Blog · Review Essay · Last reviewed June 25, 2026

Games of Empire and the Playable Machine of Power

Nick Dyer-Witheford and Greig de Peuter's Games of Empire is a book about video games, but its durable subject is the way power becomes interactive. Work becomes play. War becomes training software. Markets become virtual economies. Users become labor, audience, data, and test population. The book matters in the AI era because it shows how rule-bound worlds govern conduct while teaching people and machines what the system counts as success.

Playable power, in this review, means authority embedded in an interactive environment: rules, affordances, rewards, constraints, identity, telemetry, markets, and feedback loops that train conduct by making some actions visible, cheap, repeatable, and rewarded. The practical test is whether affected people can inspect the rules, understand the scoring, contest the records, report harm, and change the environment that is training them.

The Book

Games of Empire: Global Capitalism and Video Games was published by the University of Minnesota Press in 2009 as volume 29 in the Electronic Mediations series. The publisher lists the paperback ISBN as 9780816666119, the e-book ISBN as 9781452942704, and the print edition at 320 pages. The book's contents move through labor, capital, consoles, military simulation, World of Warcraft, Grand Theft Auto, activist games, open-source game development, metaverse speculation, and the material supply chains beneath virtual play.

The authors were not trying to defend games from moral panic or celebrate them as harmless entertainment. They read commercial video games as exemplary media of a global system in which economic, military, communicative, and cultural power become difficult to separate. That is why the book belongs beside A Hacker Manifesto, Platform Capitalism, Surveillance Valley, Coding Freedom, Hamlet on the Holodeck, and The Stack.

Its durable insight is simple: games are not only representations. They are operational environments. They organize attention, feedback, scarcity, coordination, ranking, identity, training, and reward. A game is a small constitution for action: it defines what can be seen, what can be done, what matters, what is cheap, what is rewarded, and what disappears outside the rules. In a society increasingly governed by dashboards, scores, simulations, agents, and behavioral loops, that makes games a serious object of political analysis.

The sharper definition for this review is playable power: authority that does not only command from above, but arranges an environment where people learn the approved moves by acting inside it. The same pattern appears when a workplace is run through a dashboard, when an AI evaluation becomes a leaderboard, when a platform economy turns attention into currency, and when a synthetic world becomes a training ground for agents.

Current Context

As of June 25, 2026, the book speaks to three live governance surfaces. First, game interfaces are consumer-protection and child-safety objects. The Federal Trade Commission's 2022 Fortnite action treated unauthorized charges, dark patterns, children's privacy, and default communications settings as enforceable design issues. The EU Digital Services Act similarly treats online interface design, recommender transparency, advertising, minors, and systemic risk as subjects of law rather than private taste.

Second, games have become distribution points for generative AI. Valve's January 9, 2024 Steamworks policy separates pre-generated AI content from live-generated AI content, asks developers to describe use and guardrails, says much of the disclosure will be shown to customers, and adds reporting for illegal content in games that use live generation. That does not settle copyright, labor, safety, or deception questions, but it names the right unit of governance: model use, developer claim, platform review, customer notice, reporting channel, and enforcement process together.

Third, generated worlds are becoming training and evaluation spaces for agents. Google DeepMind's Genie 3 announcement describes real-time interactive generated environments and also names limitations around action space, multi-agent interaction, geographic accuracy, readable text, and interaction duration. Those limits matter. A plausible world can be useful for research, prototyping, or rehearsal without being a validated substitute for physical, social, legal, or military reality.

The Game Engine

The book's first major move is to treat the game industry as a labor system. The word "play" does not make production light. Games depend on programmers, artists, writers, designers, moderators, hardware supply chains, marketing teams, platform owners, modders, fan communities, players, gold farmers, and many forms of unpaid or underpaid contribution. The polished surface depends on people whose work is converted into content, community, telemetry, and future sales.

That argument has become more obvious since 2009. The game industry now includes live-service pipelines, app stores, esports, streamers, asset marketplaces, modding platforms, player analytics, creator economies, and constant pressure to keep worlds fresh. But Games of Empire had already named the basic pattern: digital play can hide the conversion of enthusiasm into value.

This matters beyond games because the same pattern now appears across AI products. Users prompt, correct, rate, share, remix, test, jailbreak, flag, and socialize with systems whose owners learn from the activity. Open-source maintainers provide training data. Forum participants become retrieval material. Moderators become safety examples. Customer-service workers train bots that later supervise or replace pieces of their work. The interface says play, help, create, improve. The institution receives labor, data, and behavioral evidence.

