The Economist Who Ran a Galaxy
In June 2007, an Icelandic game studio issued a press release that read like satire: CCP Games had appointed a lead economist for EVE Online, a science-fiction world of warring spaceship corporations. Dr. Eyjólfur Guðmundsson left his post as a dean at the University of Akureyri to publish quarterly reports on inflation, growth, and price trends, for a galaxy that exists only on servers (CCP Games, 2007). He later described the job as being like “a research scientist for a central bank,” one who occasionally had to intervene to prevent inflation and market dislocations, particularly because EVE’s PLEX tokens tie the in-game currency to real money (Fast Company, 2014).
The hire looked like a curiosity. It was actually a data point in a trend that had been visible since 2001, when the economist Edward Castronova did something no one had thought to do: he calculated the gross national product of a fantasy world. Using auction data from EverQuest’s player markets, he estimated that Norrath, a place made of code, produced roughly $135 million in value per year, or $2,266 per capita, placing it between Russia and Bulgaria and above China and India at the time. The average player “earned” about $3.42 an hour, and a unit of Norrath’s currency traded above the yen and the lira (Castronova, 2001). The paper became the most-downloaded article in the history of SSRN’s economics network. The message was not that games are amusing. It was that virtual worlds generate real economic phenomena (scarcity, prices, wages, trade, inflation, inequality) under conditions of near-total data visibility.
For twenty years, the economists went to the games. The proposition this essay assesses is the inversion of that: deliberately building game-like worlds, populated by increasingly capable AI agents, designed from the start as instruments for economic research. The analogy that fits best is the flight simulator. Simulators did not replace flying; they made it possible to rehearse engine failures and windshear, events too dangerous, expensive, or rare to practice live. Macroeconomic policy has exactly this problem. There is no control group for a national economy, no rewind button on an interest-rate decision. An honest assessment of AI-driven virtual economies has to answer four questions: what did the accidental laboratories already teach us, what specifically does AI change, what does methodological rigor require, and where does the approach genuinely fall short?
What the Accidental Laboratories Already Taught Us
Before anyone built an economic simulation on purpose, the game industry built several by accident. Their track record is the empirical foundation for everything that follows.
Central banking for a galaxy
CCP’s economics program did not end with a hire; it became an institution. To this day the company publishes Monthly Economic Reports containing the money supply, the “faucets” that create the in-game currency (ISK) and the “sinks” that destroy it, the velocity of money, and price indices for minerals, ships, and modules, with the raw data downloadable by anyone (CCP Games, 2026). This is central-bank-grade disclosure for a video game, and players treat it accordingly: when the money supply roughly doubled over four and a half years, players marked up the charts and argued publicly about whether the developer needed tighter monetary policy. CCP has also run genuine policy experiments, most famously the “scarcity era” beginning in 2020, when it deliberately restricted resource supply to drain years of accumulated oversupply. The results were textbook: sharp mineral and ship price inflation, redistribution across professions, and a multi-year rebalancing effort. And the institution is expanding rather than fading: in 2025, CCP hired Stefán Þórarinsson, formerly of Iceland’s central bank, as head of economy for its next title, EVE Frontier, with the stated ambition of building an open financial system inside a virtual world (Game Developer, 2025).
A balance-of-payments problem between two games
In June 2012, Valve hired the Greek economist Yanis Varoufakis. The recruiting email from CEO Gabe Newell described the problem plainly: Valve wanted to link the economies of two virtual environments under a shared currency and was wrestling with the resulting balance-of-payments issues, which is why Newell wanted an expert on what happened to Germany and Greece inside the euro (Washington Post, 2012). Varoufakis’s first study treated Team Fortress 2 as a “peculiarly sophisticated barter economy” and made a methodological point that deserves more attention than it gets: because the platform records every transaction, he could construct an index of arbitrage potential and directly measure how far the economy sat from equilibrium, something economists studying the analog world can only estimate (Varoufakis, 2012). He also found that barter stubbornly persisted even as trading volume exploded, a wrinkle in the textbook story that money must always emerge. Three years later, the same man was Greece’s finance minister, managing a balance-of-payments crisis that was anything but virtual.
