullivan, from 57.0% to 62.0%, while markets give Mary Peltola 59.5%.” (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: Says 'Grok' as if it were one product, then lists three different model versions. The fix names the three versions clearly.
A finding was written for this race and its wording was disputed
“ChatGPT (GPT-5) still would not pick a candidate in 34.0% of questions today when asked to choose. No other chatbot topped 1.0%, and GPT-5 was up 1.4pp from yesterday.” (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: Uses jargon 'pp' for percentage points. Minor but should be plain English per the style rules.