AIPolitics 19: Comments on Arnold Kling's Comments

Where Kling is right, where he's handing me back my own Part 1, and the half-step in between

Arnold Kling wrote a post this morning called “Comments on Kyle Saunders.” When someone who’s been doing political economy since before I finished grad school puts your name in his headline and spends a thousand careful words on your framework, you really don’t get to be cool about it. Instead you read it four times, then you write back.

And, well, it’s Sunday morning, and I have some time while following my standard routine of coffee(s), reading and having the political shows on in the background. So here’s the back-and-forth, because I think the disagreement turns out to be more useful than the engagement (and flattery).

Kling read Part 18, where I argued that the cosmopolitan/populist cleavage that’s been rotating into place since 1960 is now getting cut at a right angle by a pro-AI/anti-AI split. His reply, boiled down: a Lee Corso-ish “not so fast, my friend.”

AI is salient to basically one group, the college-educated core of the Democratic Party. Everywhere else it’s background noise. And even inside that core, anti-AI doesn’t become its own thing. It gets folded into what he calls the “Omnicause,” the bundle on the activist left that already carries anti-Trump, anti-ICE, anti-Israel. The pro-AI side is too weak to pry it back out. So AI won’t be a new cleavage at the mass level. It’ll be one more plank stapled onto a coalition that was already built.

I want to do two things with that. Agree with it, because I do. And then show why agreeing with it doesn’t cost me Part 18. :)

Well, first, I already said the absorption part

Here’s the slightly awkward thing about Kling’s objection: I also said it back in the beginning of this series and then it kinda faded.

Go back to Part 1, the long January brain-dump that kicked off this whole series. The core claim, stated near the top and then just semi-ass-u-me-d everyone would remember: AI “will become politicized less because of new ideologies and more because it will activate existing cleavages.” A few thousand words later: “AI will probably not invent new ideological divides. It will activate familiar ones.” And the line I’d point to first: “AI will sort into partisan politics the same way trade and immigration did.”

That’s the absorption thesis, right there. Trade and immigration didn’t conjure new American coalitions out of nothing. They got routed through the ones we already had.

Then Part 3, from late February, where I got specific about the mass public and put down a marker I’d half- mostly-forgotten being this blunt about: “the mass public is not yet clearly sorted into hardened ‘pro-AI’ and ‘anti-AI’ camps the way it is on, say, climate or abortion. Instead, AI is being filtered through existing partisan frames.” And on whether AI would ever become a sorting mechanism at the level of actual voters rather than donors: “That is when polarization hardens, but I think that’s further off.”

“Further off.” That was February. So when Kling says the mass-level cleavage isn’t here, my honest answer is that I agree. And I’d love to leave it there, receipt in hand. But saying a thing in February and actually building it into the argument are two different jobs, and I only did the first. That’s the part Kling caught, and it’s the part worth the post—but you’ll have to decide that, dear reader.

Which raises the obvious question. If I’ve been saying since January that AI gets absorbed rather than inventing anything, what on earth was Part 18 doing with a headline about AI bisecting the cleavage?

Two levels, and the headline only described one

This is where I owe Kling for the call-out, because his objection made me notice I’d let a headline do the work of a part of the argument that had kinda fallen to the side of the road.

Part 18’s title says AI “has bisected” the cleavage. The body of Part 18 says something a lot narrower. The exact line, about the four-quadrant map I drew, is that it “describes elite factional sorting, not voter behavior at scale.” And a little later: “The AI fracture in 2026 is the perpendicular line at the elite-policy level. The mass-electoral implications are slower and less clear.”

So the piece was already running on two tracks. I just didn’t label them well, and a punchy title will paper over that distinction every single time.

Let’s do a better job labeling them now, for goodness’ sake.

There’s an elite level. Donors, staffers, the policy class, the people who write sixty-five-page strategy memos and stand up super PACs. At that level AI really is a new fracture, and it really does cut across the old cosmopolitan/populist line. Reid Hoffman and Bernie Sanders sit on the same side of the 2024 partisan cleavage and they are nowhere near each other on AI. And before someone says this is just the old growth-Democrat versus labor-Democrat fight wearing an AI jersey: it isn’t, quite. That fight always kept its lines inside the party. What AI does is reach across party, dropping a chunk of the labor left and a chunk of the populist right into the same anti-AI foxhole while splitting the donor class along a seam the cosmopolitan/populist line never drew. The cross-party rhyme is the genuinely new thing. The four-quadrant map in Part 18 is a map of that room.

