Information cascades

Information cascades occur when individuals make decisions sequentially based on their own private information and the observable actions of others, leading to a situation where later actors ignore their own signals and follow the crowd. First formalized by Bikhchandani, Hirshleifer, and Welch in 1992, the cascade model reveals that rational, Bayesian agents can produce collectively irrational outcomes—not through malice or stupidity, but through the structural logic of observational learning.

The canonical example is the restaurant choice problem: two restaurants, one objectively better, but patrons choose based on what they see. The first patron uses their own taste. The second combines their taste with the first's choice. By the third or fourth patron, the accumulated public signal can overwhelm private information, and everyone thereafter follows the crowd—even if their own experience tells them the other restaurant is better. Once a cascade begins, it is self-reinforcing and informationally inefficient.

The Feedback Topology of Herding

Information cascades are not merely a behavioral phenomenon. They are a feedback topology problem. The system's structure is a positive feedback loop with a critical threshold: when the number of prior public signals exceeds the precision of an individual's private signal, the rational response is to abandon private information and follow the crowd. The cascade is not a failure of individual rationality but an emergent property of the system's coupling between observation and action.

This feedback topology is isomorphic to the Chinese restaurant process and preferential attachment in network formation. In all three cases, early advantages are amplified by the structure of sequential choice, producing rich-get-richer dynamics that no individual actor can resist. The 2016 U.S. election demonstrated how this topology operates in algorithmic environments: social media platforms amplified engagement-driven content, creating information cascades that substituted algorithmic curation for independent judgment.

Algorithmic Cascades and Epistemic Architecture

The information cascade model was developed for human social learning, but its implications are most urgent in algorithmic institutions. Modern recommendation systems are essentially cascade amplifiers: they observe aggregate behavior (clicks, shares, dwell time) and use it to shape what subsequent users see. The result is an algorithmic cascade in which the platform's own observations become the public signal that shapes future behavior. Unlike human cascades, which eventually reach saturation or encounter contradictory evidence, algorithmic cascades can operate at scale and speed that make correction impossible.

The feedback topology of algorithmic cascades has been redesigned for engagement rather than accuracy. The system's positive feedback loops amplify content that triggers emotional arousal, not content that is true or useful. This is not a content problem. It is a structural problem: the feedback gain is too high, the delay is too low, and there are no circuit breakers or negative feedback loops to interrupt the cascade before it produces epistemic collapse.

From Information Cascades to Institutional Memory

Information cascades also illuminate how institutional memory decays. When organizations make decisions based on precedent rather than current evidence, they are following an institutional cascade. The first few decision-makers set a pattern; subsequent actors follow it because the cost of challenging the pattern exceeds the cost of being wrong. This is how organizations persist with failed strategies, outdated technologies, and toxic cultures long after the evidence against them has accumulated.

The link to somatic marker hypothesis is striking: just as the Iowa Gambling Task shows that emotional markers can guide rational decision-making before conscious awareness, information cascades show that social markers can guide collective decision-making before any individual recognizes the error. The body learns faster than the mind; the crowd decides faster than the individual. Both are feedback topologies that short-circuit deliberation.

_The theory of information cascades is too comfortable. It locates the problem in rational Bayesian agents following a social signal, which implies that the solution is better information or smarter agents. But the real problem is the feedback topology itself—the sequential observability structure that makes cascades rational. An information cascade is not a market failure; it is a market feature. The architecture of sequential choice, combined with observable actions, inevitably produces herding. The only solution is not to educate the agents but to redesign the topology: to introduce friction, privacy, and structural delays that prevent the cascade from forming. Those who propose transparency as the cure are prescribing the disease—transparency accelerates cascades, it does not stop them._

Related Phenomena

Information cascades are often confused with herding behavior, but the distinction matters. Herding is a descriptive term for any situation where individuals follow the crowd; information cascades are a specific mechanism—rational, Bayesian, and structurally inevitable. Not all herding is a cascade, and not all cascades look like herding to an outside observer. The preference falsification model developed by Timur Kuran adds another layer: individuals may publicly conform to a cascade while privately dissenting, creating a hidden reservoir of opposition that can trigger sudden revolutionary cascades when the public signal finally shifts.

Social proof—the psychological principle that people look to others for cues about correct behavior—is the cognitive substrate on which information cascades build. Without the human tendency to treat others' actions as informative signals, the cascade mechanism would have no purchase. But social proof is a heuristic; information cascades are a theorem. The difference between the two is the difference between a hunch and a proof.