Information Cascade

An information cascade occurs when individuals make decisions sequentially, observing the actions of those before them, and rationally choosing to follow the crowd even when their private information suggests a different choice. The phenomenon was formalized by economists Banerjee (1992) and Bikhchandani, Hirshleifer, and Welch (1992), and it demonstrates that locally rational behavior can produce globally irrational outcomes — herding that overrides genuine private signals.

The classic model assumes agents with private signals of varying quality who act in sequence. Early actors reveal their information through their choices. Later actors, seeing the accumulated public signal, may find it so informative that they ignore their own contradictory private signal and follow the crowd. Once this happens, the cascade becomes self-sustaining: subsequent actors see only the same public signal, and no new private information enters the public record.

Cascades and Epistemic Infrastructure

Information cascades are not merely cognitive phenomena; they are infrastructurally mediated. The speed, visibility, and topology of the network in which sequential decisions occur determine whether cascades form, how deep they run, and whether they can be broken. A social media platform that amplifies early signals through algorithmic promotion is an infrastructure designed to produce cascades — not because its designers intended herding, but because the engagement-optimization target systematically rewards high-visibility early signals.

The connection to algorithmic curation and algorithmic amplification is direct: when a platform's ranking function promotes content that is already receiving attention, it creates the informational equivalent of a sequential decision environment. Users observe what is trending, infer that others have found it valuable, and rationally attend to it — even if their own unmediated judgment would rate it as noise. The result is a filter bubble not of explicit preference but of cascade dynamics: the information environment converges on a small set of high-arousal signals, and epistemic diversity collapses. In degraded information ecosystems, cascades become the primary mechanism by which misinformation achieves population-scale saturation.

Breaking Cascades

Information cascades can be broken by three mechanisms: (1) the arrival of a highly visible signal that contradicts the cascade, (2) the revelation that early actors were poorly informed, or (3) institutional designs that protect private signals from being swamped by public ones. Scientific peer review, secret ballots, and adversarial legal procedures are all institutional technologies designed to prevent information cascades by making some private information temporarily non-public.

The design challenge for epistemic infrastructure is to maintain enough diversity in the information environment that cascades do not become permanent attractors. This requires not merely "diverse viewpoints" but diverse *discovery mechanisms* — multiple, partially decoupled channels for finding and evaluating information, so that a cascade in one channel does not immediately colonize all others.

Social Media as Cascade Infrastructure

The transformation of information cascades by social media is not merely a matter of scale — more people, faster sharing — but of architecture. Traditional cascade models assume sequential decision-making in a transparent environment where later agents can observe earlier agents' choices. Social media platforms violate both assumptions simultaneously. The environment is not transparent: algorithmic curation means that each user sees a different subset of the public signal, filtered by engagement predictions. And the decision sequence is not linear: millions of users act in overlapping waves, each influenced by different samples of the "public" signal, producing what is better described as a viral dynamic than a classical cascade.

This architectural transformation changes the mathematics of cascades. In the classical model, cascades are fragile: a single contradictory signal from a well-informed early agent can break them. On social media, the "early agents" are algorithmically selected for their predicted engagement value, not their informational quality, and the cascade is reinforced by platform amplification rather than merely observed by subsequent agents. The result is that social-media cascades are far more persistent than classical cascades, and far less responsive to corrective information. The infrastructure has been designed to produce herding; the surprise is not that herding occurs, but that anyone expected otherwise.

The systems-theoretic point: information cascades on social media are not a behavioral pathology but an architectural achievement. The platform has built a system that converts individual attention into collective convergence, and the convergence is the product. The question is not how to prevent cascades but how to build infrastructure that makes cascades visible and interruptible before they achieve population-scale saturation.