Social Simulation in Games

There's an old dream in game design that Tynan Sylvester, creator of RimWorld, calls the Simulation Dream: build a complex simulation of a world, and fascinating emergent stories will write themselves — stories as powerful as anything a human author could craft.¹ It's the dream that drove the design of SimCity, The Sims, Dwarf Fortress, and Crusader Kings. It's also a dream that keeps breaking.

Before Ultima Online launched in 1997, associate producer Starr Long described a "virtual ecology" where rabbit populations affected wolf behaviour affected dragon behaviour, cascading up to create adventure hooks automatically. Players entered the world and killed everything so fast the simulation couldn't respawn. Nobody noticed the ecology existed. They ripped it out.¹ BioShock started with a simulation-heavy design — Splicers hunting Gatherers guarded by Protectors — and shipped as a heavily scripted corridor shooter. The simulation was buried under complexity the player couldn't see.

The Player Model Principle

Sylvester's diagnosis is sharp. The problem isn't that these simulations are bad — it's that they violate what he calls the Player Model Principle: the whole value of a game is in the mental model of itself it projects into the player's mind.¹

Every game has two models. The Game Model lives in silicon — algorithms, variables, state machines. The Player Model lives in the player's head, assembled from observation, experimentation, and inference. Players can't perceive the Game Model directly. They only ever experience the Player Model. So anything in the Game Model that doesn't copy into the Player Model is waste. That's what killed the Ultima Online ecology. The simulation existed, but nobody could see it, so it degraded into noise.

But here's the flip side: what if you could put something in the Player Model that isn't in the Game Model at all? What if players could perceive drama, intention, personality — things the computer never simulated? This happens constantly. It's called apophenia.

Apophenia as Co-Author

Apophenia is seeing meaningful patterns in random or meaningless data. You see a face in a wall socket. Michotte's mid-century demonstrations showed that people perceive dominance, gender, intent, and elaborate social relationships between coloured balls bouncing on a screen. "The little ball is trying to play with the big ball, but the big ball doesn't want to play." None of those feelings exist except in the viewer's mind.¹

Sylvester tells a story about The Sims 3. A player created a virtual version of himself and his roommate. A cute redhead appeared and went straight for the roommate. The player felt jealousy, frustration. He hatched an evil plan involving a cheap stove, poor cooking skills, and wooden chairs. Fire. Revenge. The game simulated attraction and fiery death, but the jealousy, the frustration, the sorrow — the emotional core of the story — was projected entirely by the player. The simulation was a co-author, not the author. It did the logistics and generated outcomes; the player added the meaning.

This is deep Predictive Processing territory — the brain as a pattern-completion engine, hallucinating intention into mechanical systems. And it means that simulation games don't need to simulate emotion or intelligence. They need to simulate enough structure that the human mind will project emotion and intelligence onto the result.

Dwarf Fortress: The Accidental Masterpiece

If anyone has pushed the Simulation Dream further than reason would suggest, it's Tarn Adams. Dwarf Fortress is "perhaps the most complex video game ever made," and it's still only in alpha.² Its interface is a dense tapestry of ASCII characters — ♠§dg is a dog leashed to a tree, about to be mauled by a goblin — and beneath that surface is a staggering array of interlocking algorithms modelling personality, climate, economic patterns, history, material science, fluid dynamics, and dwarven psychology.

The stories aren't scripted. "We didn't know that carp were going to eat dwarves," says Tarn's brother Zach. "But we'd written them as carnivorous and roughly the same size as dwarves, so that just happened, and it was great."²

The game's unofficial slogan — "Losing is fun!" — captures something essential about what makes social simulation work. Dwarf Fortress doesn't try to guarantee interesting outcomes. It builds a system complex enough that interesting outcomes are statistically inevitable, and then it lets catastrophe do the narrative work. Stampeding elephants, famine, vampire dwarves — the game generates situations that put human values at stake (life, death, community, loss), and the player's apophenia fills in the emotional content.

