You can replace almost everything in polling with AI—just not the respondents
The liberal baby bust, the UK's angry young women, young men filling the pews but not going to college, nature is healing for Democrats, the fastest growing political YouTube channels, the NYT in data
No. 398 | April 10, 2026
🤖 The false promise of silicon samples
Following polling analysis cycle after cycle is like being stuck in the movie Groundhog Day. If there’s a polling miss, people will swear off horserace polling, but the breathless reporting of head-to-head polls always returns. Pollsters who publish outlier results or don’t weight on education are mocked, yet those pollsters persistently reappear.
This also applies to more novel approaches. In 2024, there was some buzz around a company called Aaru that promised synthetic AI respondents could be both cheaper and more predictive than real polls. The basic premise: instead of polling a real person, poll an LLM fed a diet of the latest news. Their results ended up tracking closely with the polling averages — at the time, a tell that this new AI form of “polling” was basically a copy-paste of normal polling rather than providing any sort of value-add. It ended up being just as off as the polling averages and forecasts which showed the presidential race a tossup.
They’re now back in 2026 with claims of having signed up midterm campaigns. The forecasters and aggregators aren’t letting the silicon samples anywhere near their averages. Silver Bulletin’s Eli McKown-Dawson writes this week that it’s because they’re not real polls. Nate Silver himself didn’t mince words on this during the ‘24 cycle:
I consider myself a big proponent of AI in polling. You can actually use AI to write pretty good questions. You can also use it to analyze a survey pretty well. It used to be replacement-level intern-level stuff, but you can feel it inching up the food chain with every model release. You also have several companies offering AI moderation of focus groups or in-depth interviews, which is really exciting use case if it helps scale an extremely valuable but underutilized part of opinion research.
You can probably replace almost every part of the polling process with AI, maybe even the pollsters themselves.
What you can’t replace is the respondents.
The crux of why is this: polling exists to tell us what we don’t already know about what the public thinks. LLMs are by definition a synthesis of all the things we already know. The likes of Aaru will say that they feed their silicon samples a media diet to update their preexisting beliefs. But this is a classic garbage-in, garbage-out problem. What you’ll likely get is a highly media-filtered, CW-addled version of what we think the public thinks. Which is why their results hewed so closely to conventional media polls in 2024. In the end, the results could not surpass traditional approaches in the same way that AI can replace bad writing with average writing, but doesn’t on its own produce great writing.
It might be possible for an LLM to predict what people feel about, say, the state of the economy. Congratulations: that’s 1% of polling. A ton of private polling is on pretty obscure issues where there isn’t much if any reporting yet. I got texted and took a poll like that this week. Afterwards, I Googled what was being tested—and nothing. Likewise, lots of political polling — the majority by volume — is in downballot races with next to zero media coverage. If one candidate’s numbers start moving because of some targeted direct mail or under the radar Facebook ads, it’s unlikely that will make it into media coverage or into Aaru’s AI models — but it absolutely will show up if you actually ask people.
For a long time, we’ve been promised the sexy new toy that will eventually kill off traditional polling. In 2014 and 2016, I watched as many in the industry got snookered by Cambridge Analytica promising a fancy-sounding form of behavioral targeting. The firm predicted personality traits based on conventional voting and demographic data and then used the personality types to predict political behavior. (Just using the voting and demographic data to build models directly would have been too boring.) The hype cycle here feels a lot like Cambridge Analytica before the fall.
The Silver Bulletin folks do make a concession that silicon samples might be useful for modeling. But this use case is already well covered by existing machine learning techniques, whether Silver’s election forecasting or individual-level voter scores. We already know how to project out the likely views of everyone in a population based on survey data — but they need to be based on samples of real voters.
Where AI might actually be useful to polling is in translating raw data to linguistic outputs—exactly the thing LLMs are designed to do. Let’s say I give an AI model raw scores for an individual voter across different issue areas. Could it spit out an ad or text message copy entirely personalized to them? I could see that. But it all starts with original survey work to actually get voters to tell us the things that will always remain obscured from large language models.
Need an expert for your conference or fly-in to break down the midterms or the latest trends in politics, demographics, and AI? Visit my public speaking page or get in touch directly.
