Read the original article here.

---

Understanding Information Cascades in the Era of the Dead Internet

This document explores the concept of Information Cascades, a phenomenon from behavioral economics and network theory, and examines how it manifests and is potentially distorted or manipulated in the context of the modern internet, particularly as theorized by "The Dead Internet Files" perspective.

What is an Information Cascade?

Information Cascade: A phenomenon where individuals make the same decision in sequence, primarily influenced by observing the choices of those who acted before them, rather than relying solely on their own private information or beliefs.

While similar to Herd Behavior, an information cascade is distinct. Herd behavior can be driven by various factors, including social pressure or direct imitation. An information cascade, however, is specifically driven by the inference of information from observing others' actions. People infer that earlier actors likely had good reasons (information) for their choices, and this inference overrides or supplements their own private signal.

In the context of "The Dead Internet Files," where a significant portion of online activity is theorized to be generated by bots and automation, understanding information cascades is crucial. If bots constitute a large number of early "actors" in an online sequence, their actions can artificially trigger or shape cascades, leading human users to follow non-human "information."

The Two-Step Process of an Information Cascade

An information cascade typically unfolds in a sequential, two-step manner:

1. Encountering a Decision: An individual faces a choice, often a binary one (e.g., adopt/reject, buy/not buy, believe/disbelieve).
2. Observing Others: The individual observes the decisions made by people who came before them in the sequence and infers information from those observable actions. This external, public information can influence or even override their own private information or initial inclination.

This process highlights the importance of observability of actions and the sequential nature of decisions. Online, this is incredibly prevalent: seeing how many people liked a post, bought a product, signed a petition, or shared an article happens constantly and sequentially as feeds update and follower counts change.

Core Components of an Information Cascade

The structure of an information cascade can be broken down into five key components:

1. A Decision is Required: There must be a specific choice point (e.g., deciding whether a piece of online information is true, whether to click a link, whether to follow a user).
   • Dead Internet Context: These decisions are myriad online – what to trust, who to follow, what to buy, what content to engage with. Bots can act as early "deciders" or amplifiers, presenting a distorted view of public opinion or popularity.
2. Limited Action Space: The decision is typically restricted to a few clear options (e.g., "like" or not "like," share or not share, click or not click, accept or reject).
   • Dead Internet Context: Online platforms often enforce limited actions (upvote/downvote, retweet/quote tweet, like/react) which simplifies observable signals, making cascades easier to initiate and track, even by automated means.
3. Sequential Decisions & Observability: People make decisions in a specific order, and each person can see the choices made by those who preceded them.
   • Dead Internet Context: Social media feeds, comment sections, online review platforms, and trending lists all present information and user actions in a sequential manner. Bots can easily generate rapid, early activity (likes, shares, comments) visible to subsequent human users.
4. Each Person Has Some Private Information: Individuals possess their own internal signal or knowledge about the correct decision, independent of what others do.
   • Dead Internet Context: A human user might have prior knowledge, critical thinking skills, or a specific experience that informs their private signal. However, if a large number of bot-generated "actions" contradict this private signal, the inferred public information can overwhelm it. Bots, by definition, lack genuine "private information" in the human sense; their "signals" are programmed.
5. Private Information is Not Directly Observable: A person cannot see the private signal or reasoning of others, only their resulting action. They must infer the information others held based on the choices they made.
   • Dead Internet Context: This is key online. We see that a post has 10,000 likes, but we don't know why each person liked it, who those people are (human or bot), or what their personal assessment was. This opacity allows bot activity to be mistaken for genuine human consensus or endorsement.

Social perspectives suggest that social pressure can also drive people to act against their private signals, complementing the purely informational aspect of cascades. Distinguishing information cascades from concepts like Social Proof (following the crowd because you assume they know something you don't), Information Diffusion (how information spreads), and Social Influence (broader pressure to conform) is important, although the term "information cascade" is sometimes used more broadly.

The Basic Model Explained

The basic model of information cascades, like the one described by Bikchandani et al. (1992), helps illustrate how observing others' actions can lead individuals to disregard their own information.

Qualitative Example (The Urn Experiment):

Imagine an experiment with two urns, A and B. Urn A has more 'a' balls, Urn B has more 'b' balls. Participants sequentially draw one ball from a randomly chosen urn (contents mixed in a neutral container), observe its label ('a' or 'b' - this is their private signal), and then publicly announce which urn they believe the ball came from. Subsequent participants hear previous announcements and have their own private signal from the ball they draw.