But the genius of Dwarf Fortress isn't just depth. It's also what Sylvester calls "hair complexity" — decorative details that don't affect anything but help the player build a mental image. Each dwarf has an appearance, preferences, opinions about art. None of it feeds back into the core simulation. But it gives the player material for projection. You care about a dwarf not because the AI is sophisticated, but because you know she's depressive and appreciates good masonry, and now she's standing in front of a goblin siege.

The Design of Social Simulation

Mitu Khandaker's GDC talk on social simulation offers a framework for thinking about where your game should sit on the spectrum between fully emergent autonomous behaviour and fully authored branching narrative.³ Neither end is inherently better. Emergent systems give you "tiny moments of awe" — things that happened only for you, that will never repeat exactly. Authored narrative gives you emotional depth and nuance that no procedural system can match. The art is in choosing your position on the spectrum and managing the tradeoffs.

Six considerations shape good social simulation design:³

Where on the autonomy-authoring spectrum? Shadow of Mordor starts with scripted narrative and focuses most action on the emergent Nemesis system. Blood & Laurels balances authored events with autonomous agents. The Walking Dead is almost entirely authored but creates deep emotional investment. All valid.

Communication. Even a magically nuanced simulation fails if the player can't tell what's happening. The Sims uses thought bubbles. Crusader Kings uses character sheets and opinion modifiers. Blood & Laurels lets you click on characters to see their state of mind. More mediation is good — it removes ambiguity and helps the player understand where they have agency.

People as ends, not means. If characters exist only as tools for the player's advancement (raise relationship, unlock combat stats), you lose the thing that makes social simulation interesting. Persona 4 ties social links to combat abilities, which works because the character writing is charming enough to make you care anyway. But it's a tension.

Expression through limitation. Restricting the possibility space can say more than opening it up. The Sims reduces all behaviour to object affordances — people and things interact through contextual menus. That limitation is what makes it feel like a game about the acquisition of stuff, which is the satirical point. Redshirt reduces all interaction to a parody social network, which is its commentary on online relationships.

Diversity of characters. A social simulation with identical agents is boring regardless of complexity. Different identities, motivations, and personality parameters create a wider range of possible outcomes. Crusader Kings II's five hidden personality parameters (rationality, greed, zeal, honour, ambition) combine to produce characters that feel distinct, and the emergent drama between them is what keeps players writing campaign novellas on forums.

Social simulation as supporting system. Not every game needs to centre people. Skyrim uses faction membership and a few personality axes (aggression, assistance, confidence, morality) to create emergent conflict that makes the world feel alive — Imperial soldiers and Stormcloaks fighting in the wilderness not because a designer placed them there, but because their factions hate each other.³

The Nemesis System and Emergence on a Budget

Shadow of Mordor's Nemesis system is perhaps the purest recent example of social simulation designed for maximum apophenia at minimum complexity. The orcs' only mode of interaction with each other and the player is brutal violence. They climb ranks by killing. That's it — one dimension of social behaviour.

But because they're orcs, this one-dimensionality is narratively coherent. Their motivations are believable. If they were humans, the system would feel shallow. As orcs in a Tolkien universe, it works. The result is that you end up caring about orc drama — remembering who killed you, tracking rivalries, feeling genuine satisfaction when you take down a nemesis who defeated you three times. The system generates attachment from almost nothing, because the player projects all the meaning.³

Economies as Social Systems

One of the most underappreciated forms of social simulation is the emergent economy. A brilliant essay on a survival Minecraft server traces how a player with no initial knowledge builds a complete mental model of the server's economy — price ranges, supply chains, seller tendencies — and uses that model to dominate the market.⁴

The dynamics are fascinating because virtual economies differ from real ones in precise, learnable ways. Diamond in Minecraft is nonrenewable; wheat is farmable; cactus is autofarmable. Each category follows different growth curves. The player who understands these curves can exploit the gap between perception and reality. Coloured wool was worth the same as white wool in the market because players assumed mechanically identical items should cost the same — but coloured wool was far harder to supply in bulk, making it genuinely more valuable to anyone who needed it.