And there’s a mass level. Tens of millions of voters for whom AI lands, per Pew’s own issue ranking, well down the list behind inflation, immigration, health care, abortion, crime, so far anyway. At that level there’s no new axis. There’s a bundle, and anti-AI sentiment slides into it next to everything else the coalition already hauls around. That’s Kling’s “Omnicause.” That’s my Part 1.

The two pictures look like they fight. They really don’t though, because they’re photographs of different rooms. The mistake was mine, for printing the elite-room photo with a caption wide enough to sound like the whole country.

What Kling actually adds

If all Kling had done was catch me on a caption, that’d be a footnote, not a post. He did more. He swapped out the engine under the whole realignment story, and the swap is worth me showing you my own thinking here.

My account of how the cosmopolitan/populist cleavage got built runs on ideological sorting, the bottom-up version Alan Abramowitz and I were arguing about back in the late ‘90s: voters with consistent issue positions found their way into the party that finally matched them, once the elites got clear enough to make the choice legible. Kling tells it through social identity instead. In 1960, he says, people weren’t thinking labor versus capital. They were thinking Irish, Italian, Jewish, WASP, Southern white, Black-and-not-yet-allowed-to-vote. The coalitions were ethnic and regional long before they were ideological, and the ideology arrived later, if it arrived at all. His blunt version: most people don’t have a political identity. They have a social identity that drags a political one along behind it.

Now, I have to cop to a slightly unfair advantage on this one: Alan Abramowitz and I wrote the paper. Literally. “Exploring the Bases of Partisanship in the American Electorate: Social Identity vs. Ideology,” Political Research Quarterly, 2006. The title is Kling’s (and his commenter’s) question, word for word. And our answer, running the National Election Studies from 1952 to 2004, came down on the ideology side: party identification tracked voters’ ideological preferences a good deal more tightly than their social-group memberships, and the gap had been widening for decades. So Kling and I aren’t strangers across this table. We’re sitting at opposite ends of a fight I helped pick twenty years ago.

Which is exactly why I won’t wave him off. A lot of things have happened since 2006! But let’s name two. A wave of work since has pushed back toward identity, recasting partisanship as a mega-identity, a stack of social memberships that line up, with ideology as the story we tell afterward, leaning on the relative ideological naivete of the American public. And that’s where I’d push back: the “active partisans” work we did suggests ideological engagement runs on a continuum, not a clean elite/non-elite split. But that’s an even longer story for another day.

That said, the electorate kept sorting, which makes ideology and identity harder to pry apart than they were when AA and I were fitting those models, because by 2026 your ideology is one of your social memberships. My 2006 self would defend the ideology side of the ledger. My 2026 self thinks the two co-move so tightly, at the level of a single voter, that you often can’t say which one is pulling. Scott Gibb, in Kling’s own comments, said it better than I did: “How can we really know whether people are voting on social identity versus political ideology? We probably can’t.” Right. We mostly can’t. So I’ll treat Kling’s micro-foundation and mine as two readings of one underdetermined thing, and note that the clock axis I keep drawing (educated/credentialed-cosmopolitan on one end, rooted-populist on the other) holds up whichever reading you like, because that line is a social identity and an ideological one at the same time.

AK rebuilt my Part 5 without using it

Here’s the part that made me chuckle-blush a little.

Kling, working from his own priors, goes looking for why the college-educated core hates AI specifically, and lands here: “What is the value proposition for a college credential and for those who provide such a credential, given what AI can do?” The degree was a promise. AI makes the promise wobble. And so, he writes, “many of the college-educated respond by attacking AI any way that they can.”

That’s Part 5. That’s the entire argument of “The Credential Is the Coalition,” reached from a different door and parked in two sentences. I spent four thousand words getting there. He treats it as obvious, and I’ll take it. (When someone reconstructs your mechanism without leaning on your work, the mechanism is probably real, right? Right?)

