Simulating Society

Social simulation is the use of computational agents and environments to study how individual decisions can produce collective outcomes. Instead of asking a model for one confident prediction, a simulation represents actors, relationships, information flows, and rules of interaction, then observes how different futures unfold across repeated runs. Modern systems increasingly use language models to make those agents more expressive—but believable dialogue is not the same thing as validated social behaviour.

What social simulation means

A social system contains actors who observe only part of the world, respond to other actors, and change the environment through their actions. A market narrative, election campaign, product launch, online controversy, public-health intervention, or organisational change can all be treated this way. The important feature is interaction: one person's action becomes part of somebody else's context.

Traditional agent-based models encode behaviour with explicit rules. An agent may adopt a belief when enough neighbours adopt it, buy when a threshold is crossed, or share information with a fixed probability. These models are transparent and computationally efficient, but the behavioural rules can be too rigid for settings where language, memory, identity, and interpretation matter.

Generative social simulations add language-model agents that can interpret events, retrieve memories, plan, communicate, and choose among actions. The influential Generative Agents study placed 25 agents in a small interactive town and showed that observation, planning, and reflection could support emergent coordination. That work demonstrated a compelling interface and architecture. It did not establish that any fluent agent population is a faithful model of a real one.

The components of a social simulation

The output is only as meaningful as the system beneath it. A serious design makes at least six components explicit.

1. Scenario and boundary. Define the event, time horizon, observable outcomes, and variables that remain fixed. “How will people react?” is not a simulation contract. “How does information adoption change when a trusted actor responds before the wider crowd?” is closer to one.
2. Population. Decide which named actors, institutions, communities, and representative cohorts matter. Population weights must survive compression: one synthetic retail agent should not accidentally count the same as one regulator or institution.
3. Relationships and exposure. Model who can observe whom, where information travels, and which ties carry trust or influence. A globally visible feed produces a very different world from a networked feed with partial exposure.
4. Behaviour. Agents need a decision process that includes action, response, amplification, and non-action. The simulation should distinguish deciding what to do from rendering how that decision is expressed.
5. Environment and time. State changes over rounds. Posts alter attention; reactions alter sentiment; external events arrive; memories and commitments accumulate. Old state must not be silently rewritten as the world evolves.
6. Measurement. Raw action traces, participation, network diffusion, sentiment, event timing, and end states should be recorded before a narrative summary is generated. Otherwise the evaluator is interpreting a story without an auditable ledger.

This is closely related to the distinction between attention and tone in our research on crypto mindshare and stock sentiment. A social system cannot be reduced to one number. Who acted, who saw the action, and when the response occurred are often as important as whether the language was positive or negative.

Why simulations should run many worlds

One generated trajectory is easy to overinterpret. It may be internally coherent because the same model created both the decisions and the explanation. It may also depend on activation order, sampling temperature, a small trait perturbation, or one early interaction. A single run cannot tell us which events are robust and which are accidents of the seed.

A better approach is Monte Carlo agent simulation: hold the scenario and population definition fixed, vary justified sources of uncertainty, and run multiple worlds. Each world produces an action trace and a structured scorecard. Cross-world analysis can then measure event frequency, outcome distributions, timing variance, and conditional patterns.

The result should look less like “the future will be negative” and more like this: a competitor responded critically in seven of ten worlds; early criticism was associated with a negative end state in five of five worlds; two worlds produced a rapid adoption cascade; and three worlds remained quiet. Frequency is not automatically real-world probability, but it exposes which conclusions depend on fragile branches.

A rigorous simulation workflow

1. Ground the initial state

Start from time-stamped evidence available at the simulation boundary. Preserve source items, authors, timestamps, entities, and the method used to select them. If a feed is reconstructed rather than observed, name it a proxy. Later information must not leak backward into the initial world.

2. Build the population and network

Use real actors where individual fidelity matters and representative cohorts where scale matters. Cohorts should be derived from observed distributions rather than a list of dramatic stereotypes. Influence, activity, reaction speed, and stance can vary within a cohort; the network should encode plausible channels of exposure.

3. Simulate decisions and preserve traces

At every step, an agent observes its available context and chooses whether to act. Non-action must be available, because real social systems are mostly quiet most of the time. Every action, target, timestamp, and state update belongs in an append-only trace that can be replayed.