This is social simulation without NPC AI: the simulation is the game rules, the social agents are other players, and the emergent stories are about market manipulation, cartel formation, and the discovery that game economies reward the same careful modelling of systems that any simulation rewards. "I've always loved knowable systems," the author writes. "People are messy and complicated, but systems don't lie to you."⁴

When the NPCs Talk Back

Everything above assumed that the characters in a social simulation are systems — state machines with personality parameters, opinion modifiers, relationship scores. The player projects meaning onto mechanical behavior. But what happens when the characters can actually hold a conversation?

The first real test case isn't a carefully designed AI NPC system. It's cheating in Old School RuneScape. OSRS has been plagued by bot farms for years — automated scripts that grind for in-game gold to sell on secondary markets for real cash. At any given moment, roughly a quarter to a third of the players logged in are bots. The community's most reliable detection tool was simple: talk to them. Bots couldn't hold a conversation.⁵

Then, in March 2023, bot developers started hooking their scripts up to ChatGPT. Within weeks, the bots could maintain context across conversations, remember previous interactions, and even commiserate with real players about the overabundance of bots. One bot, documented by the YouTuber SirPugger, adopted the persona of a grouchy longtimer who insulted nearby players. The community's grassroots detection system collapsed overnight. "We're quickly approaching a state where you won't be able to tell if you're talking with a bot."⁵

This raises Sylvester's Player Model Principle in a strange new light. If a bot is indistinguishable from a human player — maintains friendships, participates in social activities, has opinions about the game — does it matter that it's automated? SirPugger himself sees potential upsides: "I like multiplayer games because of the player interaction. So all of a sudden, you have more 'players' in the game because you can't tell that they're bots. Bots are suddenly kind of this positive externality."⁵

The game industry is already exploring the intentional version of this. Fan-made Skyrim mods give NPCs ChatGPT-powered memories of past interactions. Replica Studios has unveiled AI-powered NPCs for Unreal Engine that can hold complex conversations. Inworld AI reports that 99% of gamers they surveyed believe AI NPCs would positively affect gameplay.⁵ The vision: instead of clicking through scripted dialogue trees, you talk to characters who remember what you said last time and react based on their personality model.

But significant hurdles remain. Language models hallucinate facts, which makes them unreliable for advancing storylines. Free-form conversation is less likely to deliver relevant quest information than a carefully authored dialogue tree. And the economic implications are serious — if AI bots permanently warp the economics of an MMO (as they already have in OSRS, where bot farms rake in six figures annually on the black market), the social fabric of the game community changes even if individual interactions feel richer.⁵

The deeper question connects back to apophenia. Sylvester's insight was that players project emotion and meaning onto mechanical systems. But what happens when the system talks back? Does apophenia increase (because the projection has more surface area to work with) or decrease (because the illusion of real conversation creates expectations the system can't meet)? The first generation of ChatGPT-powered game bots suggests both at once: conversations are more engaging than scripted dialogue, but also more prone to breaking the fiction when the model says something obviously wrong or out of character. The simulation dream now has a new variant: not just emergent stories from mechanical interactions, but emergent stories from simulated conversation. Whether that's a step toward or away from the dream remains to be seen.

Why the Dream Persists

The Simulation Dream keeps breaking because simulation complexity is easy and player comprehension is hard. You can simulate a thousand variables, but if the player can't see the cause-and-effect chains, the stories stay buried in silicon. The path forward isn't more variables — it's what Sylvester calls "story-richness": the percentage of interactions in a game that are actually interesting to the player.¹

Real life isn't story-rich. Most days for most people are mundane. A simulation game can't be a faithful model of its subject — it has to be a narratively condensed, intensified version. Choose the minimum representation that supports the stories you want to generate. A prison sim probably doesn't need nuanced food simulation because prison stories aren't about food. A colonial survival game probably does, because starvation drives the narrative.

The games that succeed at social simulation — Dwarf Fortress, Crusader Kings, RimWorld, The Sims — all understand this. They don't try to simulate everything. They simulate the right things at the right granularity, lean heavily on the player's apophenia to fill in emotional content, and use "hair complexity" for flavour that doesn't add system load. The simulation is the scaffold. The player is the storyteller. The dream isn't about making the computer write stories — it's about making the computer and the human write stories together.