Six AIs Saw the Democratic Majority Coming. Voters Added a Verdict on Trump

Six AIs agreed that Democrats were ahead. None described what happened next.

A week before the vote, Satcove put the same question about the US midterms to six models and published the answers in our first article. They gave Democrats 79% to 90% for the House and 46% to 65% for the Senate. On 3 November Democrats won both chambers and took seats in Texas, Kansas and South Carolina, where a Democrat had not won a Senate race in decades. The AIs got the direction right and missed the scale.

This article is not a scorecard. It asks what each AI was looking at, what it left out, and what that says about the way forecasts are built. One caution first: we are reading the result backwards. Had Republicans held the Senate, the same answers could be told as sound judgment. An autopsy shows what happened. It does not prove what the AIs should have seen.

Results below are election-night projections, not certified counts.

What happened

Networks and the AP projected that Democrats won control of both chambers. In the House, the AP said early on 4 November that Democrats passed the net gain of three seats they needed. In the Senate, CBS, CNN and the AP projected a Democratic majority.

The Senate flips, state by state, according to CBS projections unless noted:

• Kansas: Adam Hamilton beat Sen. Roger Marshall. CBS says Kansas had not elected a Democrat to the Senate in almost a century, since 1932. Hamilton is a pastor, not a career politician (see the angles below).
• Texas: James Talarico beat Attorney General Ken Paxton, 55.0% to 43.1% in NBC's table with about 91% of the vote counted. CBS says Texas had not elected a Democrat to the Senate since 1988. Both men came out of primaries: Talarico beat Jasmine Crockett in March, and Paxton beat the incumbent, Sen. John Cornyn, in a May runoff. Before the vote, polls had the race close and the Cook Political Report rated it a toss-up.
• Ohio: Sherrod Brown beat Sen. Jon Husted. Brown lost narrowly to Bernie Moreno in 2024 and ran again this year. Before the vote, a PollCheck average had Brown ahead by about three points, 48% to 45%, with outside groups pledging tens of millions of dollars on each side.
• North Carolina: Roy Cooper, a former governor, beat Republican Michael Whatley. CBS says the state had not elected a Democratic senator since 2008. The seat was open because Sen. Thom Tillis chose not to run again, and he congratulated Cooper on election night. A PollCheck average had Cooper about ten points ahead in late October.
• Iowa: Josh Turek, a state representative and Paralympic wheelchair basketball champion, beat Rep. Ashley Hinson in a seat left open when Sen. Joni Ernst retired. CBS says Iowa had not elected a Democrat to the Senate since 2008 and that Trump won the state by 13 points in 2024. Cook rated Hinson a slight favourite before the vote.
• Maine: Sen. Susan Collins lost and conceded. Troy Jackson, a former state Senate president and logger from northern Maine, was the Democratic nominee. He replaced Graham Platner, who won the June primary and withdrew in July after an allegation of sexual assault that Platner denies. Democrats picked Jackson at a convention on 25 July, under a state process they had never used. Kamala Harris won Maine by almost seven points in 2024.
• South Carolina: Annie Andrews beat Sen. Darline Graham, 50.6% to 47.2% with about 93% counted, according to NPR. Graham was appointed after her brother Lindsey Graham died on 11 July. Trump endorsed her and rallied for her three times in the closing weeks. The last Democrat to win a Senate race in the state was Fritz Hollings in 1998, an incumbent of long standing.

Two of the seven seats were open (North Carolina and Iowa), one was held by an appointee (South Carolina), and in Maine the Democratic nominee changed in July. Candidates who change late are the hardest to price, which matters when reading the AIs' answers.

What each AI said on 26 October

Where the AIs agreed

All six put Democrats well ahead for the House, within 79% to 90%, and every House range started at 70% or higher. On the Senate, none of the six put Democrats below 46%, and the midpoint was about 60%.

The agreement is the first half of the story. That consensus mattered because the Senate required a net gain of four seats against a 53-47 majority. An AI could have called that unlikely. None did. The agreement was also thinner than it looks: Decision Desk HQ appeared in five of the six answers. The AIs were mostly relaying the same published forecasts.

Where they diverged

Agreeing on a direction hid six different ways of getting there. The second half of the story is in how they differed.

What the result says about each AI

Those differences were invisible on 26 October, because each answer looked reasonable on its own. The votes make them readable. Here is what the result says about each model, and then what it says about all six together.

Scores are the mean of the House and Senate Brier scores, where a range was given we used its midpoint, and the outcome counts as 1 for both chambers. A coin flip scores 0.250. The median of the six answers, 81% and 60%, scores 0.098 and beats three of the six models.

The angles the result exposed

The thread through the six angles below is one idea: voters were not only choosing a party, they were judging Trump. We did not ask the AIs to explain a sweep, so we cannot say what each would have said about it. What we can compare is what the result shows next to what they described.

Texas and Kansas: a wave, not an adjustment. The AIs priced a Democratic majority. Texas at about 55% to 43% and a Kansas flip belong to a different class of result than a narrow Senate gain. The 26 October article names Texas as a race to watch, not Kansas. A forecast that treats Kansas as safe has a blind spot at the edge of its distribution.

A rebuke of Trump, not only of the Republican Party. NBC's exit poll found 38% of voters approving of Trump. 60% said their House vote was to express support or opposition to him, and 65% said his policies had worsened economic conditions. NBC News headlined a rebuke. The AIs we asked reasoned about parties and seats. We did not ask them to separate the two.

