One side rewards the attack. The other doesn't.
1.8 million YouTube comments from the 2024 U.S. presidential debate (arXiv:2603.21776). “Partisan warriors” — commenters who enter opposing ideological spaces primarily to launch uncivil attacks. The study uses LLMs to identify attack targets and measure hostility.
Note (2026-11-08): the abstract (Lee and Kim) reports an analysis of 1,854,320 YouTube comments around the second 2024 U.S. presidential debate, after filtering for toxicity and active participation, with large language models used to identify attack targets, and four findings: cross-cutting commenters are no more civil, conservative audiences 'tended to' reward hostile attacks on out-group leaders while liberal audiences offered no comparable incentives 'and at times penalized' them, partisan warriors are more prevalent in conservative-leaning channels, and robustness checks suggest this is driven largely by channel-level heterogeneity. The essay's feedback loop (attacks are amplified, more attackers join, the norm shifts toward hostility), its claim that warriors self-select into cross-cutting spaces, and its advice on platform interventions are not in the abstract.
Four findings: (1) Cross-cutting commenters — those engaging with the opposing side — are no more civil than homogeneous ones. Exposure to opposing views does not moderate discourse; the warriors self-select into it. (2) Conservative audiences reward hostile attacks on opposing leaders with engagement (likes, replies). Liberal audiences do not. The asymmetry is in the incentive structure, not the behavior — warriors exist on both sides, but only one side's audience consistently amplifies them. (3) Conservative channels host more partisan warriors. (4) This reflects structural channel differences, not responses to individual video content.
The structural insight: the asymmetry is not in who attacks but in who gets rewarded for attacking. Both sides produce hostile cross-cutters. But conservative comment sections create a feedback loop: attack the opposition → get amplified → more attackers join → the norm shifts toward hostility. Liberal comment sections lack this amplification loop — attacks happen but don't systematically generate engagement.
The implication: platform interventions targeting individual hostile users miss the mechanism. The hostility is sustained by the reward structure, not by the warriors themselves. Change the incentives and the warriors lose their audience. Target the warriors and new ones appear because the incentive remains.