Sunstein's Colorado experiments with colleagues at the University of Chicago provided the definitive demonstration. Citizens were assembled into groups sharing general political orientations and asked to discuss charged issues — climate change, affirmative action, same-sex civil unions. Pre-discussion, participants recorded numerical views. Post-discussion, they recorded views again. Liberal groups became substantially more liberal, conservative groups substantially more conservative, and — the most consequential finding — internal diversity within each group collapsed. Before discussion, meaningful variation existed among liberals and among conservatives. After discussion, the groups had become internally homogeneous, not merely more extreme but more uniform in their extremity.
The algorithmic environment of contemporary communication functions as a polarization machine of unprecedented power. Social platforms sort users into like-minded enclaves with ruthless efficiency. The Colorado deliberations lasted hours and involved groups of eight. Current discourse involves millions, algorithmically sorted, operating continuously, with the most extreme positions receiving the widest distribution because extremity drives engagement and engagement drives distribution. The mechanism scales, and it scales badly.
The AI discourse of 2025–2026 replicated the Colorado pattern with textbook precision. Triumphalist enclaves heard disproportionately the arguments for unlimited capability; any expression of doubt became a social liability. Elegist enclaves heard disproportionately the arguments for cultural loss; any recognition of genuine benefit became a mark of collaboration. The silent middle — those whose private assessments most accurately tracked the mixed evidence — found themselves without institutional channels for expression, suffering the spiral of silence that converts moderate voices into silence or extremity over time.
The identity dimension intensifies polarization when the topic engages participants' sense of who they are. A discussion about tax rates rarely triggers this dynamic. A discussion about whether twenty years of professional mastery has been rendered economically worthless engages identity at its deepest level. When identity is engaged, challenging evidence is processed not as information to be weighed but as attack to be repelled. The epistemic question converts into a social question, and once that conversion occurs, rational deliberation diminishes sharply.
The phenomenon was first documented by James Stoner in 1961 in his study of 'risky shift' — the observation that groups tended toward riskier decisions than individual members. Subsequent research established that the shift was not specifically toward risk but toward the pole the group was already leaning toward. Sunstein's work from the 1990s onward extended the framework to political deliberation and to internet-mediated discourse, culminating in Going to Extremes (2009) and his broader analysis of how digital environments accelerate polarization.
Convergence toward extremity, not averages. Like-minded groups do not moderate through discussion; they intensify, producing post-deliberation positions no individual would have endorsed at the outset.
Two channels of influence. The mechanism operates through both informational skew (arguments surfaced in the group favor the shared leaning) and social pressure (individuals adjust positions toward the perceived group norm).
Diversity collapses internally. After polarization, groups become not merely more extreme but more uniform, eliminating the minority voices that would have provided corrective friction.
Algorithmic sorting amplifies the mechanism. Contemporary communication infrastructure sorts users into like-minded enclaves with efficiency that Colorado experimenters could not have imagined, scaling group polarization to civilizational dimensions.