🎙️ Two quick hits this week
Great to join the gang over at Central Air to talk realignment, Iran, and more:
And also at WSJ Opinion Live in DC:
👶 The liberal baby bust
Charlie Smirkley has been diving into different data sources to signal boost earlier work by the Institute for Family Studies about a growing ideological gulf in childbearing. At ages 25-35, conservative women are about twice as likely to have children as liberal women.
It doesn’t take much to project the consequences of something like this well out into the future. If these trends hold, descendants of today’s conservatives will be over 80% of the population by 2100. That is among the native-born, and there is going to continue to be a lot of international migration, but there’s a book that’s partly about why you shouldn’t bet on them being liberal either.
Now, ideology isn’t a heritable trait. Children don’t perfectly mirror their parents’ political views. But a solid majority do.
The split in child-rearing is also a function of the growing cultural, geographic, and psychological cleavage in the country. It used to be much easier to switch parties if the Democrats or Republicans were more in fashion at any given point. Now, party and ideology are increasingly aligned with innermost values and family dynamics. You also see meaningful differences in teens’ mental state based on the party their parents are in. So, parents clearly seem to be having some impact. And to the extent that parents are inculcating in their children a set of values, those might have once been incidental to politics but they’re now central.
As far as overall rates of fertility go, the native-born population is still going down, as young conservatives and liberals both have a fertility rate below 2, the replacement rate.
A piece in The Upshot says not to worry, as fertility is being delayed and women are making up for it with births in their 40s.
But that turns out to a “felony level chart-crime,” according to Dartmouth professor Paul Novosad. In raw numbers, the rise in births for older women is not even a tenth of the drop among young women.
😡 The UK’s angry young women
Going back to a common Intersection theme of everything about modern US politics actually being a global phenomenon, Scarlett Maguire has some groundbreaking new polling of Gen Z in the UK showing the hostility that young women have towards young men — and that young men actually don’t feel in return. Some of the politics here are absolutely wild:
Young women are 26 pts less favourable to capitalism than young men, and feel much more positively towards communism than capitalism.
Women u25 dislike capitalism so much, they view it as (un)favourably as fascism.
UK should pay slavery reparations by a 2-1 margin
And:
U30 women are 3x as likely to hold a negative view of young men than the other way around
Just 35% of u25 women hold a positive view, only 11% a very positive view
This is a bleak story, but not a surprising one given today’s highly gendered information spaces.
In education too, we’ve seen massive segregation by gender, with the choice to finish or even to go to college falling along gender and political party lines. The share of Republicans who say they’re not interested in graduating college has doubled since the turn of the century. And the rate for Republican men is especially high: four times that of Democratic women. So, college becomes even more female and liberal.
This plays out in actual surveys of college graduates, who have moved well to the left in recent years. Some have speculated that this is just because of compositional changes — grads becoming more female or nonwhite. The key chart in this piece from Paul Heideman is a bit hard to read, but it debunks the idea: college grads are getting more liberal because all groups within academia are getting more liberal, including white men. And that returns to the point of the chart above: even as demographics play some role, the people opting out of college are disproportionately Republicans.
⛪ Young male revival?
Talk about most things this week fitting a theme: young women zooming left and going to college, men maybe moving right and not pursuing higher ed. And now, the Gallup finding widely shared this week about young men bucking the trend towards secularization and turning to religion.
🫏 A “nature is healing” moment for Democrats
In-party favorability for Congressional Democrats has overtaken Congressional Republicans, ending (for now) a brief interlude of Republicans doing better on this score. The underlying mechanism here was Democratic base anger that their leaders weren’t fighting Trump hard enough at the outset of his term, reversing the common pattern where this was always more the Republican complaint about their leaders. YouGov finds this has reversed, possibly a sign of base Democrats predictably rallying around the flag for the midterms. (If you can call net negative favorability for your party in Congress that…)
🏛️ The Big Sort of Congressional districts
Tom Wood’s most recent visual shows the realignment at the Congressional district level, where wealthier districts now vote Democratic, and poorer districts vote Republican.
📺 More of the fastest growing YouTube channels lean right
Kyle Tharp has the breakdown of YouTube subscriber growth in Q1, where Turning Point USA added over 1.5 million subscribers in one quarter alone.
📰 2.2 million NYT articles as data
A new freely available dashboard makes it easy to view trends in 2.2 million articles published by The York Times since 2000.