And the convergence earns its keep, I think, which is why I’m not just spiking the football. The credential is the one wire that runs from the elite room down to the mass floor. Most elite fractures stay elite. This one doesn’t, because the thing under threat isn’t a policy preference, it’s an identity, the degree that sorted you into the coalition in the first place. That’s why the Democratic numbers on AI (roughly 20 percent favorable, 56 unfavorable this spring, and the gap’s only widened since) read so differently from the Republican ones (a still-even 33/33). The Republican fight over AI is an interest fight: who pays for the data center, what it does to my power bill. Negotiable. The Democratic one is closer to a wound. You can cut a deal on an interest. You can’t really cut a deal on “this machine makes the central credential of my adult life worth less.”

Now the complication, and it’s one I raised back in Part 4 and then (also) conveniently forgot here. The Democratic anti-AI number isn’t coming from the young credentialed class. It’s older, whiter Democrats driving it, while the younger, more-credentialed, more-online Democrats, the ones whose degree is arguably most exposed, are actually a touch more favorable to AI. That cuts against a clean credential-threat story, or at least roughs it up. My best read is that the wound is sharpest where credential, age, and institutional loyalty stack on top of each other, and the young credentialed are exactly the group a trigger would have to convert before any of this goes fully mass.

And I won’t pretend the credential does all the work. Some of that 20/56 is plain negative partisanship: when the visible owners of AI are Musk, Thiel, and a Sacks-run policy shop, you don’t need a theory of credentials to explain why Democrats recoil. The credential is the part that makes it sticky. The out-group cue is the part that makes it fast.

One of Kling’s readers, Roger Sweeny, put it in a sentence: “Since the heart of the Democratic Party is in academia, and ai threatens to take away some of the reason for schooling, of course it will be booed.” Of course it will. That’s the credential doing its work, straight from a comment box. And along with the credential comes an identity, which comes with motivated reasoning that’s only strengthened among sophisticates.

The asymmetry is real, with one caveat

Kling’s sharpest move is about asymmetry, and it’s one I’d been circling in Part 18’s subtitle without naming it so plainly: the passion is one-sided. Or rather, one kind of passion is. There’s a hard, identity-motivated, anti-AI bloc inside the Democratic coalition, and nothing symmetric to it on the right. No Republican faction experiences AI as an attack on who they are. Most Republicans, as Kling says, would sit through an AI pitch at their 2028 convention without much cheering or booing either way. That identity asymmetry is real.

Benjamin Gilad, another commenter, supplied the mechanism: anti forces feel more strongly than pro forces, just about everywhere. Anger turns out. Mild approval stays home. Sweeny’s “Burn it down” energy is the affective version of the same observation. The anti-AI Democratic pole has mass reach not because more people hold the position but because the people who hold it hold it like a grievance, and grievances vote.

So Kling’s right, and I’d only redraw the map a touch. The Republican coalition isn’t pro-AI in any unified way. It’s three different fractures wearing one coat: the Thiel-style accelerationists, the national-security hawks, and a populist right that’s plenty angry about electricity bills and small-town data centers. What it lacks is a mass identity wound to convert those elite fractures into heat. And the fractures themselves are real, not cosmetic. Look at the Vatican dust-up from a few days ago, where Peter Thiel is on record fretting that JD Vance, whose campaigns Thiel funded, has gotten “too close to the pope” on exactly this cluster of questions. That’s not a party at rest. That’s a party whose internal AI argument hasn’t found its trigger yet.

(The caveat I owe you, of course: that last claim leans harder on Thiel and Vance than on polling. The mass numbers on a Republican anti-AI constituency are thin. I’m reading elite tea leaves there, and I’d rather flag it than dress it up.)

But the identity framing hides something, and a commenter in Kling’s thread flagged it: there’s a second kind of mass passion, and on that one the asymmetry runs the other way. AI is an elite-driven phenomenon. Elites build it, elites pocket most of the upside, and in that it rhymes with social media a decade back. But elite-driven has never meant populism-proof. Social media was elite-built and elite-enriching too, and it still curdled into a populist grievance, one that, by the time it matured, came out right-coded: Big Tech, censorship, the algorithm against the little guy. AI’s distributive grievance is rhyming the same way.

