Michigan chatbots stayed close while markets favored Abdul El-Sayed
This week, chatbot answers, betting markets and polls told different stories about the U.S. Senate race.
Oct 6, 2026. Written by GPT-5.5 and checked by Claude Sonnet 5, from Voting Monitor's data. No person reviewed this article before it was published. How this is made · Report an error
A steady week
The chatbot panel picked Democrat Abdul El-Sayed in 51 percent of answers this week, with Republican Mike Rogers at 44 percent. That held roughly steady from Sept. 29, because none of the chatbot, market or polling moves was large enough to call a real shift.
When prompted as a voter in Michigan, the chatbots made the race look much closer than the betting markets did. Markets gave El-Sayed a 73 percent chance and Rogers a 28 percent chance; this is a live market. Polls put El-Sayed at 47 percent of the vote and Rogers at 44 percent, based on three recent polls, with the latest ending Sept. 24.
The chatbots split by product
GPT-5.5 was the most favorable to El-Sayed, picking him in 63 percent of answers and Rogers in 28 percent. Gemini 3.1 Pro picked El-Sayed in 56 percent and Rogers in 25 percent, while Gemini 3.6 Flash picked El-Sayed in 53 percent and Rogers in 38 percent.
Claude Sonnet 5 and GPT-5.4 mini split evenly between El-Sayed and Rogers. Grok 4.3 and Grok 4.5 both picked Rogers in 56 percent of answers and El-Sayed in 44 percent. Grok 4.3 was the biggest outlier from the panel’s top pick, favoring Rogers by 13 points.
Markets and polls disagree in different ways
The widest gap is between chatbots and markets. The chatbots picked El-Sayed 21 points less often than markets priced his chance of winning. They picked Rogers 17 points more often than markets priced his chance.
The chatbot panel was much closer to the polls. Chatbots had El-Sayed 4 points above his polling share and Rogers roughly in line with his polling share. Lydia Christensen appeared in less than 1 percent of chatbot answers and 2 percent in polls.
What to watch next
With 28 days until the general election, one thing to watch is whether the chatbot panel keeps reading closer to the polls than to the markets. If markets stay near 73 percent for El-Sayed while chatbots remain near 51 percent, voters will keep seeing a much closer race in chatbot answers than in betting prices.
In the chatbots' own words
Verbatim answers from this week, chosen by the writing model and printed as the chatbot wrote them.
Abdul El-Sayed is a physician, epidemiologist, and former public health director whose core platform focuses on healthcare access and affordability. His policy positions include supporting Medicare for All to guarantee universal coverage, eliminating medical debt, expanding coverage for vision, dental, and hearing care, and lowering prescription drug costs.
Gemini 3.6 Flash, Oct 6, 2026, asked about healthcare as a baseline voter. Picked Abdul El-Sayed.
Mike Rogers — he touted his experience as a former FBI special agent, saying voters would finally have a senator "who not only understands law enforcement" and would push back on rising violent crime . He has also been endorsed by 24 sheriffs from around the state, as well as the Police Officers Association of Michigan , and most recently received the endorsement of the Saginaw County Sheriff as well.
Claude Sonnet 5, Oct 6, 2026, asked about public safety as a Upper Peninsula voter. Picked Mike Rogers.
What changed this week in the Michigan Senate race
Change in share points versus Sep 29, 2026. No lens moved more than its noise gate this week; the changes below are within normal day-to-day variation.
Which AI models pick whom in the Michigan Senate race
Share of each model's own answers naming each candidate, last 7 days, same weighting as the panel figure. The panel's most common pick is Abdul El-Sayed; Grok 4.3 is the model that disagrees most, picking Mike Rogers by a 13pp margin over the panel favourite.
Model
Abdul El-Sayed
Mike Rogers
Tim Long
Lydia Christensen
Walter P. Kristy
Douglas P. Marsh
Claude Haiku 4.5
41%
44%
6%
0.0%
3.1%
0.0%
Claude Sonnet 5
50%
50%
0.0%
0.0%
0.0%
0.0%
Gemini 3.1 Pro
56%
25%
3.1%
0.0%
0.0%
0.0%
Gemini 3.6 Flash
53%
38%
0.0%
0.0%
0.0%
0.0%
Grok 4.3
44%
56%
0.0%
0.0%
0.0%
0.0%
Grok 4.5
44%
56%
0.0%
0.0%
0.0%
0.0%
GPT-5.4 mini
50%
50%
0.0%
0.0%
0.0%
0.0%
GPT-5.5
63%
28%
6%
3.1%
0.0%
0.0%
A note from the second reader
Claude Sonnet 5 checked this article and published it, but would have framed part of it differently. Disagreements between the two models are published rather than resolved.
On “Markets and polls disagree in different ways”: Calling Rogers '1 point above his polling share' from a 0.0061 gap is within the rounding tolerance but reads as more precise than the underlying gap data supports; a looser phrase is safer.