A market-design failure in fast-forward
In 2012, Blizzard launched a real-money auction house inside Diablo III, creating an efficient market for the game’s loot. The efficiency was the problem. Buying an item outperformed playing for it, and the game’s production director eventually conceded that the auction house had short-circuited the core reward loop, “kill monsters to get cool loot,” by making the market the optimal path to power (Blizzard Entertainment, 2013). The auction house closed on March 18, 2014. The general lesson is one mechanism designers know but rarely get to observe so cleanly: an efficient market can destroy the incentive structure that gives its goods value. A design error, its measurable consequences, and its correction played out in about eighteen months, at zero social cost.
A financial crisis in miniature, one year early
Ginko Financial, an in-world “bank,” offered depositors interest that at one point reached 0.145% per day, roughly 60–70% annualized, funded by undisclosed investments, many tied to in-world casinos. When Linden Lab banned gambling in 2007, Ginko’s asset base cratered, depositors ran, and the bank collapsed in August 2007 owing about 200 million Linden dollars, roughly $750,000 in real money, with remaining balances force-converted into “perpetual bonds” (TechCrunch, 2008). After further bank failures, Linden Lab prohibited interest-bearing accounts from any entity without a real-world banking charter, effective January 22, 2008 (Linden Lab, 2008). Read the sequence back: unregulated shadow banking, yield-chasing, contagion from an adjacent policy shock, a bank run, an improvised resolution, and new prudential regulation, the full arc of a financial crisis, compressed into eighteen months, concluding almost a year before Lehman Brothers failed.
When an epidemic escaped its dungeon
And not only economics. In 2005, World of Warcraft’s “Corrupted Blood” plague escaped its intended dungeon via player pets and fast travel, infecting whole cities. Epidemiologists later argued in The Lancet Infectious Diseases that the incident captured behavioral dynamics (curiosity-driven exposure, altruistic healers acting as vectors, deliberate spreaders) that standard epidemic models omit entirely (Lofgren & Fefferman, 2007). Virtual worlds were already functioning as behavioral observatories for disciplines beyond economics.
The honest caveat is that all of these were natural experiments: uncontrolled, unreplicated, confounded by software patches and community quirks, and run for entertainment rather than inference. They prove that virtual economies generate real economic phenomena and real policy problems. They do not, by themselves, constitute a laboratory. Closing that gap requires building such worlds deliberately, which raises the question of what to populate them with.
The Missing Ingredient Was Always the Agents
Economics has a mature tradition of designed simulation: agent-based modeling (ABM). The Santa Fe Institute’s artificial stock market showed in the 1990s that markets of heterogeneous, adaptive traders endogenously generate bubbles, crashes, and fat-tailed returns, phenomena that rational-expectations models struggle to produce at all (Arthur et al., 1997). Leigh Tesfatsion’s agent-based computational economics program systematized the approach (Tesfatsion, 2006). After 2008 exposed the fragility of equilibrium models under stress, Farmer and Foley argued in Nature that the economy needs agent-based modeling precisely because its assumptions do not fail in a crisis (Farmer & Foley, 2009). Institutions responded in places: the Bank of England built an agent-based model of the UK housing market to study macroprudential tools such as loan-to-value caps (Baptista et al., 2016), and Axtell and Farmer’s recent survey in the Journal of Economic Literature documents a field that has grown from provocation to established method (Axtell & Farmer, 2025).
But ABM has always carried one standing weakness, and its critics have never let it forget: the agents’ behavior is hand-written. A modeler encodes rules (how agents consume, when they panic, what they imitate) and skeptics can reasonably ask whether the interesting conclusions were smuggled in through those behavioral assumptions. Calibrating scripted rules against the richness of human decision-making is notoriously hard. What agent-based modeling lacked, for forty years, was a defensible model of the decision-maker.
That is precisely the thing that just changed.
What AI Actually Changes: Three Layers of Evidence
The claim that AI transforms economic simulation is no longer speculative. It rests on peer-reviewed test cases at three distinct layers.
Reinforcement learners that discover policy. The AI Economist recovers optimal-tax theory, then improves on it, while its workers invent tax-gaming nobody coded.