4. Analyse before summarising

Compute participation, volume, sentiment, diffusion, cohort differences, and turning points programmatically where possible. Use language models for bounded classification and synthesis, not for counts the system can calculate exactly. Keep the evidence behind every extracted event.

5. Aggregate across worlds

Compare trajectories rather than selecting the most interesting one. Report common outcomes, contested outcomes, edge cases, and the early conditions associated with divergence. A useful report makes uncertainty visible instead of averaging it into a bland central story.

How social simulations should be validated

Believability is a user-experience metric, not sufficient scientific validation. A simulation can sound convincing while choosing the wrong actors, overproducing activity, or reproducing the model's general stereotypes. Validation should happen at several levels.

• Micro validity: do individual agents choose actions resembling held-out behaviour under comparable context?
• Relational validity: does information travel through plausible ties, and do influence and exposure operate as intended?
• Macro validity: do aggregate participation, diffusion, polarisation, and timing resemble known data or accepted stylised facts?
• Counterfactual validity: do controlled changes produce sensible directional effects, and are those effects stable across seeds and model versions?
• Sensitivity: does a small arbitrary prompt change overturn the conclusion? If it does, the result should be labelled fragile.

The 2024 study LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals offers one useful benchmark design: compare simulated answers with held-out responses and with each participant's own test-retest consistency. Its results show that rich interviews and surveys can produce meaningful predictive signal. They also show why simulation claims need an explicit human reliability ceiling rather than a vague claim of realism.

Common failure modes

Persona theatre is the most visible failure. A detailed biography creates the impression of depth, but it may encode stereotypes and unsupported traits. Authentic behaviour and context are stronger evidence than decorative psychographics. Our companion guide to behavioural simulation explains this actor-level problem in detail.

Activity bias occurs when every agent is prompted to contribute. Real actors ignore most stimuli. A world in which everyone has a polished opinion every round is a panel discussion, not a social system. Model monoculture is related: agents may have different names but share one underlying model's priors, verbosity, and reasoning habits.

Evaluation circularity occurs when a language model generates the world, judges the world, and writes the final causal account. Programmatic measurements, independent baselines, human audits, and held-out historical replays reduce that risk. Finally, simulations must not be used to launder invented events into evidence. Synthetic scenarios are useful for stress tests; they are not observed ground truth.

Interventions and counterfactuals

The most valuable simulation questions usually compare worlds. What changes when a trusted intermediary responds early, a platform limits redistribution, an institution releases more evidence, or one network bridge disappears? The intervention should be precise enough that the baseline and treatment worlds differ in one documented way. If the population, prompts, model, and information supply all change together, the comparison cannot identify a mechanism.

Randomness should be controlled as well as varied. Paired runs can use the same initial seed and actor population on either side of an intervention, reducing noise from unrelated differences. Other runs should vary seeds deliberately to test whether the effect survives plausible paths. Report both: the average difference across worlds and the range of trajectories that produced it.

These experiments do not automatically establish real-world causality. The result is conditional on the modelled actors, exposure network, behavioural rules, and environment. A simulation can demonstrate that a mechanism is sufficient to create an outcome inside the model; historical or experimental evidence is still needed to show that the mechanism operates with the same strength outside it. Every intervention report should place assumptions and sensitivity tests next to its headline result.

What social simulation is useful for

Social simulations are strongest when the question is conditional and mechanism-oriented: which actors create a cascade, which interventions reduce spread, what produces divergence, where coordination breaks, and which outcomes remain stable across assumptions. They can help with launch rehearsal, policy exploration, market-narrative research, crisis planning, organisational design, and the discovery of scenarios that deserve closer human analysis and more direct empirical investigation.

They are weakest when presented as precise forecasts of one open-ended future, especially when the population or information environment is poorly observed. For market work, simulations should complement live evidence such as our stock attention tools, crypto market sentiment, and continuing research library—not replace primary data or decision-makers.

Sources and further reading

• Park et al.: Generative Agents—Interactive Simulacra of Human Behavior
• Park et al.: LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
• Li et al.: BehaviorChain benchmark for continuous human behaviour simulation
• Schwager et al.: operational validity in simulated social-media users