---
title: "Human behaviour and society · Simulations · Simulation Hypothesis"
url: https://simulation-hypothesis.com/simulations/human-society
description: "Decisions, crowds, opinions, markets, past societies and whole populations. Part of a taxonomy of the simulations people actually run."
license: https://creativecommons.org/licenses/by/4.0/
credit: "Ross Allan Dahl, Simulation Hypothesis (simulation-hypothesis.com)"
---

[Simulations](https://simulation-hypothesis.com/simulations)domain 8 of 18

# Human behaviour and society

Decisions, crowds, opinions, markets, past societies and whole populations.

Here people are modelled as decision-makers, crowds, markets and populations. The models are kept simple on purpose. Often the point is to show how little individual intent it takes to produce a large social pattern. Anthropologists and archaeologists use the same tools to rebuild societies that no longer exist. That makes them the closest real analogue of the ancestor simulations in the hypothesis.

## Minds and decisions

### Cognitive architectures

A computational model of human memory, attention and reasoning, run on the same task given to people in the lab.

Question it answers: Does the model make the same mistakes, at the same speed, as people do?

Leaves out: Fitting human reaction times does not prove the model uses the brain's actual mechanism.

Runs on: about 10^12 FLOPS rough a desktop workstation running an architecture on a lab task Since: 1983 Newell's development of a unified cognitive architecture, Soar

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)floor: Rules

### Models of decision-making

Evidence building up at random until it crosses a threshold, used to explain choices and reaction times.

Question it answers: Why do people trade speed for accuracy the way they do?

Leaves out: Assumes evidence builds up at a constant noisy rate, a simplified neural picture.

Runs on: about 10^9 FLOPS rough a laptop fitting a drift-diffusion model to choice data Since: 1978 Ratcliff's drift-diffusion model of decision-making

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Models of learning

Agents learning from reward and error, fitted to how people and animals behave in learning experiments.

Question it answers: What learning rule best explains how people adapt to changing odds?

Leaves out: Captures average behaviour well but often misses individual differences in learning.

Runs on: about 10^10 FLOPS rough a desktop workstation fitting a reward-learning model Since: 1972 Rescorla and Wagner's model of associative learning

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Social cognition and theory-of-mind models

An agent's picture of what another believes, wants or knows, nested inside itself, used to explain how people predict and manipulate each other's behaviour.

Question it answers: How many levels deep do people actually reason about what others are thinking?

Leaves out: Most models cap belief-nesting at a small number of levels, although real reasoning is less regular.

Runs on: about 10^10 FLOPS rough a desktop workstation running a nested belief-reasoning model Since: 1978 Premack and Woodruff's theory of mind, later formalised computationally

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Rules

### Emotion and affect modelling

A simulated appraisal of a situation, such as threat or loss, turned into an emotional state that then biases the agent's later decisions.

Question it answers: How does feeling afraid or angry actually change the choices that follow?

Leaves out: Appraisal rules are the modeller's theory of emotion, not a direct read of anyone's feelings.

Runs on: about 10^10 FLOPS rough a desktop workstation running an appraisal-based emotion model Since: 1988 Ortony, Clore and Collins's cognitive model of emotion

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Rules

### Choice under risk and uncertainty modelling

Gambles and decisions run through a model of how people actually weigh gains, losses and probabilities, rather than how a purely rational agent would.

Question it answers: Why do people's choices under risk depart so consistently from the rational textbook answer?

Leaves out: Fits average biases well, but any one person's choices can depart from the fitted curve.

Runs on: about 10^10 FLOPS rough a desktop workstation fitting a prospect-theory choice model Since: 1992 Tversky and Kahneman's cumulative prospect theory of choice under risk

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Rules

### Language-model survey respondents

A language model is prompted with a demographic profile and asked to answer survey questions as that person would.

Question it answers: Can a model stand in for a survey sample to predict group responses?

Leaves out: Inherits bias and homogeneity from the model's training text, and is weak on groups it saw little of.

Runs on: about 10^14 FLOPS rough One model call per simulated respondent Since: 2022 Out of One, Many: using language models to simulate human samples (2022) [source](https://arxiv.org/abs/2209.06899)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Language-model behavioural experiments

Language models are prompted as many different participants to replay classic experiments in psychology and economics.

Question it answers: Does a model reproduce human results in a classic study?

Leaves out: Replication may reflect memorised studies, and the model may not generalise to new designs.

Runs on: about 10^14 FLOPS rough Many model calls per experiment Since: 2022 Using large language models to simulate multiple humans and replicate human subject studies (2022) [source](https://arxiv.org/abs/2208.10264)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

## Social dynamics

### Residential sorting

Households moving when too few neighbours are like them, Thomas Schelling’s model of how mild preferences produce sharp divides.

Question it answers: How much individual preference does it take to produce a divided city?

Leaves out: Real segregation also involves income, discrimination and history mild preference alone cannot explain.

Runs on: about 10^9 FLOPS rough a laptop running Schelling's segregation model on a grid Since: 1971 Schelling's original hand-run segregation model, later computerised In the simulations tree: [Schelling segregation](https://simulation-hypothesis.com/trees/simulations?n=soc-schelling)

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Run the interactive model](https://simulation-hypothesis.com/simulations/run#toy-schelling)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Bodies

### Opinion dynamics

Agents nudging each other’s views through contact and networks.

Question it answers: When does a group reach consensus, and when does it split into camps?

Leaves out: Real opinions involve emotion and identity that simple averaging rules do not capture.

Runs on: about 10^10 FLOPS rough a desktop workstation updating opinions across a social network Since: 1974 DeGroot's model of iterative opinion pooling, later simulated on networks

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Bodies

### Cooperation and norms

Repeated games of trust and cheating between many players, with reputations and punishment.

Question it answers: How do cooperation and fairness norms become stable?

Leaves out: Real reputations and punishment are messier and slower than clean repeated games.

Runs on: about 10^10 FLOPS rough a desktop workstation running repeated-game tournaments Since: 1980 Axelrod's computer tournaments of the repeated prisoner's dilemma

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Diffusion of ideas

Innovations, habits and rumours spreading through social networks.

Question it answers: Why do some ideas catch on and others never spread?

Leaves out: Predicts adoption curves better than it explains why any one person switches.

Runs on: about 10^10 FLOPS rough a desktop workstation simulating adoption across a social network Since: 1969 Bass's mathematical model of new product diffusion, later simulated

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Social network formation

People making and breaking ties by rules such as friends of friends and similarity.

Question it answers: Why do large networks show short path lengths between almost all pairs of nodes?

Leaves out: Real tie-formation involves geography and institutions abstract rules skip.

Runs on: about 10^10 FLOPS rough a desktop workstation growing a network under an attachment rule Since: 1999 Barabasi and Albert's preferential attachment model of network growth

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Bodies

### Language evolution

Agents inventing words and grammar to communicate, and passing them on.

Question it answers: How can a shared vocabulary arise with nobody in charge?

Leaves out: Real language has grammar and history far richer than the small vocabularies agents evolve.

Runs on: about 10^10 FLOPS rough a desktop workstation running agents inventing a shared vocabulary Since: 1995 Steels's early simulations of agents inventing a shared vocabulary

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Cultural evolution

Skills, tools and traditions copied, improved and lost across generations.

Question it answers: Why do bigger, better-connected populations keep more complex technology?

Leaves out: Treats culture like genes passed between people, missing how ideas get reinvented.

Runs on: about 10^10 FLOPS rough a desktop workstation simulating trait transmission across generations Since: 1981 Cavalli-Sforza and Feldman's models of cultural transmission

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Populations

### Language-model agent societies

Dozens of software agents driven by a language model, each with memory and plans, live in a small simulated town and form social behaviour.

Question it answers: What social patterns arise when agents plan and talk in natural language?

Leaves out: Behaviour comes from the language model's training, and runs are costly and hard to reproduce.

Runs on: about 10^15 FLOPS rough Thousands of model calls per simulated day Since: 2023 Generative Agents: interactive simulacra of human behavior (2023) [source](https://arxiv.org/abs/2304.03442)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Populated social-system prototypes

A language model generates a community of simulated members with posts and replies so designers can test a social platform before launch.

Question it answers: What behaviour might a new community design attract?

Leaves out: Output varies between runs, and bad actors are only as realistic as the prompt.

Runs on: about 10^13 FLOPS rough Many model calls per community Since: 2022 Social Simulacra (2022) [source](https://arxiv.org/abs/2208.04024)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Rules

### Language-model social platform simulation

A platform with feeds, follows and recommendation is populated by up to a million model-driven users to study information spread and herding.

Question it answers: How do posts, opinions and misinformation spread on a platform of model-driven users?

Leaves out: Agents are far more uniform than real users, and real-world validation is limited.

Runs on: about 10^17 FLOPS rough Large-scale model inference Since: 2024 OASIS: open agent social interaction simulations with one million agents (2024) [source](https://arxiv.org/abs/2411.11581)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Rules

### Recommender feedback-loop simulation

Simulated users and a recommender system interact over many rounds to show how recommendations change what users choose.

Question it answers: Do recommendation systems narrow what people consume over time?

Leaves out: User behaviour is a stylised model, not measured preference.

Runs on: about 10^10 FLOPS rough Cheap repeated rounds Since: 2017 How algorithmic confounding in recommendation systems increases homogeneity and decreases utility (2017) [source](https://arxiv.org/abs/1710.11214)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Language-model user simulation for recommenders

Language-model agents with taste profiles browse, rate and quit suggested items to stand in for users in tests of recommendation algorithms.

Question it answers: How would users react to a new recommendation algorithm?

Leaves out: Agent taste is only an approximation of real tastes.

Runs on: about 10^14 FLOPS rough Many model calls per simulated user Since: 2023 On generative agents in recommendation (2023) [source](https://arxiv.org/abs/2310.10108)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Misinformation and rumour spread

People in clusters decide to share a story according to how many of their contacts have shared it, so spread depends on network structure.

Question it answers: What network structure lets a false story go viral?

Leaves out: Platform algorithms and the motives of sharers are simplified.

Runs on: about 10^7 FLOPS rough network contagion on a workstation. Since: 2018 Toernberg, echo chambers and viral misinformation: modelling fake news as complex contagion. [source](https://doi.org/10.1371/journal.pone.0203958)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Populations

### Migration under climate stress

Households decide whether to leave a region based on income, climate shocks and social ties, and the flows between areas are summed.

Question it answers: How many people move if rainfall or farm income falls?

Leaves out: Decision rules are assumed and migration is multi-causal.

Runs on: about 10^8 FLOPS rough regional agent runs on a workstation. Since: 2011 Kniveton and colleagues, agent-based model simulations of future changes in migration flows for Burkina Faso. [source](https://doi.org/10.1016/j.gloenvcha.2011.09.006)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Populations

### Human mobility and migration models

The flow of people between places is predicted from populations and distances, using gravity or radiation-type rules.

Question it answers: How many people travel between two places?

Leaves out: Ignores transport costs, borders and individual differences.

Runs on: about 10^6 FLOPS rough flow matrix calculation on a laptop. Since: 2012 Simini and colleagues, a universal model for mobility and migration patterns. [source](https://doi.org/10.1038/nature10856)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Populations

### Synthetic population generation

A full set of artificial households and individuals is built to match the known totals of a real city or country, giving each person an age, job, home and family.

Question it answers: How can we get a realistic stand-in for everyone in a region without using anyone's real record?

Leaves out: Only the totals the census publishes are matched, so correlations among traits that were not measured are invented.

Runs on: about 10^9 FLOPS rough hours on a workstation for a country. Since: 1996 Beckman, Baggerly and McKay's creation of synthetic baseline populations. [source](https://doi.org/10.1016/0965-8564%2896%2900004-3)

[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Populations

## Crowds

### Pedestrian and crowd dynamics

Each walker steering around obstacles and other people by social forces.

Question it answers: Where do crowds bottleneck in this station, and how do we prevent a crush?

Leaves out: Real crowds panic and behave unpredictably in ways smooth force models miss.

Runs on: about 10^13 FLOPS rough a university cluster simulating tens of thousands of walkers Since: 1995 Helbing and Molnar's social force model of pedestrian movement

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Continuum fields](https://simulation-hypothesis.com/simulations/methods#m-continuum)floor: Bodies

### Crowd emotion and behavioural contagion

A mood or action such as panic, cheering or laughter spreading from person to person through a dense crowd, faster than anyone decided to join in.

Question it answers: How does a feeling or an action sweep through a crowd of strangers?

Leaves out: Contagion rules are usually guessed from video footage after the fact, not measured live.

Runs on: about 10^11 FLOPS rough a desktop workstation simulating mood spread through a crowd Since: 2000 early agent-based models of emotional contagion in crowd panic

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Bodies

### Collective action and tipping-point models

Each person willing to join in only once enough others already have, so a crowd's move from quiet to acting together depends on the mix of individual thresholds.

Question it answers: How many people acting first does it take before everyone else follows?

Leaves out: Assumes everyone's threshold is fixed in advance, when real ones shift once a crowd starts moving.

Runs on: about 10^9 FLOPS rough a laptop running a threshold cascade over a population Since: 1978 Granovetter's threshold model of collective behaviour

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Collective estimation and crowd wisdom

Many independent, individually noisy guesses at a single number, pooled to see how much closer the average lands to the truth than any one person's guess.

Question it answers: When does averaging many guesses beat asking one expert?

Leaves out: The averaging effect fails once individual guesses cease to be independent.

Runs on: about 10^9 FLOPS rough a laptop running a Monte Carlo aggregation of guesses Since: 1907 Francis Galton's analysis of a crowd's guesses at an ox's weight

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Protest mobilisation network models

People deciding whether to join a protest based on how many of their own social ties are already going, spreading participation through a network rather than a crowd in one place.

Question it answers: How does a protest movement grow from a few committed people to a mass turnout?

Leaves out: Treats joining as a single yes-or-no decision, missing the small steps of real commitment.

Runs on: about 10^10 FLOPS rough a desktop workstation simulating recruitment across a network Since: 1997 Kim and Bearman's network model of protest participation

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Mass gathering disease risk modelling

The unusually dense, temporary mixing of people at a festival, pilgrimage or sporting event, modelled to estimate how fast an infection could spread and burn out once the crowd disperses.

Question it answers: How much extra disease risk does bringing this many people together for a few days create?

Leaves out: Once the crowd disperses, tracking where any resulting infections go next is largely out of reach.

Runs on: about 10^12 FLOPS rough a university cluster simulating contacts across one gathering Since: 1987 modelling of a meningitis outbreak among pilgrims at a mass gathering

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Bodies

## Economies and markets

### Macroeconomic models

A whole economy as representative households, firms, banks and a central bank reacting to each other.

Question it answers: What happens to jobs and prices if interest rates rise?

Leaves out: Representative-household models can miss dynamics from inequality.

Runs on: about 10^12 FLOPS rough a desktop or small cluster solving a structural macroeconomic model Since: 1955 Klein and Goldberger's early computerised macroeconometric model

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Populations

### Agent-based economies

Thousands of individual households, firms and banks buying, selling, borrowing and failing.

Question it answers: How do crises start from ordinary behaviour?

Leaves out: Behaviour rules for the agents are chosen by the modeller, not measured from people.

Runs on: about 10^13 FLOPS rough a university cluster simulating thousands of interacting households and firms Since: 1993 early agent-based computational economics models at the Santa Fe Institute

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Financial markets

Traders with different strategies placing orders into a simulated market.

Question it answers: Where do crashes and bubbles come from?

Leaves out: Real markets have regulation, news and rare events toy trading rules leave out.

Runs on: about 10^13 FLOPS rough a university cluster simulating many trading strategies interacting Since: 1994 Palmer, Arthur and colleagues’ Santa Fe artificial stock market [source](https://deepblue.lib.umich.edu/items/18b0197f-305e-40c5-b9bf-98ca56549ca5) In the simulations tree: [Artificial stock market](https://simulation-hypothesis.com/trees/simulations?n=soc-santafe)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Bodies

### Financial risk

Thousands of possible futures for a portfolio, a bank or an insurer.

Question it answers: How bad is a one-in-a-hundred-year loss?

Leaves out: Assumes the future looks statistically like the past, which breaks down in a crisis.

Runs on: about 10^13 FLOPS rough a bank or insurer's cluster running millions of Monte Carlo scenarios Since: 1994 risk management groups' early adoption of Monte Carlo value-at-risk models

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Populations

### Market and auction design

Bidders with strategies competing under different auction rules.

Question it answers: Which set of rules gets the fairest price with the least gaming?

Leaves out: Real bidders do not always behave as rationally as the strategies assume.

Runs on: about 10^10 FLOPS rough a desktop workstation testing bidding strategies under auction rules Since: 1961 Vickrey's auction theory, later tested by simulation and experiment

[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Artificial societies

Agents gathering, trading and reproducing on a landscape of resources, from which wealth distributions and trade emerge.

Question it answers: Which simple rules are enough to grow a society’s large-scale features?

Leaves out: Simple rules can grow realistic-looking patterns for the wrong underlying reason.

Runs on: about 10^10 FLOPS rough a desktop workstation running an agent landscape economy Since: 1996 Epstein and Axtell's Sugarscape model In the simulations tree: [Sugarscape](https://simulation-hypothesis.com/trees/simulations?n=soc-sugarscape)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### AI-designed economic policy simulation

Learning agents act as workers in a simulated economy while a trained planner searches for tax policy that balances equality and productivity.

Question it answers: What tax rule do learning agents and a planner converge to?

Leaves out: Real people differ from learning agents, and economy structure is simple.

Runs on: about 10^15 FLOPS rough Reinforcement learning at scale Since: 2020 The AI Economist (2020) [source](https://arxiv.org/abs/2004.13332)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Agent-based exchange simulators

A simulated stock exchange with a full order book and time-stamped messages where trading agents, including learning ones, interact.

Question it answers: How would a trading strategy or rule change affect a market?

Leaves out: Agent behaviour is stylised, and market impact is hard to calibrate.

Runs on: about 10^11 FLOPS rough Event-driven runs of thousands of agents Since: 2019 ABIDES: towards high-fidelity market simulation for AI research (2019) [source](https://arxiv.org/abs/1904.12066)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)floor: Rules

### Housing market agent-based models

Households with incomes, banks with lending rules and a limited housing stock interact to show how leverage and credit change prices and defaults.

Question it answers: How do lending rules produce housing booms and busts?

Leaves out: Household behaviour is rule-based and the data are of limited detail.

Runs on: about 10^9 FLOPS rough hundreds of thousands of agents on a workstation. Since: 2012 Geanakoplos and colleagues, getting at systemic risk via an agent-based model of the housing market. [source](https://doi.org/10.1257/aer.102.3.53)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Populations

### Dynamic stochastic general equilibrium models

A representative household and firm choose over time under random shocks, and the model is used to trace how such disturbances and policy affect the economy.

Question it answers: How does the economy respond to a shock or a change in interest rate?

Leaves out: Assumes rational representative agents and cannot easily generate crises.

Runs on: about 10^6 FLOPS rough perturbation or Bayesian estimation on a laptop. Since: 1982 Kydland and Prescott, time to build and aggregate fluctuations. [source](https://doi.org/10.2307/1913386)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Insurance catastrophe models

Thousands of simulated storms, quakes or floods are applied to a property portfolio to estimate the losses an insurer might face.

Question it answers: What is the loss exceeded once in 200 years?

Leaves out: Rare extremes rest on short records and the vulnerability of buildings is estimated.

Runs on: about 10^11 FLOPS rough large event catalogues on a cluster. Since: 1972 Friedman, insurance and the natural hazards. [source](https://doi.org/10.1017/s0515036100005699)

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Populations

### Financial contagion in interbank networks

Banks linked by loans are shocked, and the failures that spread from one to the next are traced to measure systemic risk.

Question it answers: How far does one bank failure spread through the system?

Leaves out: Exposures are only partly observed and bank behaviour is fixed.

Runs on: about 10^7 FLOPS rough clearing algorithms on a workstation. Since: 2001 Eisenberg and Noe, systemic risk in financial systems. [source](https://doi.org/10.1287/mnsc.47.2.236.9835)

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Rules

### Kinetic wealth exchange models

Agents trade random amounts of money pairwise like colliding gas particles, and a distribution of wealth emerges.

Question it answers: Do simple random exchanges produce the observed inequality?

Leaves out: Exchange rules ignore production, credit and institutions.

Runs on: about 10^6 FLOPS rough simulation of millions of exchanges on a laptop. Since: 2000 Dragulescu and Yakovenko, statistical mechanics of money. [source](https://doi.org/10.1007/s100510070114)

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Particles and N-body](https://simulation-hypothesis.com/simulations/methods#m-particles)floor: Rules

### Disaster economic ripple models

An input-output model of how industries supply each other is hit by lost capacity after a disaster, and shortages cascade through the supply chain.

Question it answers: What is the total economic cost when a disaster halts some industries?

Leaves out: Fixed technical coefficients overstate rigidity and ignore firms switching suppliers.

Runs on: about 10^6 FLOPS rough matrix models on a laptop. Since: 2008 Hallegatte, an adaptive regional input-output model applied to the economic cost of Katrina. [source](https://doi.org/10.1111/j.1539-6924.2008.01046.x)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)floor: Rules

### Electricity market agent-based models

Generators bid into a market repeatedly, learning from profit, and the price pattern shows how the rules affect outcomes.

Question it answers: Do auction rules lead to lower prices or to collusive bidding?

Leaves out: Learning rules for firms are assumed.

Runs on: about 10^7 FLOPS rough repeated-auction simulation on a workstation. Since: 2001 Bower and Bunn, experimental analysis of the efficiency of uniform-price versus discriminatory auctions. [source](https://doi.org/10.1016/s0165-1889%2800%2900036-1)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)[Evolutionary](https://simulation-hypothesis.com/simulations/methods#m-evolutionary)floor: Populations

### Coupled epidemic and economy models

An epidemic model is linked to an economy model, so that infection reduces work and spending, and economic choices change its spread.

Question it answers: What is the trade-off between lives saved and output lost?

Leaves out: Behaviour is assumed to be rational and uniform.

Runs on: about 10^6 FLOPS rough dynamic optimisation on a laptop. Since: 2020 Eichenbaum, Rebelo and Trabandt, the macroeconomics of epidemics. [source](https://doi.org/10.3386/w26882)

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Populations

### Flash-crash market models

Several types of trading agents, including fast automated sellers and slow fundamental buyers, interact in a model of a futures market to show how a sudden price collapse can build and recover.

Question it answers: Can a large sell order and thin liquidity alone trigger a market-wide price plunge?

Leaves out: The agents' strategies are simple stand-ins for real firms whose algorithms are secret, so the story it tells is plausible rather than verified.

Runs on: about 10^11 FLOPS rough hours on a cluster for ensembles of trading days. Since: 2012 Paddrik and colleagues' agent-based model of the E-Mini S&P 500 futures market applied to the 2010 flash crash. [source](https://doi.org/10.1109/cifer.2012.6327800)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Populations

### Algorithmic pricing collusion

Pricing programs that learn from sales results are placed in a simulated market to see whether they come to charge high prices together without ever talking.

Question it answers: Can independent sellers using learning algorithms come to keep prices above competitive levels?

Leaves out: The markets are tiny, with a few firms and identical goods, and real firms use richer information and design.

Runs on: about 10^9 FLOPS rough millions of repeated rounds on a workstation. Since: 2019 Calvano, Calzolari, Denicolò and Pastorello's study of AI algorithmic pricing and collusion. [source](https://doi.org/10.2139/ssrn.3304991)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Populations

### Language-model macroeconomic agents

Language-model agents act as households and firms in a simulated economy, deciding in words how much to work, spend and save, and overall output and inflation are measured.

Question it answers: Do agents that reason in language reproduce the patterns of real macroeconomies?

Leaves out: The agents' behaviour comes from text they learned and may mirror what people write about economics rather than how they act.

Runs on: about 10^16 FLOPS rough millions of model calls for a hundred agents over decades of simulated time. Since: 2023 Li and colleagues' EconAgent, language-model agents for simulating macroeconomic activities. [source](https://arxiv.org/abs/2310.10436)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Populations

## Anthropology and archaeology

### Reconstructed ancient societies

The farming households of a past society, living on a simulated landscape rebuilt from climate and soil records, such as the ancestral Puebloan villages of the American Southwest.

Question it answers: Why was this valley settled, and why was it abandoned?

Leaves out: Depends on climate and crop-yield reconstructions that carry their own big uncertainties.

Runs on: about 10^12 FLOPS rough a university cluster running centuries of simulated households Since: 2002 Axtell, Epstein, Dean, Gumerman and colleagues’ model of the Kayenta Anasazi in Long House Valley [source](https://profiles.wustl.edu/en/publications/population-growth-and-collapse-in-a-multiagent-model-of-the-kayen/) In the simulations tree: [Artificial Anasazi](https://simulation-hypothesis.com/trees/simulations?n=soc-anasazi)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Human dispersal

Early human populations spreading across continents as climate and coastlines change.

Question it answers: When and by which routes did people reach each continent?

Leaves out: Routes and dates depend on climate and genetic estimates that keep changing.

Runs on: about 10^12 FLOPS rough a university cluster simulating population spread across continents Since: 2008 by 2008, early agent-based simulations of out-of-Africa human dispersal

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Cellular automata](https://simulation-hypothesis.com/simulations/methods#m-cellular-automata)floor: Populations

### Hunter-gatherer economies

Bands foraging, moving camp and sharing food.

Question it answers: Why did sharing food become universal among foragers?

Leaves out: Ethnographic data from a handful of surviving groups stands in for a wider ancient variety.

Runs on: about 10^11 FLOPS rough a desktop workstation simulating a band's foraging and sharing decisions Since: 1994 early agent-based models of forager food-sharing decisions

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Ancient trade and exchange

Settlements linked by paths and sea routes, with goods and ideas moving between them.

Question it answers: How did obsidian, pottery or ideas travel so far?

Leaves out: Surviving artefacts are a thin and biased sample of everything that actually moved.

Runs on: about 10^11 FLOPS rough a desktop workstation modelling exchange across a settlement network Since: 2007 by 2007, early network models of Roman-era trade routes

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Demographic microsimulation

A synthetic population living out births, partnerships, migrations and deaths, year by year.

Question it answers: How will the population age, and what does that mean for pensions and care?

Leaves out: Errors in individual life-event probabilities compound across a simulated lifetime.

Runs on: about 10^13 FLOPS rough a university cluster simulating a synthetic population's life events Since: 1957 Orcutt's original build of a microsimulation model of the American population

Models: [Human](https://simulation-hypothesis.com/entities/human)

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Bodies

### Monument construction and engineering reconstruction

The structural loads, labour and logistics of an ancient monument, tested by simulating candidate construction methods against what the physical remains can and cannot support.

Question it answers: Could this monument actually have been built the way tradition or a leading theory says?

Leaves out: Several competing construction methods often fit the surviving evidence equally well.

Runs on: about 10^11 FLOPS rough a desktop workstation running a structural and labour-logistics model Since: 1997 early simulations testing how Egyptian pyramid construction ramps could have worked

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

## Policy and organisations

### Tax and benefit microsimulation

A proposed change to taxes or benefits applied to a large sample of real households.

Question it answers: Who gains and who loses from this change, and by how much?

Leaves out: Assumes people do not change behaviour in response to the policy, which they sometimes do.

Runs on: about 10^12 FLOPS rough a university cluster applying rules to a large household sample Since: 1976 Orcutt and colleagues' DYNASIM microsimulation of US tax policy

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Rules and states](https://simulation-hypothesis.com/simulations/methods#m-rules)floor: Bodies

### World models

Population, food, industry, resources and pollution as stocks and feedback loops on a global scale.

Question it answers: What happens to a world with finite resources under continued growth?

Leaves out: A handful of global stocks hides all regional and national variation.

Runs on: about 10^10 FLOPS rough a desktop workstation solving a global stocks-and-flows model Since: 1971 Forrester’s World2, published in World Dynamics; World3 for The Limits to Growth followed in 1972 [source](https://en.wikipedia.org/wiki/World3) In the simulations tree: [World3](https://simulation-hypothesis.com/trees/simulations?n=soc-world3)

[System dynamics](https://simulation-hypothesis.com/simulations/methods#m-system-dynamics)floor: Populations

### Organisations and teams

Members, problems and decisions moving through an organisation.

Question it answers: Why do organisations make the decisions they do?

Leaves out: Real organisational politics and personality are hard to reduce to decision rules.

Runs on: about 10^10 FLOPS rough a desktop workstation simulating decisions moving through an organisation Since: 1963 Cyert and March's simulation models of organisational decision-making

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Bodies

### Computable general equilibrium modelling

A whole economy's industries, households and government rebuilt as a system of markets that must all clear at once, then shocked with a tax or trade policy to see where prices and output settle.

Question it answers: How does a whole economy rebalance itself after this policy change?

Leaves out: Assumes markets settle into a new equilibrium cleanly, neglecting the disorderly transition in practice.

Runs on: about 10^11 FLOPS rough a small cluster solving a large system of market-clearing equations Since: 1972 Shoven and Whalley's applied general-equilibrium model of tax policy

[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Populations

### Institutional corruption and integrity modelling

Officials and citizens deciding whether to bribe, report or comply, each responding to how common corruption already is and how likely getting caught seems.

Question it answers: Why does corruption stay rare in some institutions and become the norm in others?

Leaves out: Real corruption is hidden by design, so the numbers feeding the model are rough estimates.

Runs on: about 10^10 FLOPS rough a desktop workstation running an agent-based compliance model Since: 2002 early agent-based models of how corruption spreads within an institution

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Rules

### Policy diffusion network models

One government's adoption of a law or programme spreading to others through the ties between them, whether neighbours, trading partners or shared professional networks.

Question it answers: Why do some governments copy a policy quickly, and others never adopt it at all?

Leaves out: Copying a policy on paper is easier to model than whether it works the same way elsewhere.

Runs on: about 10^10 FLOPS rough a desktop workstation simulating adoption across a network of governments Since: 1990 by 1990, early network models of policy diffusion between governments

[Network models](https://simulation-hypothesis.com/simulations/methods#m-network)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Populations

### Language-model agent organisations

Language-model agents take on roles such as manager, programmer and tester and collaborate to produce software.

Question it answers: Can a team of role-playing agents carry out a multi-step organisational task?

Leaves out: Output depends on prompts, and real human organisational friction is absent.

Runs on: about 10^14 FLOPS rough Many model calls per project Since: 2023 ChatDev: communicative agents for software development (2023) [source](https://arxiv.org/abs/2307.07924)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)floor: Rules

### Language-grounded negotiation agents

Agents combine a language model with strategic planning to negotiate and form alliances in a multi-player game by talking to each other.

Question it answers: Can an agent persuade and plan with humans in an open-ended negotiation?

Leaves out: Only one game's rules, and it cannot reason about long-term trust.

Runs on: about 10^16 FLOPS rough Large training plus planning at play time Since: 2022 Human-level play in the game of Diplomacy by combining language models with strategic reasoning (2022) [source](https://doi.org/10.1126/science.ade9097)

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Rules

### Opioid epidemic system dynamics

Stocks of prescription users, misusers, heroin users and the dead are linked by feedback and policy levers such as supply limits and treatment.

Question it answers: Which policy mix reduces opioid deaths?

Leaves out: Illicit supply changes faster than models adapt.

Runs on: about 10^5 FLOPS rough stock-and-flow simulation on a laptop. Since: 2013 Wakeland and colleagues, modelling the impact of simulated educational interventions on pharmaceutical opioid use. [source](https://doi.org/10.1177/1090198113492767)

[System dynamics](https://simulation-hypothesis.com/simulations/methods#m-system-dynamics)floor: Populations

### Exploratory modelling under deep uncertainty

A model is run across thousands of combinations of unknown inputs and policies, to find strategies that are robust rather than optimal for a single forecast.

Question it answers: Which policy is acceptable across the widest range of futures?

Leaves out: The range of futures explored is still chosen by the analyst.

Runs on: about 10^10 FLOPS rough ensembles of runs of a base model on a cluster. Since: 1993 Bankes, exploratory modelling for policy analysis. [source](https://doi.org/10.1287/opre.41.3.435)

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Rules

### Supply-chain dynamics and the bullwhip effect

Orders and stock flow through factory, distributor and retailer with delays, and small changes in demand grow into large swings upstream.