Language models as computational stand-ins for people. A thousand agents reproduce their own interviewees at 85% fidelity; economies manifest the Phillips curve and Okun’s law unprompted.
Whole societies left to their own devices. Specialization, taxation, democratic voting, herding, and market crashes emerge without being written in.
Layer one: agents that learn. The AI Economist (Zheng et al., 2022, Science Advances) is the cleanest existing demonstration of machine-designed policy. In a two-level deep reinforcement learning framework, a population of worker agents learns to gather, trade, and build, developing labor-supply behavior, while a planner agent simultaneously learns a tax schedule to maximize social welfare. The validation step matters most: in simplified economies that satisfy the assumptions of optimal-tax theory, the learned schedule converges on the analytical solution derived from Emmanuel Saez’s framework. The simulation recovers known theory before it is trusted with anything new. In richer spatial economies where the theory’s assumptions break, the learned policy then improves the equality-productivity tradeoff by roughly 16% over the Saez-derived baseline, and, notably, the worker agents spontaneously discovered tax-gaming strategies nobody programmed. Two points deserve emphasis. First, this is a “policy feedback loop” (reinforcement learning at the regulator level, tuning parameters against macro objectives) realized and published, not hypothesized. Second, the environment is a stylized grid-world, not an actual economy; the result is a proof of method, not a policy recommendation, and the authors are careful to say so.
Layer two: agents that resemble us. Reinforcement learners optimize; they do not necessarily behave like people. Large language models changed the terms of that problem. Horton, Filippas, and Manning argue that LLMs, by virtue of being trained on the recorded outputs of human judgment, function as implicit computational models of humans, homo silicus, that can be endowed with preferences, information, and budgets and dropped into scenarios exactly the way theorists use homo economicus (Horton et al., 2023). Their replications of canonical experiments (Charness and Rabin’s social-preference games, Kahneman, Knetsch, and Thaler’s fairness judgments, Samuelson and Zeckhauser’s status quo bias) produce qualitatively similar results to the human originals, and the deviations tend to be interesting rather than random. The practical upshot: simulated subjects as cheap pilot studies, letting researchers search the design space before spending money and participant goodwill on human experiments.
The Stanford generative-agents line pushed this from individual subjects to populations. The first paper gave 25 LLM agents memory, planning, and reflection inside a small simulated town and observed believable emergent coordination (Park et al., 2023). The 2024 follow-up is the one that should change research practice: the team conducted two-hour qualitative interviews with 1,052 real, consenting participants, built an agent from each interview, and then tested the agents against their human counterparts. The agents reproduced their individuals’ General Social Survey answers at 85% of the accuracy with which the participants replicated their own answers two weeks later, performed comparably on Big Five personality measures and canonical economic games (dictator, trust, public goods), and, critically, showed smaller accuracy gaps across racial and ideological groups than agents built from demographic descriptions (Park et al., 2024). That reframes what “simulated people” can mean: not stereotype-driven personas, but measured, individually grounded models built inside a consent framework.
Macro-level evidence followed the same pattern. EconAgent placed 100 LLM-driven agents in a simulated economy for 240 months of work and consumption decisions. The resulting aggregates stayed within plausible real-world ranges, and, without being instructed to, the simulation manifested the Phillips curve (correlation −0.62) and Okun’s law (correlation −0.92), the two regularities most commonly used to validate macroeconomic simulations. The rule-based baseline produced a Phillips relationship with the wrong sign (Li et al., 2024). The regularities emerged from decisions; they were not wired in. Subsequent frameworks extended the result: SimCity models households, firms, a central bank, and a government as LLM agents and recovers the Beveridge curve and Engel’s law (arXiv:2510.01297), while a 2025 replication study using the Shachi framework re-ran EconAgent across multiple LLM backends and found that the Phillips and Okun regularities survive the model swap, though slopes and intercepts shift, a sensitivity any serious user of these tools must report (arXiv:2509.21862).
