The Politicians Who Treat Congress as a Launchpad for Media Celebrity Win More Coverage and Accomplish Less

Two point two million public statements. Floor speeches, press releases, tweets, newsletters, every word uttered or published by every member of the 118th US Congress, fed into a large language model and sorted, roughly, into a taxonomy. Policy discussion. Critical debate. Personal attacks. The researchers at Dartmouth College who ran this analysis expected to find the usual ugliness of modern American politics. What they didn’t quite expect was how perfectly rational the ugliness turns out to be.

The study, published this week in PNAS Nexus, introduces a category that feels immediately recognizable: the conflict entrepreneur. Not a legislator who occasionally reaches for a cheap shot during campaign season, but one who has reoriented their entire rhetorical output around personal insults: attacks on the character, intelligence, or integrity of opponents, rather than anything resembling policy disagreement.

The classification itself required some ingenuity. Standard sentiment analysis tools, the kind used in most computational political science work, can’t reliably distinguish between a senator calling a colleague’s economic proposals reckless and a senator calling the colleague personally corrupt. Both read as negative. Marc Jacob, Yphtach Lelkes, and Sean Westwood used GPT-4o to make that distinction instead, sorting roughly two-sentence text chunks from across the full Congress by whether they targeted a person’s character or their actions in office. Validation against two human PhD coders returned 97% accuracy for personal attack identification, which is, in political science terms, rather good.

What they found: most politicians don’t do this very much.

Thirty-five legislators (about 6%) never deployed a personal attack at all across the study period. Nearly two-thirds used them in fewer than 1% of their communications. A small tail, however, was doing something qualitatively different: devoting well over a fifth of their social media output to insults, character assassination, the full theatrical repertoire. The more a legislator leaned into personal attacks, the less time they spent on policy. That tradeoff was consistent across both chambers and both parties, which suggests it reflects a real choice, not just a stylistic tic.

Republican members engage in this behaviour at higher rates than Democrats, though the gap is a matter of degree rather than kind; both parties have their conflict entrepreneurs. The pattern sharpens at ideological extremes: more conservative Republicans are more likely to go personal, and among Democrats there’s an initial rise with extremity, though the most progressive members actually pull back slightly. Personal attacks also spike sharply in the months before elections, which is less surprising. And then, notably, the rate of substantive policy debate collapses and doesn’t recover afterward.

Media attention is where the model really clicks into focus. A legislator spending 5% of their communications on personal attacks receives, in terms of cable news coverage, the rough equivalent of a colleague devoting 45% of their output to critical policy debate. The 25 most conflictual members of Congress collected more cable news airtime than the 75 least conflictual members combined. On X, posts containing personal attacks were shared roughly 2.5 times as often as posts featuring policy criticism, and received about three times the likes. This is, it seems, what the media environment actually rewards: not expertise or legislative diligence, but the reliable production of conflict, the political equivalent of a recurring segment.

Here’s what doesn’t follow. No fundraising advantage. Despite the widespread assumption among political consultants that negativity rallies donors, conflict entrepreneurs raise neither more nor less money than their more civil colleagues. No electoral advantage either, once you account for how safe their districts are. Conflict entrepreneurs disproportionately come from safe districts, so their higher vote margins reflect geography, not rhetoric. No legislative advantage; they’re less likely to sit on prestigious committees. And, perhaps most striking, no personal financial gain during their time in office. Net worth changes showed no relationship whatsoever to rhetorical style.

The constituent alignment finding deserves particular attention. You might assume that legislators who specialize in personal attacks are simply giving their voters what they want, mirroring the partisan hostility of their districts. The data, drawn from MRP analysis of 140,000 survey interviews, say otherwise. There is effectively no correlation between how aggressive a legislator’s rhetoric is and how hostile their constituents are toward the opposing party. Many of the most vitriolic members come from districts that are, by survey measure, relatively moderate.

So why do it? The authors suggest the answer lies not in electoral theory but in the economics of attention. Political scientists have long assumed legislators are driven by three core goals: re-election, influence within the chamber, and good public policy. This research points toward a fourth, which has perhaps always existed but has become structurally viable only in the current media environment. Call it media celebrity. Visibility as an end in itself, decoupled from legislative achievement, from constituent demand, from anything resembling political power in the traditional sense.

The long-run financial picture remains murky. A conflict entrepreneur who builds enough name recognition might, after leaving office, command a media contract, a book deal, a speaking circuit. The study’s data don’t stretch that far, and that’s an honest limitation. But what the data do show is a political incentive structure that media organizations, social media platforms, and perhaps party leadership have collectively allowed to emerge. Personal attacks receive attention entirely out of proportion to how often they actually occur; the authors estimate their prevalence figures represent a lower bound, since implicit insults, visual content, and coded language all fall outside what text classification can catch.

The system, in short, appears to function as designed, even if no one designed it this way. Something produced a class of legislators for whom visibility and celebrity are the point, not the means to an end. Whether that turns out to matter for democratic governance is rather the question.

DOI / Source: https://doi.org/10.1093/pnasnexus/pgag038

Frequently Asked Questions