The on-ramp here isn’t the diploma, it’s the meter: who pays for the data center, whose water it drinks, whose electricity bill it doubles. And that backlash has already happened, more red and rural than blue and credentialed. Festus, Missouri threw out half its city council over a data center. Rural Virginia and Georgia are fighting siting right now. I spent Part 3 on exactly that “you pay, they profit” frame, and it’s bipartisan by construction, but the people walking through that door first have been on the right. So the map has two poles, not one. Democrats own the identity wound, stickier and slower. The right owns the distributive one, already live. Different doors. Different timelines. Same populism.

What would actually flip it

And then Kling hands me the most useful thing in the post: a falsifiable prediction. He thinks that by 2028, “anyone who dares to make a pro-AI statement at the 2028 Democratic convention will be booed out of the hall.” Yep.

That’s a claim about the mass level hardening, the exact event Part 3 filed under “further off.” And it gives me something better to do than nod, because I think I know what would have to happen first.

Back in Part 15 I spent a while with Andy Hall’s “Politics of Jobless Prosperity,” and the most useful thing in it was a trigger condition: the real populist backlash to AI shows up when unemployment climbs something like two points with a clear story that AI is the reason. Not exposure. Not anxiety. A number moving, plus a villain.

Two honest caveats on that trigger. First, “a clear story that AI is the reason” is doing heavy lifting. In an actual downturn, unemployment rises with five villains in the room at once, rates and tariffs and whatever shock started it, and AI is one strand of the argument, not the certified cause. Whether the story comes out “clear” is itself a political fight. Second, it’s a labor trigger, and it’s not the only gate in the building. The distributive gate, the meter one, is already ajar, which is why the mass heat that’s actually shown up so far has been red and rural rather than blue and credentialed.

So there are two ways this goes mass. Hall’s labor trigger floods the elite Democratic fracture down the credential wire and Kling gets his booing. Or the distributive backlash keeps widening on the right and the bisection arrives from the other side first. Both are live. Neither has fully fired…yet.

Let’s put both predictions on the same shelf, then, where the rest of us can check them. Kling: pro-AI gets booed off the 2028 convention stage. Hall: the gate is a two-point unemployment move with AI’s name attached. My addition: watch the Vorland faction, the lonely pro-AI Democrats I mapped in Part 18. If that crowd gains ground inside the party between now and the platform fight, the absorption held and Kling’s a shade too pessimistic. If they get run out (from today’s perspective, the likely outcome, but we’ll see), the gate opened. That’s a testable thing, not a vibe, and I’d rather be wrong on the record than vague off it (lol).

Where that leaves us

Kling is the second person to engage this series at length as an interlocutor rather than a fan or a hater, and that’s welcome and awesome. Andy Hall was the first. Kling’s the more useful kind, because he made me go back, and you know, actually think and read through what I’d already said/written. And, well, I’d been derp-running a two-level argument under a one-level headline since at least Part 3.

So, the concessions, plainly. At the mass level, today, he’s right: AI is getting absorbed into the coalition we already have, not splitting it. I said as much in February, and the data hasn’t moved since. The “Omnicause” swallows the plank.

At the elite level the fracture is real and it’s new, and there are two wires running down into mass politics: the credential, for the Democratic identity pole, and the meter, for the distributive one already humming on the right. The bisection goes wide if either gate opens, Hall’s two-point labor number on one side or the data-center revolt metastasizing on the other. And here’s the falsifiable version, because a two-level model that can absorb any outcome isn’t worth much: if one of those gates opens and the mass electorate still doesn’t sort on AI within a cycle or two, then there was never a bisection to begin with, just a donor-class spat I oversold with a good headline.

I honestly don’t know which way it breaks. That’s not a dodge, it’s the actual state of the evidence, and anybody who tells you they’re certain is selling something. What I’d say is that the wire’s in place and the room’s wired. We’re waiting to see whether anyone throws the switch.

Anyway. Thanks, Arnold. Someone of your gravity can put me in their headline anytime. Cheers.