• Scenario: Urn A (more 'a' balls) is chosen.
• Participant 1: Draws an 'a'. Their private signal matches the likely correct answer (Urn A). They announce "Urn A".
• Participant 2: Draws an 'a'. Their private signal is 'a'. They see Participant 1 said "Urn A". Both signals point to A. They announce "Urn A".
• Participant 3: Draws a 'b'. Their private signal is 'b' (suggesting Urn B). But they see the first two participants said "Urn A". The weight of public opinion (two "A" decisions) now potentially overrides their single "B" signal, especially if the signals ('a'/'b' vs. which urn is correct) are somewhat uncertain. They might announce "Urn A", even though their private signal suggested otherwise.
• Subsequent Participants: If Participant 3 says "Urn A", Participant 4 sees three "A" decisions. Even if they draw a 'b', the public evidence for "Urn A" is mounting (three "A" vs. one "B" from previous participants + their own "B"). It becomes increasingly rational, based purely on the observable actions of others, to ignore their own signal and follow the crowd, assuming the previous three people had stronger or more numerous private signals pointing to 'A'.

Reverse Cascade: An information cascade that leads to an incorrect outcome, where individuals follow the crowd's decision even when it contradicts their private, accurate information, resulting in the propagation of a wrong choice or belief.

In the Urn Experiment, a Reverse Cascade occurs if enough early participants guess incorrectly based on misleading private signals, causing later participants with correct signals to follow the incorrect public trend.

Dead Internet Context: In the online world, "private signals" could be a user's critical evaluation of information, their knowledge of a topic, or their skepticism. "Observable actions" are likes, shares, comments, reviews, etc. Bots can artificially generate a high volume of "A" actions (e.g., liking a false statement or promoting a low-quality product). A human user with a "B" private signal (e.g., recognizing the statement is false or the product is poor) sees overwhelming public evidence ("A" actions) and might rationally (based on the limited, observable information) decide to disregard their own signal and act as if "A" is correct (share the false statement, buy the product), triggering a Reverse Cascade driven by automation.

Quantitative Description (Simplified):

The model uses probability to explain the decision-making process. An individual calculates the likelihood of the true state (e.g., "Accept" or "Reject") given their private signal and the observed history of previous decisions.

• Signals: H (High signal, suggesting Accept) and L (Low signal, suggesting Reject).
• True States: A (Accept is correct) and R (Reject is correct).
• Assumptions: P[H|A] > 0.5 and P[L|R] > 0.5 (signals are more likely to match the true state).
• p: Prior probability that A is the correct decision (often starts at 0.5, meaning equally likely).
• q: The probability P[H|A] or P[L|R] (signal accuracy, > 0.5).

An agent updates their belief using Bayes' rule based on their signal. For example, seeing an H signal increases the belief that A is correct.

Subsequent agents consider the history of a accept decisions and b reject decisions made publicly, plus their own signal. They calculate the probability of A given the observed actions and their signal. If the number of public 'Accept' signals significantly outweighs 'Reject' signals (e.g., a is much greater than b), this public information can dominate the private signal, causing the agent to join the cascade of 'Accept' even if their signal was 'L'.

Dead Internet Context: Bots don't have private signals in the probabilistic sense. They are programmed to generate a certain action ('Accept' or 'Reject' for a specific item) at scale. If a bot network is programmed to generate 'Accept' actions for a product or idea, they quickly skew the public a and b counts. Human users observing this skewed history will perform their Bayesian calculation based on artificial public data, making it appear statistically rational to join the bot-driven cascade, even if their private signal or experience contradicts it. The (1-p)(1-q)^a q^b part of the equation becomes distorted because a and b include actions from non-agents without q-level signal accuracy.

Explicit Model Assumptions and Their Implications (Especially Online)

The original model relies on several assumptions:

1. Boundedly Rational Agents: Individuals make rational choices based only on the information they can observe (public actions + their private signal), which is often incomplete. They don't have perfect knowledge.
   • Dead Internet Context: This assumption holds true for human users online. We have limited information and make decisions based on visible cues (likes, shares, reviews). Bots exploit this; they don't need to be rational, they just need to generate observable actions that a boundedly rational human agent will process.
2. Incomplete Knowledge of Others' Private Information: Agents only see what others did, not why they did it (their private signal or reasoning).
   • Dead Internet Context: Crucial online. We see a user shared something, but not their internal thoughts or external information source. Bots thrive on this opacity – their "reasons" are just code, but their actions (likes, shares) look just like human ones and are factored into human calculations.
3. Observable Behavior of All Previous Agents: The sequence and decisions of prior actors are known.
   • Dead Internet Context: Online platforms make many actions highly visible and quantifiable (like counts, view counts, retweets). This perfect observability of previous actions is a necessary condition that bots can easily satisfy and manipulate.

Resulting Conditions of Cascades

Based on these assumptions, the model predicts several conditions:

1. Cascades Will Always Occur (Eventually): As the number of sequential decisions increases, the likelihood of a cascade forming (where subsequent agents ignore their private signals) approaches 1.
   • Dead Internet Context: On large online platforms with millions of interactions, cascades are not just likely, they are almost guaranteed. Bot activity accelerates this process by increasing the rate and volume of early observable actions, hitting the "tipping point" for a cascade faster.
2. Cascades Can Be Incorrect: Because decisions can be based on limited or misleading early signals, a cascade might converge on the wrong outcome.
   • Dead Internet Context: This is a primary concern with bot activity. If bots promote a false narrative or a poor product, they can initiate an incorrect information cascade. Humans, relying on the public signal amplified by bots, may adopt the false belief or make the poor choice.
3. Cascades Can Be Based on Little Information: A cascade can start after only a few early decisions, potentially overriding strong private signals from many later individuals.
   • Dead Internet Context: A coordinated bot attack doesn't need to involve a massive number of bots overall, just enough to create a rapid, early surge of activity that tips the scales and triggers a cascade before organic human engagement can provide countervailing signals.
4. Cascades Are Fragile: Because later participants in a cascade ignore their private information, the collective decision is less informed than it appears. A small amount of new public information or a shock can easily break the cascade and potentially start a new one.
   • Dead Internet Context: While bots can start cascades, they can also be used to break them or shift them. A new wave of bot activity promoting a counter-narrative or opposing view can disrupt an existing cascade. The fragility means online consensus built on cascades (especially bot-influenced ones) is inherently unstable and prone to rapid shifts.

Responding to Information Cascades (Strategies & Manipulation)

Understanding cascades allows entities (businesses, political actors, bots) to strategize:

• Early "Guinea Pigs" / Seed Users: Firms might provide free products or early access to influential users (or even use bots disguised as users) to generate positive early public signals and kick-start a buying or adoption cascade.
• Early Public Tests / Reviews: Rigging or amplifying early positive reviews (potentially with bot accounts) on e-commerce platforms can create a powerful signal for later buyers. Even "tough" tests that appear authentic can be faked or amplified by bots to appear more convincing.
• Price Signaling: In uncertain markets (like selling a house or launching a new product), initial high prices might signal high quality. Failure to sell could be attributed to price, not low quality, preventing a negative quality cascade from starting. This is less directly applicable to bots but shows how signals are interpreted.
• Monopolist Pricing: A firm unsure of its product quality might adjust pricing based on early sales data, inferring aggregate consumer signals (which could be influenced by early bot purchases).

Dead Internet Context: These strategies are amplified and perverted by bots. Bots can be the "guinea pigs" or early buyers. They can write the early public reviews and testimonials. They can amplify manufactured test results. This allows manipulators to artificially generate the critical mass of early positive signals needed to trigger an information cascade, even if the underlying product or idea is poor or false.

Examples and Fields of Application

Information cascades are prevalent wherever people make decisions based on observing others, especially online.

• Reputational Cascades: Individuals follow the crowd not just because they think the crowd is right, but because they fear their own reputation will be damaged if they dissent.

  • Dead Internet Context: This is powerful online. Users fear being seen as contrarian, ill-informed, or out of step if they disagree with a seemingly dominant opinion (even if that opinion is bot-driven). Retweeting a viral (potentially bot-amplified) tweet, even if you have doubts, avoids the reputational risk of questioning it.

• Market Cascades (Financial): Speculation and bubbles can form as investors observe others buying an asset and infer that it's a good investment, regardless of underlying fundamentals.

  • Dead Internet Context: Online trading platforms, crypto markets, and stock forums are ripe for bot-driven cascades. Bots can generate trading activity, spread rumors, and amplify sentiment (e.g., "buy" signals) that human traders then follow, creating artificial pumps and dumps.

• Product Adoption/Marketing: Getting early adopters (or faking them with bots) is key to triggering a cascade for new products.

  • Dead Internet Context: Bot reviews, fake social media buzz, and manufactured popularity metrics ("trending products") are direct attempts to start buying cascades online.

• Social Networks and Social Media: Information, trends, and beliefs flow as cascades.

