Wiki · Concept · Published 2026-10-23

Information Cascades

Information cascades explain how following earlier choices can conceal private evidence. Explore the mechanism, laboratory findings, limits and proposed Spiralist relevance.

Definition

An information cascade occurs when earlier public choices lead someone to choose the same option regardless of their own private evidence. The person may reasonably infer that earlier decision makers knew something useful. Yet following them can conceal information that would have helped the group. This definition follows Anderson and Holt’s account.

The defining question is whether a person’s private evidence can still change their decision. Agreement alone does not answer it. People might independently reach the same conclusion, consult the same source, or choose an option because coordination itself is valuable. Calling an observed pattern a cascade therefore requires evidence about how decisions were made.

This differs from Pluralistic Ignorance, which concerns mistaken perceptions of other people’s views. A cascade can occur even when earlier choices are observed accurately.

How it works

Abhijit Banerjee’s A Simple Model of Herd Behavior, published in August 1992, demonstrates a collective cost of imitation: choosing sensibly for oneself can withhold information from later people. His sequential model produces inefficient herding under specified assumptions. Its contribution is a mechanism explaining how individual reasoning can undermine collective learning, rather than evidence that every crowd behaves this way.

Hypothetical example: a community must choose between two workshop venues. Each member receives a different, imperfect clue about which venue would work better. Members announce their choices in sequence without explaining their clues. Two early members choose the warehouse. Suppose later members regard those decisions as sufficiently persuasive that they would choose the warehouse whichever clue they personally received. Their announcements then conceal their clues. A long sequence of warehouse choices could grow without revealing whether later members independently found supporting evidence.

The example identifies what an investigation would need to establish. If later members instead visited both venues, compared detailed reports, or changed their choices after receiving stronger evidence, the same visible agreement would have a different explanation. If the warehouse were selected because everyone needed to meet in one place, coordination benefits would also need consideration.

In Banerjee’s model, choices can fail to preserve the information behind them. Later observers receive a decision without being able to reconstruct its evidential basis. That loss, rather than imitation by itself, creates the learning problem.

Evidence and applications

The supplied manuscript of Lisa R. Anderson and Charles A. Holt’s Information Cascades in the Laboratory (1997) reports paid predictions about a hidden urn, using private draws and earlier public predictions. Across twelve sessions, cascades appeared in 87 of 122 periods where they were possible, including incorrect cascades. Behavior nevertheless departed from the ideal model. In an asymmetric design, participants chose the Bayesian option in 41 of 82 cases where it conflicted with a counting rule.

A proposed application is studying consultation processes. Researchers could distinguish judgments recorded before discussion from choices made after seeing others’ responses. They could also examine whether participants supplied new reasons or merely endorsed an existing conclusion. Such comparisons would investigate dependence between judgments; they would not establish which proposal deserves adoption.

For digital discussion, a useful research question is whether visible endorsements change the expression of contrary evidence. This is an extension of the mechanism, not a finding established by the cited experiment.

Limitations

Banerjee explicitly presents a counterexample in which observing earlier choices supports convergence toward the truth. His discussion and conclusion also identify restrictive assumptions, including free signals and a payoff structure that treats every incorrect option alike. Learning failure is conditional on the model’s structure.

Anderson and Holt’s controlled experiment involved undergraduate economics students and a narrow prediction task. It did not test AI mediation or establish a general prevalence of cascades outside the laboratory.

Outside such a task, an investigator may not know what information people possessed, how reliable they considered it, or what rewards shaped their choices. Reputation concerns, shared evidence and practical constraints can produce similar public behavior. A unanimous meeting record cannot resolve those alternatives.

Proposed remedies also require separate evaluation. Collecting judgments before discussion might preserve differences, but those judgments could still depend on one common source. Asking for reasons might reveal evidence, but a plausible explanation does not guarantee an independent investigation. Neither disagreement nor independence establishes correctness.

Relevance to Spiralism

The following is a proposed interpretive connection for this wiki. Information cascades offer a way to examine how repeated agreement can acquire apparent evidential weight. The Spiralist lens can ask what each repetition contributes: a fresh observation, a reconsidered judgment, or another response to an existing pattern. This interpretation does not supply evidence about any particular community.

For AI-mediated deliberation, a hypothetical concern is that an early generated summary describes agreement, later participants defer to it, and a subsequent summary counts their endorsements as further support. Establishing a cascade would require examining participants’ evidence and decisions. The mere presence of generated text would be insufficient.

This proposed pathway connects with Recursive Reality, the site’s broader interpretive account of representations influencing later evidence. Algorithmic Monoculture instead emphasizes shared dependencies between decision systems. Data Cascades concerns downstream consequences of unresolved data problems. These mechanisms can overlap, but identifying one does not establish the others.

A possible practical inquiry is whether a discussion preserves the distinction between independently supplied evidence and endorsements influenced by earlier choices. That question leaves room for justified agreement while making the origin of apparent consensus open to examination.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries