Skip to content

Lecture 4 — Emergence of Collective Patterns (Klein)

Paper: Douven, I. & Hegselmann, R. (2021). Mis- and disinformation in a bounded confidence model. Artificial Intelligence, 291, 103415.

Type: Methodological — the highest level in Newell's hierarchy that this course covers. Where L2 modelled individual choices and L3 modelled individual–machine interaction, L4 zooms out to societies of agents and the emergent patterns their interactions produce.

One-paragraph framing

Some properties of an open society are visible only at the collective level — polarization, segregation, inequality, social norms, gendered division of labour. Individuals act and politics defines the playing field, but the resulting macro-pattern can be counter-intuitive (the "sum is more than its parts"). To reason about this rigorously you need a method that lets the macro-pattern emerge from explicitly modelled micro-interactions: agent-based modelling (ABM). The lecture works through ABM theory, then applies it to opinion dynamics — specifically the Hegselmann–Krause bounded-confidence model and its mis-/disinformation extension (Douven & Hegselmann 2021).

Key concept 1 — Emergence and Coleman's bathtub

Emergence: complex social patterns arise from the interactions of individuals; the resulting patterns may have properties that do not immediately follow from individuals and their properties. Slogan: "The sum is more than its parts."

Coleman's bathtub (explanation in the social sciences):

   Macro-Structure ───────────────────────►  Updated Macro-Structure
        │                                              ▲
        │ Situation                          Aggregation │
        ▼                                              │
       Actor  ────── Selection ────►  Action ──────────┘

Methodological individualism: every explanation of a social-level phenomenon needs to refer to individual agents — how their perception of the environment and their (inter)actions give rise to the social-level phenomenon. ABM operationalises this.

Toy explananda

  • Walking on the right-hand side of the street
  • Panic buying
  • "Everyone using the same messenger app"

Examples of macro-only patterns

  • Polarization (Hegselmann & Krause 2002; Douven & Hegselmann 2021)
  • Segregation (Schelling 1969)
  • Inequality (Klein et al. 2020)
  • (Social) Norms (Bicchieri 2005; Lisciandra et al. 2013)
  • (Gendered) Division of Labour (O'Connor 2019)

Key concept 2 — Five emergent-pattern phenomena to remember

These are the analytic vocabulary for talking about ABMs:

  1. Emergence itself — patterns at the macro that don't sit "in" any individual.
  2. Path dependencies — small random effects early in a model can have large lasting effects. Examples: emergence of firms, preferential attachment in social networks.
  3. Tipping points — minuscule changes in parameters can have drastic effects on output; the input → output function need not be continuous. Examples: climate change, sudden emergence of trust/revolutions.
  4. (Non-)monotonicity — more of an ingredient can sometimes reverse the effect direction. Loosening agents' confidence interval can increase polarization, not decrease it.
  5. Direction of effect — how individual motives relate to the emergent pattern. Four possibilities:
  6. Aligned with individual motive (and may even be stronger than it — Schelling segregation: weak in-group preferences → strong segregation)
  7. Accidental (bar example — everyone happens to converge by chance)
  8. Opposed to individual motivation (Adam Smith's invisible hand — selfish behaviour produces collectively useful market)
  9. No individual counterpart at all (trends, flocking)

Key concept 3 — Hegselmann–Krause bounded-confidence model

A canonical opinion-dynamics ABM (Hegselmann & Krause 2002).

Setup

  • Stance on a [0, 1] scale.
  • Agents have a private starting position.
  • They want to learn from each other but must decide whom to listen to.
  • Confidence level ε: each agent only listens to others whose opinions are within ε of its own.

Update rule

Each step, new opinion is the average over all opinions within ε:

pos_new^i = (1 / |{j : |pos_old^j − pos_old^i| < ε}|) · Σ_{j : |pos_old^j − pos_old^i| < ε} pos_old^j

Key findings (base model)

  • No assumption of a "correct" answer. Suitable for opinions where truth is irrelevant (beauty, moral value) or where information is hard to get.
  • Polarization can emerge even when every agent is unbiased and rational — simply because clusters drift outside one another's confidence interval and lose contact.
  • Non-monotonicity (openness monotonicity): changes in ε usually decrease cluster count but can sometimes increase it, and can also move the position of consensus (Duijf & Klein, in progress).