The hard line is not between voluntary creativity and exploitation. Modding, streaming, bug reports, community guides, playtests, chat moderation, and ratings can be meaningful participation. They become governance problems when a platform depends on them while denying contributors visibility, bargaining power, exposure protection, compensation, or appeal. That is why the book belongs beside Ghost Work and Work Without the Worker: the machine often looks smooth because people are doing repair work at the edge of the interface.

Game development also previews the workplace politics of artificial intelligence. The promise is creative augmentation. The risk is accelerated production under tighter managerial measurement. If a tool lets studios generate more assets, dialogue, animation, and tests, the critical question is not only whether the output is good. It is who gains time, who loses craft, who absorbs review labor, who can refuse a tool, and who owns the pipeline. That is the same question raised by algorithmic management and by accounts of platform labor: automation often moves labor into review, repair, moderation, and compliance rather than making it vanish.

Military Simulation

The book's chapter on Full Spectrum Warrior remains one of its clearest bridges to the present. Dyer-Witheford and de Peuter read the game through the entanglement of entertainment software, military training, recruitment, weapons culture, and the translation of war into interface. The point is not that every player becomes a soldier. The point is that combat can become a manageable screen problem: squads, objectives, fields of vision, cover, constraints, and feedback.

That framing is now central to AI in warfare and to the governance of world models. Military and security institutions increasingly think through simulations, autonomous systems, synthetic training environments, sensor fusion, target recognition, drone interfaces, and decision-support tools. A training world is never neutral. It encodes what counts as a threat, a civilian, a target, a success condition, an acceptable loss, a clean action, and a manageable uncertainty.

The game form makes this problem legible because it turns violence into procedures. The user learns the world through possible actions. The interface does not simply show a conflict; it teaches what can be done inside it. In the AI era, the same design question returns with higher stakes: what behaviors are agents rehearsing in simulated worlds, what reward systems are shaping them, and what institutional assumptions become invisible because they are embedded as rules?

A model trained in a game-like environment can learn competence without learning the politics of the world it is asked to act within. That is why simulation is not just a technical matter. It is a form of institution-building by other means. Safety analysis has to ask whether the simulation preserves the variables that matter for real people, and whether human oversight can interrupt the procedure when the game board is wrong.

A serious simulation safety case should therefore document omissions as carefully as capabilities. What classes of civilian, weather, sensor failure, uncertainty, adversarial action, legal constraint, and command breakdown are absent? What does the scoring function reward? What does it make impossible to notice? Without that file, simulation becomes a persuasive interface rather than evidence.

Virtual Economies

Games of Empire is also useful because it refuses to keep the virtual and the material separate. Its analysis of online worlds and gold farming treats virtual economies as real economies with labor, currency, scarcity, fraud, status, extraction, and cross-border inequality. A sword, skin, coin, account, avatar, server, mod, or reputation score may be digital, but the labor and infrastructure around it are not imaginary.

This is one of the book's most useful lessons for AI-era media. The digital object often looks weightless from the user's side. The system, however, depends on hardware, energy, supply chains, moderation, data cleaning, annotation, cloud contracts, payment rails, legal rights, and workers distributed across the world. The fantasy of virtuality can make those dependencies easier to ignore.

Games sharpen the problem because they make abstraction pleasurable. A user can inhabit a world where labor has been converted into progress bars, skins, levels, markets, leaderboards, achievements, cooldowns, quests, and social rank. That same grammar is now everywhere: productivity dashboards, learning apps, fitness trackers, creator analytics, workplace metrics, loyalty programs, reputation scores, and model evals.

Once institutions adopt that grammar, people begin acting toward the score. The system does not merely measure behavior. It produces behavior organized for measurement. That is why virtual economies belong next to The Tyranny of Metrics: both show how a proxy can become the environment people have to survive.

The current governance example is not abstract. The Federal Trade Commission's Fortnite action treated game-interface design, unauthorized charges, children's privacy, default communications, and dark patterns as consumer-protection issues. The European Union's Digital Services Act similarly treats online interface design, recommender transparency, advertising, minors, and systemic platform risk as governable surfaces. In both cases, the question is not merely what content appears on a screen. It is how the screen arranges choice.

The review question for a game economy should therefore be concrete: What currencies exist? Can they be converted, traded, refunded, or obscured? Are minors present? Are odds, fees, cooldowns, subscriptions, cancellations, and battle-pass deadlines clear? Can a player appeal a ban, recover an account, export records, or challenge a charge? Who receives marketplace fees and creator payouts? Virtual property is not weightless when money, children, labor, identity, and status are routed through it.

Recursive Reality

This is where the book becomes a theory of recursive reality. Games model social worlds, then players act inside those models, then platforms measure the action, then developers redesign the world around the measurements, then players adapt again. The representation becomes an environment. The environment becomes data. The data becomes the next representation.

The loop is not limited to entertainment. A school dashboard turns learning into visible metrics, then students and teachers adapt to the dashboard. A workplace tool turns effort into tickets and throughput, then workers adapt to the tool. A recommender turns culture into engagement signals, then creators adapt to the recommender. A benchmark turns model quality into a leaderboard, then labs adapt to the benchmark. A synthetic training world turns reality into reward structure, then agents adapt to the reward structure. This is the same failure mode described by reward hacking and by the site's page on AI evaluations: a proxy can become a curriculum.

Games of Empire helps because games make the loop explicit. Rules, rewards, constraints, maps, inventories, enemies, markets, and avatars are all built. Nothing about the world is natural, even when the player becomes fluent inside it. The danger is forgetting that institutional interfaces are also built worlds.

The strongest AI systems increasingly operate through worlds made for them: browser environments, coding sandboxes, robotics simulators, enterprise connectors, game-like evals, synthetic users, virtual offices, generated classrooms, and multi-agent arenas. Each world teaches the model what reality is like. Each world also teaches users what machine action should feel like. This is why environment design is part of AI governance, not a downstream user-experience detail.

The AI Reading

Read in 2026, Games of Empire is not mainly about whether video games are good or bad. It is about the institutional power of interactive systems. An AI agent, like a game character, acts inside an environment with rules, affordances, permissions, objectives, state, feedback, and memory. The politics lives in that environment as much as in the model.

That shifts the governance question. Asking whether an AI system is intelligent is less useful than asking what world it has been given to act in. What can it see? What can it change? What counts as reward? Which actions are cheap? Which harms are invisible? Who is modeled as user, adversary, worker, customer, target, student, patient, or obstacle? Who can appeal when the system's game board misrepresents them?

The book also clarifies why "gamification" is not a superficial design trick. When an institution turns conduct into points, ranks, badges, alerts, streaks, dashboards, or objectives, it is building a behavioral machine. The user may experience motivation, progress, or fun. The operator receives legible action.

For AI, this matters in both directions. Humans are gamified into data-producing behavior, and machines are trained through gamified reward structures. The same culture that learned to make people chase scores now trains models to chase scores. The result can be competence without judgment: systems that are excellent at the game because the game was easier to specify than the world.

Generative AI inside games makes the argument more literal. A platform can now host a game that used AI during development, a game that generates material while it is running, or a generated world used to train another agent. Steam's 2024 AI disclosure policy separates pre-generated and live-generated uses, asks developers to describe guardrails, surfaces much of the disclosure to customers, and adds reporting for illegal content in games with live-generated AI. That is not a complete safety regime, but it shows the right governance object: model, store, developer promise, customer notice, reporting channel, and review process together.

World-model research raises the same issue at a higher level. Google DeepMind's Genie 3 announcement describes real-time interactive generated environments for agents while explicitly naming limits around action space, multi-agent interaction, geographic accuracy, text rendering, and interaction duration. The lesson is not to treat generated worlds as proof of general intelligence. It is to ask what a synthetic environment can actually support, what it cannot verify, and how quickly a convincing rehearsal can be mistaken for the world.

An AI game or generated world should carry an environment card: the rules of the world, the model components involved, the action space, reward signals, telemetry, content-generation points, moderation process, retention period, child-safety assumptions, reporting route, and known failure boundaries. This is the playable-world version of model documentation. It tells reviewers what the world is training, not only what it displays.

Governance and Safety

The governance lesson of Games of Empire is that a platform world is a safety system whether or not it calls itself one. The rules, store policy, age gates, chat defaults, virtual currency, account system, moderation queue, recommender, telemetry, ranking, generated content tools, and appeal routes all shape conduct. If those layers are treated as ordinary product design, power has already moved where the policy team is not looking.

For games, consumer protection begins at the interface. Purchases, loot, skins, battle passes, virtual currencies, cancellation paths, parental controls, privacy defaults, voice chat, and reporting flows are not decorative details. They are the mechanisms through which children, players, parents, creators, and workers encounter the institution. A dark pattern in a game economy is not less political because it is attached to a costume or coin.