Layer three: whole societies at scale. Project Sid ran 10 to 1,000+ LLM-driven agents in persistent Minecraft worlds and observed spontaneous economic structure: agents specialized into farmers, guards, artists, and traders without role assignments. In a governance experiment, agent societies obeyed a tax law requiring deposit of 20% of inventory, then amended it through democratic voting after exposure to pro- and anti-tax influencer agents, with individual tax compliance tracking the amended law in both directions (AL et al., 2024). TwinMarket simulated 1,000 LLM investors in a socially networked financial market and reproduced the stylized facts of real markets (fat-tailed returns, volatility clustering, herding) from individual behavior grounded in behavioral-finance theory (Yang et al., 2025). AgentSociety scaled the paradigm to study polarization and universal-basic-income-style interventions across large simulated populations (Piao et al., 2025). And the flight-simulator property the analogy promises is now concrete: EconAgent’s 240 simulated months run in days. Time compression, longitudinal observation, total data transparency (the properties the original argument for this methodology predicted) are demonstrated, not projected.
Doing It Rigorously: A Methodological Standard
If this is to become an instrument rather than a demo genre, the methodology needs the same discipline as any measurement device. Four requirements stand out.
Ground the agents in data, not vibes. The Park et al. result is the template: agent behavior anchored in structured elicitation from real, consenting individuals, with fidelity measured against those same individuals. Where individual grounding is impractical, calibration should target measured behavioral distributions and historical microdata, and the calibration step must be reported as carefully as the findings.
Use hybrid architectures deliberately. The evidence favors composition: LLMs where context-sensitive, heterogeneous judgment matters; reinforcement learning where optimization pressure is the point (the AI Economist’s planner); explicit behavioral-economics scaffolding (loss aversion, bounded rationality, present bias) where the empirical record demands features that language models express unreliably; and memory and planning modules of the kind the generative-agents architecture introduced. Frameworks such as Concordia now exist specifically to make this “generative agent-based modeling” systematic (Vezhnevets et al., 2023).
Exploit the one advantage reality can never offer. A simulated economy can be forked. Identical worlds, identical histories, one policy changed: treatment and control at the level of an entire economy, with replication across seeds essentially free. Interventions map cleanly onto the practitioner’s toolkit that game economists already use: currency faucets and sinks are open-market operations; EVE’s dynamic bounty adjustments are automatic stabilizers. Simulation experiments should be pre-registered like any other experiment, with metrics (price stability, output growth, Gini coefficients, surplus measures, volatility) specified in advance.
Validate in a hierarchy, and report fragility. The credibility ladder runs: (1) reproduce known stylized facts (Phillips, Okun, Beveridge, fat tails) without hard-coding them; (2) retrodict historical episodes out of sample; (3) demonstrate robustness across model backends, random seeds, and prompt paraphrases: the Shachi finding that regularities survive but coefficients shift is exactly the kind of sensitivity that belongs in every results table; (4) benchmark side-by-side against human subjects on a subsample, in the Horton manner.
A result that exists in one model version at one temperature is not a finding. It is an anecdote with a confidence interval of zero.
Limitations That Deserve Respect
The case for this methodology is strongest when its weaknesses are stated plainly, because several of them are structural.
Transferability, with the Lucas critique attached. Agents distilled from records of past human behavior may not adapt to genuinely novel policy regimes the way living humans do, which is precisely Lucas’s objection to naive econometric policy evaluation, now applicable to simulation. Results from virtual economies are hypothesis generators and stress tests. They are not forecasts, and any use of them as forecasts should be treated as a category error.
Contamination. Language models have read the textbooks. An LLM agent may “replicate” the Phillips curve in part because descriptions of the Phillips curve are in its training data. Partial mitigations exist (novel and rephrased scenarios, held-out experimental designs, testing against studies published after a model’s training cutoff), but the problem is only partially solvable, and papers in this literature should say so more often than they do.
Homogenization and bias. LLMs over-produce modal answers and under-produce the variance that makes populations interesting; their training corpora over-represent English-speaking, educated, online populations. Interview grounding demonstrably reduces group-level accuracy gaps (Park et al., 2024), but reduction is not elimination, and a simulation that flattens human heterogeneity will flatten exactly the distributional questions (who gains, who loses) that policy analysis exists to answer.
Fragility as a reproducibility hazard. Results shift across model versions, prompts, and sampling temperatures. This is a failure mode traditional ABM never had: the behavioral core of the model can silently change when a vendor ships an update. Version pinning, multi-model replication, and full prompt disclosure should be table stakes for publication.
Goodhart’s law and performativity. A policy optimized against simulated agents can overfit the simulator. The AI Economist’s tax-gaming workers are charming inside the simulation and a warning outside it: real populations will discover exploits the simulation never surfaced, and a policy tuned to simulated compliance may be tuned to nothing at all.
Cost, access, and ethics. Large-scale LLM simulations are computationally expensive, concentrating the capability in well-resourced institutions. The data question cuts deeper: the defensible model is Park’s (consented, compensated participants with withdrawal rights), not scraped behavioral profiles. And the machinery is dual-use. The same simulation that stress-tests a tax reform can optimize manipulative pricing or engagement traps. Methods this powerful import the ethical obligations of human-subjects research even when no human is directly in the loop.
Where This Is Heading
Four near-term directions follow from the evidence, plus one inversion that changes the stakes.
Hybrid human-AI experiments. The natural next step is mixed markets: human participants trading with and against grounded agent populations inside classic experimental-economics paradigms such as the double auction. This preserves human ground truth while gaining the scale and repeatability of simulation, and it converts the venerable experimental-economics laboratory into something closer to a wind tunnel.
Institutional adoption. Central banks already run agent-based models of specific markets (Baptista et al., 2016). Upgrading scripted agents to empirically grounded generative agents is an engineering project, not a conceptual leap. The realistic role is complementary: LLM-grounded simulations sitting beside DSGE models and econometrics as a scenario-generation and stress-testing layer, valued precisely for producing the heterogeneity and disequilibrium dynamics that equilibrium models suppress.
Behavioral fidelity as a research program. Interview-grounded populations, explicit modeling of measured heterogeneity, and integration of psychological and sociological structure (the original promise of behavioral emulation) now have a working template and clear metrics for progress.
Automated policy search under human audit. The AI Economist’s outer loop generalizes: regulator agents proposing mechanisms, grounded agent societies responding, humans auditing the winners and validating the survivors in narrow human trials. Treat the simulation as a design-search instrument, and the human experiment as the confirmation stage. Teaching deserves a sentence here too: a classroom that can pause a recession, rewind it, and rerun it under a different policy is a pedagogical instrument no problem set can match.
The inversion nobody planned for. The most consequential virtual economies of the next decade may not be simulations at all. As autonomous AI agents begin transacting with real money over open protocols, researchers at Google DeepMind have proposed treating the emerging agent economy as a design problem: deliberately constructed “sandbox economies,” characterized by their origins (intentional versus emergent) and their permeability to the human economy, precisely so that instabilities, flash-crash dynamics among them, do not propagate into real markets (Tomašev et al., 2025). Read alongside CCP hiring a central-bank economist in 2025 to build an open financial system inside a game, the direction of travel is clear:
The tools built to study virtual economies and the tools needed to govern real machine economies are converging into the same discipline.
Conclusion
In 2007, a game studio hiring an economist read as a publicity stunt. In hindsight it was an early observation of a two-decade regularity: virtual economies kept becoming real enough to need economists, real enough to produce inflation that required intervention, bank runs that required regulation, and market-design failures that required repeal. The new development is that the arrow now points both ways. Populated with agents that learn, agents measured against real people at 85% of their own test-retest consistency, and agent societies that spontaneously produce the Phillips curve, Okun’s law, and the stylized facts of financial markets, virtual worlds are close to functioning as instruments rather than anecdotes.
The flight-simulator analogy should be kept honest at both ends. Simulators never replaced flight, and simulated economies will not replace econometrics, natural experiments, or theory. What simulators did was make rare catastrophes rehearsable, and economics has never had that. Used with the discipline the method demands (validated against known regularities, replicated across model versions, bounded by the Lucas critique, and grounded in consented data), AI-driven virtual economies supply the layer our discipline has always lacked: a controlled, replicable middle ground between the blackboard and the world.
In 2007, a game needed an economist. The question for the next decade is how soon the economists will need a game.
References
AL (Altera) et al. (2024). “Project Sid: Many-Agent Simulations Toward AI Civilization.” arXiv:2411.00114.