Kansas: a moral candidate, not a polling category. Adam Hamilton leads the largest United Methodist congregation in the United States and ran as a centrist on unity. Religion News Service reported he prayed for Trump and for his opponent in the final days. Republican-aligned groups called him a "woke megachurch pastor", and Trump and Vice President JD Vance went to Kansas to campaign for Marshall, which shows the race was seen as close. Polling-based forecasts have no variable for a clergy candidate who runs on character. None of the answer summaries we kept mentions the idea.

South Carolina: Trump's own pick. Trump won the state by nearly 18 points in 2024, according to NPR. This time he endorsed Darline Graham and rallied for her three times in the closing weeks. She lost, and she ran roughly 10 points behind the Republican candidate for governor on the same ballot, according to Fortune as relayed in the reports we found. Ticket-splitting is one explanation and not proven. The contrast is the point: the same voters backed a Republican for governor and rejected the candidate Trump chose. It is the closest thing in this election to a test of his personal pull.

Economy and inflation: a shared cause. NBC reports that just under 8 in 10 voters said inflation is causing their family hardship, and 62% said higher prices are more because of Trump's policies. Cost of living is the cause that every analysis in this article can source. Mistral's geopolitical angle sits beside it: coverage of the same polls lists the Iran war among voters' worries, so the assumption was not baseless. We did not measure whether it moved votes.

Trump's personal enrichment: a symptom, not a proven cause. A Reuters/Ipsos poll in August found 63% of Americans saying Trump and his family had inappropriately profited from cryptocurrency, as reported by Axios. A poll commissioned by an outlet covering crypto found it ranked as a midterm issue for 1% of voters, far behind cost of living at 36%. We read enrichment as one sign of a wider distrust, and we have no source showing it moved votes on its own.

What the AIs would say now

This section is an editorial reconstruction by Satcove. These are not statements by the models. We wrote one sentence per model from the record of its 26 October answer.

• ChatGPT: "I had the direction right, but my 50% for the Senate was a synthesis with no model behind it, and the election punished that."
• Gemini: "I gave a wide Senate range of 46 to 56% and no assumptions, so nobody can tell what I was betting on."
• Perplexity: "Every source I cited could be checked, and every one of them was cautious on the Senate."
• Mistral: "I named the Iran war while others stayed with the polls, and I still did not see how far the swing would go."
• Grok: "I said about 65% for the Senate while the Poliwave page I cited said 57%, so my best score rests on a number I cannot trace."
• Claude: "I named a polling error in Democrats' favour as a risk, but I answered after reading the other five, and Kansas was not on my list."

What we take from it

1. Direction was consensual, amplitude was not. All six agreed that Democrats were ahead. None described a result that reached Texas, Kansas and South Carolina.
2. Confidence was rewarded this time, not reliably. Grok, the most confident, scored best, and ChatGPT, closest to a coin flip, scored worst. A closer election would have punished the same confidence.
3. Quality of reasoning and accuracy of the score are different measures. The best score came with a figure that did not match its source, and the best-sourced answer finished fourth.
4. The comparison of angles is the useful part. Each AI left a different blind spot: no stated reasons, no commitment, anchoring, a source mismatch. Together they mark where a single forecast is thin.
5. The rejection of Trump himself is the angle the answers lacked. The exit poll puts 60% of voters as casting their House vote to express support or opposition to him. The summaries we kept reason about parties and seats, and none separates the two.

The deeper question

The AIs are not the main problem. They mostly relayed forecasts, and those forecasts did not describe a wave either. Decision Desk HQ sat behind five of the six answers, and the Poliwave model that two of them cited says on its own page that it has not been calibrated against past elections. Six assistants drawing on one pool of sources behave like one forecaster with six voices.

So the useful question is not which AI was best. It is what the forecasts themselves cannot see: a clergy candidate running on character, a national mood aimed at one person, a Republican incumbent running ten points behind her own ticket. A model fed by polls can add confidence to what polls already say. It cannot add the thing polls do not measure. We do not know why the published models missed the scale, and this article does not claim to. It claims only that an AI placed on top of them inherits the gap.

Our 2 November audit points the same way from a different side. When the same six AIs read one fixed file of 152 documents on Trump's record, none gave an overall verdict: they graded the deficit lowest and declined to grade most of the rest (see that article). Voters did give a verdict the next day. The AIs were cautious where the evidence was thin and still underestimated the size of the vote, which suggests caution about the record and calibration about the result are separate skills. Two articles and one election are not proof of that.

That is a reason to ask several AIs and read where they split, as we did here, and a reason not to count their agreement as independent evidence.

Limits

• Results were projections on 4 November. We did not verify certified totals or seat counts.
• One attempt per model, different tools and modes, no fixed temperature.
• Midpoints of ranges stand in for point forecasts.
• Claude is a Satcove research tool and also a scored model, and it read the other answers before replying.
• Observations about what each AI missed come from the answers we recorded on 26 October. We did not re-ask the AIs after the result.
• Exit-poll figures are early and may be revised. Causes of the result are not established by this article.

What comes next

The open question is whether an AI can learn to doubt the consensus it relays. If we repeat the exercise, we will fix the quality criteria before the vote, keep a running score per model, and ask each AI to name the one thing its sources cannot measure. To put the same question to several AIs yourself and see where they split, try Satcove Open.