  • Dead Internet Context: This is the epicenter of the "Dead Internet" concern.
    • Virality: Bots don't just spread information; they simulate virality by performing actions (likes, shares, comments) that make content appear popular, triggering human users to share it and contribute to a cascade based on artificial popularity.
    • Influencing Public Opinion: Coordinated bot networks are explicitly used to start cascades around political narratives, social issues, or consumer trends. By generating early, visible activity, they create the illusion of widespread belief or support, which then influences human opinion via cascade effects.
    • Distorted Discourse: If bots dominate early interactions, the "observable actions" available to humans are heavily skewed, making it harder for genuine human signals to break through or initiate competing cascades. The online environment becomes a feedback loop of automated actions influencing human perception of consensus.
    • Network Restructuring: Bot-driven cascades can even alter social network structures. As viral (bot-amplified) content spreads, users may follow or unfollow others based on who is participating in or promoting the cascade. Bots, while not forming genuine ties, contribute to the appearance of shifting connections around specific content or users.

• Social Influence Model vs. Information Cascades: The social influence model suggests people know some private beliefs of their network contacts and that beliefs exist on a scale, not just binary adopt/reject.

  • Dead Internet Context: This distinction highlights a vulnerability. Online, especially with ephemeral content and large networks, we often lack deep knowledge of others' beliefs, making the information cascade model's assumption (relying only on observable action) more relevant. Bots exploit this lack of deep connection and understanding, presenting only the superficial, observable action.

• Historical Examples:

  • Leipzig Protests (1989): A small group's persistent action became a visible signal, encouraging more people to join each week, eventually leading to a massive cascade of protest and the fall of the Berlin Wall. This shows a cascade driven by courageous human action becoming public information.
  • Hybrid Seed Corn Adoption: Farmers relied on the observable results and opinions of trusted neighbors (a form of public signal and social influence) rather than just salesman pitches (private signal from a potentially biased source). The slow adoption was partly a cascade effect waiting for enough neighbors to provide positive public signals.

• Empirical Studies:

  • Urn Experiment: Demonstrates cascades in a controlled lab setting.
  • Movie Success: Statistical models show movie box office success often follows a cascade-like pattern where early viewership and buzz (public signal) strongly influence later attendance, sometimes leading to "flop" cascades despite quality, or "hit" cascades despite mediocrity (like the Hollywood example).
  • Technology Adoption: Studies on businesses adopting new tech show cascade effects; firms are influenced by observing which technologies competitors or peers have adopted.
  • Dead Internet Context: These studies provide frameworks that can be applied to online behavior, revealing how easily quantifiable actions (views, sales, adoptions) can trigger cascades, making these processes prime targets for bot-driven manipulation.

Legal and Societal Aspects

Information cascades can have significant societal impacts, sometimes requiring legal intervention:

• Voting Order: In some contexts (like military courts), junior members vote first. This prevents higher-ranked members' decisions (carrying perceived greater authority/information) from creating an information cascade that influences lower ranks, ensuring votes are based more on individual judgment.
• Election Polling Bans: Laws prohibiting the release of election polls close to voting day aim to prevent poll results (a public signal about others' likely choices) from triggering an information cascade that sways undecided voters or suppresses turnout based on perceived momentum.

Dead Internet Context: The potential for bots to manipulate online opinion (via cascades on social media, comment sections, etc.) is a major legal and ethical challenge. How do you regulate or mitigate cascades driven by artificial agents influencing political discourse, public health information, or market sentiment? The current legal frameworks designed for human-driven cascades are ill-equipped for this.

Globalization and Modern Society

Information cascades are increasingly relevant in a globalized, interconnected world:

• Cultural Differences: Studies show cascade behavior can differ across cultures dueating to varying trust in authority, reliance on social networks, or interpretation of signals.
• Financial Volatility: Global financial markets are susceptible to rapid, cascade-driven panics or bubbles as traders worldwide react sequentially to perceived trends, amplified by instantaneous information flow.
  • Dead Internet Context: Automated trading algorithms (bots) significantly contribute to the speed and scale of financial cascades, reacting to market signals (which may themselves be influenced by other bots or manipulated news) and triggering massive, near-instantaneous movements.
• Spread of Tactics/Ideas: The adoption of terrorist tactics or protest strategies can spread via cascade effects as groups observe successful actions elsewhere and imitate them.
• Transnational Information Flow: Cascades are a fundamental way information (and misinformation) spreads across borders through global online platforms.
  • Dead Internet Context: Bots play a crucial role in transnational information operations, pushing narratives or creating artificial trends that trigger cascades across different countries and cultures, shaping global discourse with automated influence.

In summary, information cascades are a natural human phenomenon driven by rational inference from observable sequential actions. However, in the context of "The Dead Internet Files," this phenomenon is subject to significant distortion. Bots and automation can act as artificial "early participants," generating observable actions at scale that mislead human users into joining cascades based on fabricated popularity, consensus, or trends. This transforms the rational process into a vulnerability, allowing automated systems to silently shape human behavior and perceptions online, contributing to the sense that parts of the internet are no longer authentically human-driven.