For AI-enabled games and generated worlds, governance has to follow the lifecycle. Useful controls include documented training sources and rights, clear separation of pre-generated and live-generated material, customer-facing disclosure, provenance or labeling where synthetic media could mislead, guardrail evidence for live generation, abuse reporting, escalation to human review, incident logs, and the ability to pause a feature when evidence fails. The relevant companion pages are AI governance, content provenance, synthetic media, the EU AI Act, and human oversight.

For military, robotics, and agent training, the standard should be stricter: a simulation should be treated as evidence only within a documented safety case. It should say what the world model preserves, what it omits, how it was validated against real environments, what failure cases are known, what distribution shifts are expected, and who can stop deployment when the rehearsal no longer matches reality. NIST's AI Risk Management Framework is useful here because it frames AI risk as a design, development, use, and evaluation problem rather than a one-time release checklist.

A minimum playable-world file should name purpose, affected users, child exposure, rules and affordances, reward and monetization design, virtual currency and refund paths, generative AI use, training or source data, telemetry, moderation, human review, appeal, incident triggers, rollback, and shutdown authority. If a vendor, studio, school, agency, or commander cannot produce that file, the world is being governed by product instinct rather than accountable design.

Where the Book Needs Friction

The book's central weakness is also its ambition. "Empire" can become too large a category. If every object, platform, genre, workplace, supply chain, and player practice is folded into one total system, the analysis can begin to feel as if the answer arrived before the evidence. Several reviewers make versions of this criticism while still treating the book as important.

The historical moment also shows. A 2009 account could not fully absorb mobile free-to-play, Twitch, Discord, Roblox, Fortnite as platform, loot boxes, battle passes, cloud gaming, AI-generated assets, generative NPCs, or the politics that became visible through Gamergate. The authors' 2021 postscript partially updates the frame by naming climate crisis, platform proliferation, organizing, reactionary game culture, and gaming's violent intersections as part of any renewed research agenda.

The book is strongest on production, labor, military-entertainment entanglement, and the material underside of virtual economies. It is weaker when a reading of a game's representational content has to carry too much political weight. A game can represent empire, but the more durable question is how the whole stack works: studio, engine, console, store, cloud, account system, payment rail, community, labor process, mod scene, data pipeline, and player habit.

That limit is useful for AI criticism. Do not stop at reading the chatbot's output, the agent's screen, or the generated image. Read the stack that made the output possible and the workflow that will use it.

What This Changes

The practical lesson is to treat interactive systems as political environments, not neutral containers for action.

When a system becomes game-like, ask what it makes playable. Does it make work feel voluntary while increasing extraction? Does it make war feel procedural? Does it make surveillance feel like personalization? Does it make ranking feel like merit? Does it make virtual property feel weightless while hiding labor and infrastructure? Does it make a benchmark feel like intelligence?

Then ask what the system is training. It may be training a player, worker, soldier, student, patient, moderator, customer, or model. It may also be training an institution to see only what the interface can score.

Finally, ask who can change the rules. A playable system is safer when affected people can understand its incentives, contest its records, opt out where possible, report harm, trigger review, and see evidence that complaints change the system. A world with no appeal is not just immersive. It is administrative.

Games of Empire matters because it shows that play is not outside power. Play can be a route through which power becomes intuitive. In an AI era of simulated training worlds, agent benchmarks, synthetic users, gamified labor, platform economies, and dashboards that turn conduct into reward signals, that is not a side topic. It is one of the places where the future is rehearsed before it is called inevitable.

Source Discipline

This review separates four kinds of claims. Book metadata comes from the publisher and bibliographic records. Scholarly reception comes from reviews, essays, and the authors' later postscript. Current governance claims come from official sources: the FTC, Valve, EUR-Lex, Google DeepMind, and NIST. The interpretive claim is this review's argument: interactive systems govern by turning rules into lived environments.

The examples do not prove that every game, platform, simulation, or AI world is harmful. They show where the evidence burden belongs. A game economy, live-generated AI feature, training simulation, or agent world should be evaluated through its rules, incentives, labor chain, data flows, reporting process, and appeal structure, not only through visible content or benchmark scores.

This article makes no claim that AI systems are conscious, divine, inevitable, or already general intelligence. It treats games, agents, and world models as sociotechnical systems: code, labor, data, institutions, markets, rules, interfaces, and people acting inside them.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books