Paper 4 — Douven & Hegselmann (2021): mis- and disinformation in a bounded-confidence model

What the paper adds

The base HK model has no notion of truth. Douven & Hegselmann extend it by:

  1. Introducing a correct state τ. Some opinions are right; others are wrong.
  2. Adding rare signals about τ to the social signals.
  3. Three agent types:
  4. Free Riders — update only from social signals, as in base HK.
  5. Truth Seekers — pull toward τ: pos_new^truthseeking = (1−α)·pos_new^social + α·τ
  6. Campaigners — hold a fixed position ρ regardless of τ.

Misinformation vs. disinformation

  • Misinformation: campaigners aim to make the public believe a falsehood about a topic.
  • Disinformation: campaigners aim to impede or distract the public from believing a truth.
  • Logical relation: for logically consistent agents, misinformation implies disinformation, but not vice versa.

Key counter-intuitive observations

Without influence of truth (no truth-seekers): - Having a more extreme campaigner position hurts the campaigners — they isolate themselves outside others' ε. - Having more / stronger campaigners can impede their own campaign (saturation/repulsion effect).

With truth, no free riders: - Light deviation (ρ close to τ) is more successful than strong deviation — a subtle disinformer beats a bold one. - Now having more campaigners can help — the previous reversal flips back.

With truth + free riders: - Free riders have, if anything, a negative influence on truth-seekers (they pull them off truth, not toward it). - Against truth-seekers, disinformation can increase when free riders are present — but misinformation never can. - Free riders themselves can end up drawn to campaigners, to truth-seekers, or trapped in the middle.

The bigger argument

Iterated social learning combined with campaigning produces qualitatively counter-intuitive effects. Campaigners may help themselves by being few and subtle. These findings are qualitative not quantitative — the model is heavily idealised: - No real underlying social-network structure - No substantive arguments exchanged, just numeric positions - Same ε for everyone - Double-counting issues

These idealisations are why the model is a thinking tool, not a forecasting tool. (See extensions: Wu et al. 2022 add network structure; Lorenz 2008 adds multi-topic discussion; Betz 2023 combines Hegselmann–Krause with LLMs that produce substantive arguments.)

Key concept 4 — Agent-based modelling as an AI tool

Klein's pitch: when a phenomenon has tipping points and path dependencies, analytic (closed-form math) solutions are often intractable and human intuitions are often wrong. ABMs are the AI tool for this regime.

ABM = a set of agents + their interactions, simulated forward in time. Many free, easy-to-use packages exist (NetLogo, Mesa, etc.). Recommended intro: Klein, Marx & Fischbach (2018), Agent-Based Modeling in Social Science, History, and Philosophy: An Introduction, HSR.

Why this matters for an open society

  • Polarization & filter bubbles (Democracy & Good Governance pillar): bounded-confidence dynamics show polarization can emerge without bad actors — pure homophily + selective attention suffices. Then add Douven-Hegselmann campaigners, and you have a formal model of how a small disinformation operation can fracture public opinion more effectively than a loud one.
  • Norms & inequality (Equity & Diversity): emergent patterns are how durable inequality and entrenched norms persist even when individuals do not directly intend them.
  • Crowd safety & "Open Cities": panic buying, evacuation flows, public-space congestion are emergent patterns governed by tipping points — ABMs are the methodological link from Klein's lecture to L6's pedestrian digital twins (Bontje).

Likely essay-question angles

  1. "Explain emergence, path dependency, tipping points, and the four 'directions of effect'. Illustrate each with one example."
  2. "Describe the Hegselmann–Krause bounded-confidence model. How can polarization emerge from unbiased, rational agents? What policy implication does this carry?"
  3. "Douven & Hegselmann distinguish misinformation from disinformation and find that a subtle disinformer often outperforms a loud one. Walk through the mechanism. What does this imply for how we should regulate online influence campaigns?"
  4. "Agent-based modelling has both methodological and ethical advantages over closed-form social theory. Discuss, naming the major limitations of ABM as a forecasting tool."

Quick self-test

  1. Define emergence and state Coleman's bathtub diagram in your own words.
  2. Distinguish path dependence from tipping points; give one example of each.
  3. In the Hegselmann–Krause model, what is ε and how does it determine who an agent listens to? State the update rule.
  4. List the three agent types in Douven & Hegselmann (2021) and what each one does at each update step.
  5. Define misinformation vs disinformation as in the paper. Which logically implies the other?
  6. Two counter-intuitive D&H findings: (a) what does more-extreme positioning do to a campaigner without truth-seekers; (b) what happens when truth-seekers are added?
  7. Why is ABM described as the "AI tool" appropriate to collective patterns?

Source slides

Open AIOS_lecture4_Klein.pdf in new tab ↗