The Consensus Trap

Abstract

AI’s growing capacity to synthesize evidence and settle arguments creates an illusion: that settling what is true will also settle what people do. It will not. Behavior is governed by affect — shame, interest, threat, loyalty — not by the truth-value of a conclusion, and no increase in synthetic power changes that. This paper argues that AI chatbots operate top-down, moving from question to evidence to conclusion, while human beings operate bottom-up: a precognitive affective verdict, rooted in scripted history and lateral habenula activity, arrives before conscious appraisal is even complete. Drawing on Silvan Tomkins’ affect theory and Donald Nathanson’s Compass of Shame, the paper traces how this asymmetry plays out at the scale of an online argument going quiet, and at the scale of institutions increasingly deferring to AI-generated answers. The central risk identified is not classical, coercive authoritarianism but a “totalitarianism of preference” — a consolidation of agency that requires no enforcement because deference to AI is simply cheaper, affectively, than the discomfort of unresolved disagreement. The paper also addresses how this dynamic compounds as institutional habits formed around technical questions migrate into value questions, where no single computable answer exists. Keywords: artificial intelligence, affect theory, Silvan Tomkins, Compass of Shame, lateral habenula, sycophancy, AI alignment, agency, deference, totalitarianism