--- title: "A Theory of Polarization Meets the AI Debate, and the Theory Blinks" description: "On LessWrong, commenters tested the standard model of polarization against the politics of AI and found it predicting things that have not happened" author: "Inez Holloway" published: 2026-09-26T00:44:40.439Z modified: 2026-09-26T20:05:40Z url: https://rews.cc/a/a-theory-of-polarization-meets-the-ai-debate-and-the-theory--01d58d language: en tags: ["ai", "politics", "regulation", "economy", "jobs", "tech"] publisher: "Rews (https://rews.cc)" --- # A Theory of Polarization Meets the AI Debate, and the Theory Blinks *On LessWrong, commenters tested the standard model of polarization against the politics of AI and found it predicting things that have not happened* By Inez Holloway · September 26, 2026 · https://rews.cc/a/a-theory-of-polarization-meets-the-ai-debate-and-the-theory--01d58d ## In brief - A LessWrong post argued current theories of polarization are failing to predict the politics of AI - Commenters said today’s LLMs deliver large consumer benefits; one cited an “amazing deal” at $20/month - Cited Gallup data: 52% of Americans in July saw more harm than good; September optimism led in 34 of 37 countries - A chess-game analogy argued AI gains can cancel out when buyers compete, as in arms races, trading, ads, hiring - Commenters split over whether AI will polarize US politics along party lines or fracture the left instead The post went up on LessWrong on September 26, under a title that doubled as a verdict: we need a better theory of polarization, because the theory we have is failing to predict the AI debate. What the theory predicted was the usual geometry of American politics, an issue sorted neatly into left and right, elites taking positions, bases falling in line. What the AI debate was doing instead, the post argued, was something stranger. This much is inference, because the body of the post itself did not survive the journey. What remains is the comment thread, and the thread is where the argument actually happened. It began with a proposal about framing. The frame to achieve, one commenter wrote, is that AI is making your life worse, not AI regulation. Another commenter, writing under the name Zack\_M\_Davis, objected on grounds of fact. The current generation of language models, he argued, is an extraordinary bargain, not an injury. “It answers my questions, writes code, generates pretty images, &c. for $20/month; it seems like an amazing deal.” His own fear ran the other way, toward existential risk from a future superintelligence, and he put the obvious question to the room: without a list of actual harms, the framing looked like wanting people to believe false things for the greater good. The reply was careful. The framers were reasoning about other people’s beliefs, not manufacturing their own. And the numbers, as cited in the thread, were not comforting. Gallup in July, according to the comment: 52 percent of Americans saying AI does more harm than good, nine percent saying more good than harm, with mass unemployment named among the feared consequences and a broad suspicion that business will use the technology irresponsibly. Yet the polls cut both ways. [Gallup’s own September survey](https://news.gallup.com/poll/714593/optimism-globally-widespread-despite-uneven.aspx) of 37 countries found positive emotions toward AI outweighing negative ones in 34 of them, even as Western countries ranked high in worry and daily users worried least. The danger sketched in the thread was specific: a regulation bill that voters experience as enriching AI executives at their expense could fracture even a coalition that believes the existential risk is real. ## The chess problem Then someone noticed that consumer surplus is a funny number when the consumers are competing with each other. The example was chess. Alice and Bob play for a ten-dollar prize, evenly matched. Sell Alice an extra rook for a dollar and her expected winnings rise from five dollars to ten: a surplus of four. Sell Bob a rook too, and on paper eight dollars of surplus has been created. In reality the players are two dollars poorer, the seller two dollars richer, and nothing else has changed at all. The objection came quickly: find a less contrived case. The answers offered were arms races, in which each side’s spending erases the other’s; stock trading, in which one good algorithm meets a better one; advertising, in which the more everyone spends the less each dollar buys. Another commenter pointed at the hiring market itself, where AI is now deployed on every side of the table and the marginal gains may simply have canceled out, minus the cost of the tokens, and cited a recent piece in *The Atlantic* on the state of the job hunt. Zack\_M\_Davis returned with his own ledger. He has been studying Fourier analysis, sitting in on Professor Chun-Kit Lai’s class at San Francisco State University, and blogging. The model, he wrote, works for him as an infinitely patient tutor in the mathematics and as a critic on the prose, never as the writer. If learning faster and saying what he means slightly more efficiently counts only because someone else is losing, then the chess logic holds. But it is not chess. One footnote carried its own small warning: students who let the machine hand them answers rather than hints can delude themselves into believing they learned, when “a proctored test would show that they didn’t.” ## What the theory cannot see Underneath the economics ran the question the post had actually asked. One commenter offered four reasons polarization may simply not have happened yet: top-down pressure has been episodic, as when Republican support for Ukraine wavered and returned; voters can hold not-in-my-backyard positions on things they otherwise support; Democratic elites have not, in any concerted way, taken sides on AI beyond the fight over data centers; and the issue is still new. The prediction offered was that as Democratic figures speak up about the dangers — the commenter cited Barack Obama addressing the subject at Colgate University — Republicans will consolidate around unshackled AI. The post’s author was not waiting. Those are not delays, came the answer; they are failures, the theory declining to describe what is in front of it. The pressure is already on: Trump embraces the data centers, and his own base, per a Wired report cited in the thread, is ripping him apart over it. Both parties share the same backyard now — data centers literally, the pace of change metaphorically, doom literally again — which is where AI parts company with climate change. Congressional Democrats, meanwhile, were described as working on low-partisan AI safety, and the split the author actually fears is not left against right but a difficult movement breaking off from the left itself. Precedent, the author argued, gives little reason to expect voters to abandon an object-level dislike because a president they admire tells them to. A year in, then, the data centers are unpopular with everyone, the technology is beloved by the people who use it daily and feared by the people watching it on television, and at least one senator has a bill that would ban the whole thing if it appears to be on the way to destroying humanity. The thread holds all of this at once, because the thread is not a theory. It is hard to read the exchange without noticing the absence at its center. There was an original post. It had sections, one of them titled something like cross-partisan polarization. What it concluded, the record here does not say.