Arthur, W. B., Holland, J. H., LeBaron, B., Palmer, R., & Tayler, P. (1997). “Asset Pricing Under Endogenous Expectations in an Artificial Stock Market.” In The Economy as an Evolving Complex System II. Addison-Wesley.
Axtell, R. L., & Farmer, J. D. (2025). “Agent-Based Modeling in Economics and Finance: Past, Present, and Future.” Journal of Economic Literature, 63(1), 197–287.
Baptista, R., Farmer, J. D., Hinterschweiger, M., Low, K., Tang, D., & Uluc, A. (2016). “Macroprudential Policy in an Agent-Based Model of the UK Housing Market.” Bank of England Staff Working Paper No. 619.
Blizzard Entertainment (2013). “Diablo III Auction House Update.” Battle.net, September 17, 2013. (Auction house closed March 18, 2014.)
Camerer, C. F., Loewenstein, G., & Rabin, M. (2004). Advances in Behavioral Economics. Princeton University Press.
Castronova, E. (2001). “Virtual Worlds: A First-Hand Account of Market and Society on the Cyberian Frontier.” CESifo Working Paper No. 618.
Castronova, E. (2005). Synthetic Worlds: The Business and Culture of Online Games. University of Chicago Press.
CCP Games (2007). “EVE Online Appoints In-World Economist.” Press release, June 27, 2007.
CCP Games (2026). “Monthly Economic Report.” EVE Online dev blogs, ongoing series (e.g., May 2026 edition), eveonline.com.
Farmer, J. D., & Foley, D. (2009). “The Economy Needs Agent-Based Modelling.” Nature, 460, 685–686.
Fast Company (2014). “Meet the Alan Greenspan of Virtual Currency in EVE Online.” January 2014.
Game Developer (2025). “CCP Hires New Economy Head to ‘Legitimize’ EVE Frontier’s In-Game Economy.” March 2025.
Horton, J. J., Filippas, A., & Manning, B. S. (2023, rev. 2026). “Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?” NBER Working Paper No. 31122.
Li, N., Gao, C., Li, M., Li, Y., & Liao, Q. (2024). “EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities.” Proceedings of ACL 2024, 15523–15536.
Linden Lab (2008). “New Policy Regarding In-World ‘Banks.’” Second Life blog, January 8, 2008.
Lofgren, E. T., & Fefferman, N. H. (2007). “The Untapped Potential of Virtual Game Worlds to Shed Light on Real World Epidemics.” The Lancet Infectious Diseases, 7(9), 625–629.
Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). “Generative Agents: Interactive Simulacra of Human Behavior.” Proceedings of UIST 2023.
Park, J. S., Zou, C. Q., Shaw, A., Hill, B. M., Cai, C., Morris, M. R., Willer, R., Liang, P., & Bernstein, M. S. (2024). “Generative Agent Simulations of 1,000 People.” arXiv:2411.10109.
Piao, J., et al. (2025). “AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents.” arXiv:2502.08691.
Tesfatsion, L. (2006). “Agent-Based Computational Economics: A Constructive Approach to Economic Theory.” Handbook of Computational Economics, Vol. 2, 831–880.
Tomašev, N., Franklin, M., Leibo, J. Z., Jacobs, J., Cunningham, W. A., Gabriel, I., & Osindero, S. (2025). “Virtual Agent Economies.” arXiv:2509.10147.
Varoufakis, Y. (2012). “Arbitrage and Equilibrium in the Team Fortress 2 Economy.” Valve Economics blog, June 2012.
Vezhnevets, A. S., et al. (2023). “Generative Agent-Based Modeling with Actions Grounded in Physical, Social, or Digital Space Using Concordia.” arXiv:2312.03664.
Washington Post (2012). “The Economics of Video Games.” September 28, 2012.
Yang, Y., Zhang, Y., Wu, M., et al. (2025). “TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets.” arXiv:2502.01506.
Zheng, S., Trott, A., Srinivasa, S., Parkes, D. C., & Socher, R. (2022). “The AI Economist: Taxation Policy Design via Two-Level Deep Multiagent Reinforcement Learning.” Science Advances, 8(18), eabk2607.
Additional frameworks cited in text: Shachi (arXiv:2509.21862); SimCity (arXiv:2510.01297).