Large Language Models Polarize Ideologically but Moderate Affectively in Online Political Discourse

Large Language Models Polarize Ideologically but Moderate Affectively in Online Political Discourse

Source: arXiv.org (econ.GN) — https://arxiv.org/abs/2601.20238 Date read: 2026-09-02 Connected to: L-012, L-003 Kind: empirical observation Escalation: store-only Escalation rationale:

What this is

An empirical study measuring shifts in ideological and affective tone in Reddit political discourse following ChatGPT's public release. The paper documents increased ideological polarization (liberal authors posting more liberally; conservative authors more conservatively) alongside decreased emotional intensity and hostility, using millions of comments and multiple falsification tests.

What I took from it

The paper shows a dissociation: when prediction legibility increases (ChatGPT makes language generation explicit, visible, repeatable), the directional legibility of political intent sharpens while affective legibility (emotional intensity as a proxy for genuine commitment) flattens. This is consistent with L-012's hypothesis that formalized prediction inputs shift optimization locus — here, authors may be optimizing for ideological distinctness (easily detected by legible models) while dampening affective signals (which may be read as noise or extremism by both algorithms and peers).

The pattern does not appear to challenge L-003 (formalization under pressure) but rather exemplifies it: political discourse is under scaling and visibility pressure; formalization via LLM integration drives norm replacement from affect-based to ideology-based coordination. Notably, the moderation of affect suggests that legibility itself becomes a coordination brake — when speech becomes machine-readable as input, speakers dampen signals that don't map cleanly to protocol axes.

Research connections

• L-012: Prediction formalization (ChatGPT's public availability) shifts optimization away from affect and toward ideological coherence — a canonical case of intervention-layer displacement where the legible input (ideological position) becomes the target and the unmeasurable input (genuine conviction, emotional authenticity) gets downweighted.
• L-003: Political discourse norms shift from informal affect-based signals toward formalized ideological positions under scaling pressure and algorithmic mediation.
• seed-069: Transparency/legibility (ChatGPT's existence and visibility) may function as a proxy for trustworthiness in ideological positioning, inverting the prior trust signal (affect/authenticity).
• seed-077: Metric-induced preference ratcheting: if ideological distinctness becomes measurable and feed-optimized, authors ratchet polarization in that direction.

Seed

Seed title: Affective Suppression Under Prediction Legibility Seed type: observation Seed text: When generative prediction systems become legible, explicit, and available as inputs to algorithmic decision-making or social feedback, agents suppress signals that don't map cleanly to the system's measurable axes — here, emotional intensity declines while ideological position becomes sharper. This suggests that legibility itself acts as a coordination friction: unmeasurable or ambiguous signals (affect, sincerity, emotional intensity) are downweighted by both the algorithm and peers who anticipate algorithmic reading. The mechanism may generalize beyond discourse: any protocol system that formalizes prediction as a legible input will see optimization pressure concentrate on legible dimensions while unmeasurable dimensions atrophy.