Question it answers: Why do small demand changes cause huge order swings upstream?

Leaves out: Real firms respond to information that the model leaves out.

Runs on: about 10^5 FLOPS rough stock-and-flow simulation on a laptop. Since: 1958 Forrester, industrial dynamics: a major breakthrough for decision makers. [source](https://doi.org/10.1007/978-3-642-27922-5%5F13)

[System dynamics](https://simulation-hypothesis.com/simulations/methods#m-system-dynamics)[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)floor: Rules

## Law, justice and elections

### Electoral and voting system simulation

The same set of voters and candidates run through different voting rules, such as first-past-the-post, ranked choice or proportional systems, to compare who would actually win.

Question it answers: Would a different voting system have changed who won this election?

Leaves out: Assumes voters' preferences stay fixed even though a different system would change how people vote.

Runs on: about 10^10 FLOPS rough a desktop workstation running elections under alternative voting rules Since: 1975 early computer simulations comparing the outcomes of different voting rules

[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)floor: Rules

### Court and case-flow simulation

Cases arriving, waiting for hearings and moving through stages of a court system, used to see where a backlog builds and what more judges or courtrooms would do to it.

Question it answers: How long will it take to clear this court's backlog, and what would actually speed it up?

Leaves out: Real case complexity and human discretion vary more than a queueing model assumes.

Runs on: about 10^10 FLOPS rough a desktop workstation running a discrete-event model of case flow Since: 1976 early discrete-event simulations of US court case backlogs

[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Rules

### Sentencing and recidivism risk modelling

A person's recorded history run through a statistical model that scores how likely they are to reoffend, used to inform parole, bail or sentencing decisions.

Question it answers: How likely is this person to reoffend, based on people with a similar record?

Leaves out: A score built from group statistics can be wrong for the one person it is applied to, in either direction.

Runs on: about 10^11 FLOPS rough a desktop workstation fitting a risk model to case records Since: 1928 Burgess's actuarial scale predicting parole success, an early risk model

[Learned emulators](https://simulation-hypothesis.com/simulations/methods#m-learned)[Monte Carlo](https://simulation-hypothesis.com/simulations/methods#m-monte-carlo)floor: Rules

### Policing and patrol deployment simulation

Patrol units and reported incidents across a city, used to test how different ways of assigning officers to areas change response times and coverage.

Question it answers: Where should patrols be sent to cut response times the most?

Leaves out: Historical crime data can embed past biases in where officers were deployed.

Runs on: about 10^10 FLOPS rough a desktop workstation running a queueing model of patrol deployment Since: 1972 Larson's hypercube queuing model for police patrol deployment

[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)[Discrete event](https://simulation-hypothesis.com/simulations/methods#m-discrete-event)floor: Bodies

### Legislative bargaining and coalition simulation

Parties with positions on a policy space negotiating and forming coalitions to pass a bill or form a government, using rules for who proposes and who must agree.

Question it answers: Which coalition will actually form, and what compromise will it settle on?

Leaves out: Real coalitions form around personalities and history that a spatial model leaves out.

Runs on: about 10^10 FLOPS rough a desktop workstation solving a spatial coalition-bargaining model Since: 1980 early simulations of spatial voting and coalition formation in parliaments

[Game theory](https://simulation-hypothesis.com/simulations/methods#m-game-theory)[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)floor: Rules

### Jury decision-making simulation

Individual jurors with different initial views, deliberating under rules such as majority or unanimity, to see what verdict and how much persuasion the group ends up producing.

Question it answers: Does the verdict reflect the strength of the evidence, or only the dynamics of the deliberation?

Leaves out: Lab juries deciding mock cases may not behave like real juries facing real stakes.

Runs on: about 10^9 FLOPS rough a laptop running a social decision-scheme deliberation model Since: 1983 Davis and colleagues' social decision-scheme models of jury deliberation

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Stochastic processes](https://simulation-hypothesis.com/simulations/methods#m-stochastic)floor: Rules

### Crime hotspot dynamics

Offenders move through a city and are drawn to places with past crime, producing hotspots that grow, move and respond to policing.

Question it answers: How do hotspots form and can patrols move them?

Leaves out: Offender motives are simplified and crime reporting is imperfect.

Runs on: about 10^7 FLOPS rough agent-based city runs on a workstation. Since: 2008 Short and colleagues, a statistical model of criminal behaviour. [source](https://doi.org/10.1142/s0218202508003029)

[Agent-based](https://simulation-hypothesis.com/simulations/methods#m-agent-based)[Differential equations](https://simulation-hypothesis.com/simulations/methods#m-equations)floor: Populations

[◂ Agriculture and food](https://simulation-hypothesis.com/simulations/agriculture-food)[Engineering and technology ▸](https://simulation-hypothesis.com/simulations/engineering)