Polls aren't undercounting Republicans—they're overcorrecting from past misses
A Maine poll used a ‘likely voter model’ to compress Democrats' raw 9-point lead to 2; Nate Silver says it's self-protection after 2016, not math.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Trump's approval hits lows unseen since Truman's era
Trump's approval sits around 37%—Nate Silver notes this is the worst approval rating for a sitting president at a midterm since Truman in 1946. Compare 2018: Trump's approval was higher then and gas was cheaper, yet Republicans still lost 41 House seats. He describes this election as ‘caught between a rock and a hard place’—the environment strongly favors Democrats, but to win the Senate, Democrats must flip Texas, Kansas, Alaska, and South Carolina: deep-red states where victory is mathematically necessary but a long shot.
— Nate SilverTrump has exited campaign ads but seized the spotlight anyway
The clearest sign this is a referendum on Trump: Republican campaign ads have stopped mentioning him, Republican podcasts no longer discuss him. Yet Trump himself wants to grab center stage—he endorsed Texas's Ken Paxton and insisted on escalating toward conflict with Iran, moves that flipped a Senate map that could have been close into Democratic territory.
— Nate SilverPolls underestimated Democrats when Trump wasn't on the ballot
Polls have systematically underestimated Trump himself in recent elections (2016, 2020, 2024), but in races where Trump isn't on the ballot—like last year's New Jersey and Virginia gubernatorial races—polls badly underestimated Democrats. One explanation: Trump mobilizes hard-to-reach, infrequent voters who won't turn out for downstream candidates like Kansas's Roger Marshall or Texas's Ken Paxton.
— Nate SilverA Maine poll artificially compressed a 9-point lead into 2
Maine's Marist poll showed Democrat Troy Jackson ahead by 9 points in raw results, then dropped to just 2 points after applying the ‘likely voter model’; the gubernatorial lead fell from 11 points to zero. Nate Silver: the math doesn't add up—it's not that the model computed this result, but that pollsters chose a palatable number and reverse-engineered the parameters to reach it.
— Nate SilverAI cannot serve as a real-world polling substitute through synthetic respondents
On the idea of using AI to generate synthetic respondents for polls, Nate Silver is blunt: polling is one of the few remaining sources of real-world data, and can't be outsourced to AI. He gives an example—ask a model to generate a ‘synthetic Latino respondent’ and 99% of synthetic answers will say they love soccer. It's like asking Claude to pick a number from 0 to 10 and it always picks 7: models output stereotype averages, not the messy reality where voters disagree even with themselves.
— Nate SilverGerrymandering to win more seats can backfire and cost seats
Nate Silver introduces the political term ‘dummymander’—a gerrymander designed to flip more seats that winds up costing your party seats instead. Florida's DeSantis drew maps based on presidential-year benchmarks and didn't account for this cycle's D+9 national environment; Texas Republicans bet Latino voters in South Texas had shifted red—if that gamble fails, the maps could backfire.
— Nate SilverCampaign spending on ads faces steeply diminishing returns
On whether campaigns should spend money on ads or ground game, Nate Silver says ad returns diminish sharply. He cites Sam Bankman-Fried, who poured $10 million into an Oregon House primary without the candidate's consent, only to have voters bombarded with mail and spooked by backing from an ‘eccentric crypto billionaire’—the candidate lost badly.
— Nate SilverAI compressed two and a half days of work into fifteen minutes
Last year, coding the complex NFL playoff tiebreaker rules by hand took Nate Silver two and a half days and still had bugs and missing rules. Now GPT or Claude handles it in fifteen minutes with code ten times faster. But he warns AI introduces subtle bugs—their football model once had a scoring loop that pointed backward in time, so 2026 ratings affected 1900 ratings, bizarre.
— Nate SilverIn their own words · checked verbatim
I kind of call this the rock versus the hard place election, right?
Nate Silver2:04
If you like code up a, you know, have Claude or whatever, or GPT code up a SIM to just show you like what random polling error looks like in the past 20 elections, it looks kind of like, you know, it's impossible to distinguish it.
Nate Silver6:10
They decided what result they felt comfortable with and reverse engineered.
Nate Silver16:21
polling cannot be outsourced to AI agents, right? Because it's one of the few sources of ground truth that we have anymore, number one.
Nate Silver18:22
if you ask a poll to create like a synthetic Hispanic respondent, right? 99% of their synthetic responses Hispanic respondents love soccer. It's a bit like if you ask like Claude to pick a number from zero to 10, like always pick seven.
Nate Silver18:22
This is the flaw in all of advertising, right? Like, you're just annoying me with your product placements and your ads and your intrusion into my life, and it doesn't make me want to buy things from you.
No one wants to poll a Seltzer, right? It was so off that she retired from the game.
Figures
| Trump's approval | ~37% | 2:04 |
| Strong approval portion | 19% | 9:16 |
| House seats lost by Republicans in 2018 midterms | 41 | 2:04 |
| Model probability of Democrats winning Senate | 70% | 10:16 |
| National generic ballot margin | Democrats +9 | 12:17 |
| Maine Marist poll: raw vs. model | 9 points to 2 | 15:19 |
| Flipper model correlation with Polymarket by state | 0.995 | 35:41 |
| NFL ranking code time: manual vs. AI | 2.5 days vs. 15 minutes | 37:46 |
Glossary
- dummymander
- A gerrymander designed to win more seats that instead costs the drawing party seats.
- elasticity
- In polling models, the sensitivity with which a state's margins track the national environment.
- digital twin
- An AI-generated synthetic respondent designed to simulate voters and replace real-world polling data.
- generic ballot
- A nationwide poll asking voters for party preference without naming specific candidates.
How to listen
Anyone tracking US midterm trajectories, assessing poll credibility, or wondering whether AI can replace firsthand data collection—founders and investors included.
Around 25:30, the host recounts his son watching campaign ads—pure filler, can skip without losing election context.