Asymmetries between increasing and decreasing polarization create ratchet effects where divisions easily intensify but prove difficult to reduce. The research found that algorithms can both increase and decrease polarization, but real-world conditions favor the former while impeding the latter.
Among over 1,000 users during the 2024 presidential election, increasing divisive content in feeds produced polarization increases, while decreasing such content reduced polarization. The effects appeared roughly symmetrical in the controlled experiment. But real-world platforms face asymmetric incentives.
Business models reward engagement, which divisive content generates effectively. This creates constant pressure toward polarization-increasing algorithms. Meanwhile, reducing polarization would require resisting these incentives and accepting lower engagement, which platforms have little motivation to do voluntarily.
Psychological factors also create asymmetry. Negative emotions and threat perceptions activate more intensely than positive emotions and perceptions of safety. Content promoting outrage and fear thus generates stronger reactions than content promoting understanding and cooperation. Algorithms following engagement signals therefore naturally favor polarization even without deliberate intent to divide.
Social dynamics compound these asymmetries. Polarized communities create social pressure for conformity and punish members who engage constructively across partisan lines. This makes depolarization socially costly for individuals even when they might prefer less hostile politics. Algorithmic interventions reducing polarization must overcome not just platform incentives but also social dynamics that resist depolarization.
Breaking ratchet effects requires deliberate interventions that account for asymmetries. Regulations might mandate depolarization efforts that platforms won’t undertake voluntarily. Social movements might create communities that reward rather than punish cross-partisan engagement. Technical interventions might amplify constructive content more aggressively than current algorithms amplify divisive content, counteracting natural asymmetries.