How one studio is building game AI to replicate a human storyteller

Artificial intelligence in games is often designed to be an actor, not a director.

Typically, AI drives characters in games to behave and respond to the player in a way that feels believable and entertaining, but remains within the constraints of a world designed by humans. This is my open-world game, say the human developers, and here are the things players may do within it; a player may choose how and when to engage with the game's various facets, but the game will never create new stories tailored to the player's interests; It can't.

But what if it could? What if you could design a game capable of generating its own narratives, via an AI that watches what the player does and tailors the game to their actions?

That's the pitch for Toska, a new "procedural AI narrative engine" that Canadian developer Evodant Interactive has been working on, in some form or another, for nearly a decade. Whether they pull it off remains to be seen, but the effort alone is intriguing.

"The intention behind Toska is that it provides for a computer game experience what a good human DM provides on a tabletop experience," Evodant narrative lead (and tabletop roleplaying afficionado) Ryan Fitzgerald told me at GDC earlier this month. "There's story and there's gameplay, but we really want to use Toska to try and twine these together. The game observes what you do. It watches the kind of behaviors you engage in, what you do unconsciously...and then it provides you with the challenge and the narrative framework to give you that which will satisfy you."

Setting up stories that tell themselves

The game he's referring to is Gyre: Maelstrom, Evodant's upcoming steampunk action RPG. This will be the first game designed around Toska, though Fitzgerald says the studio has been working on its procedural narrative generation tech for roughly eight years, pitching it to various publishers along the way. Only now is the Canadian studio able to build a full game around it thanks, in large part, to a recent $1.15 million grant from the Canada Media Fund. 

"It's given us the runway to actually produce a complete and polished game," Fitzgerald says. "We're hoping that this is another way to do narrative. You can have tragic stories [in Gyre.] You can have satirical stories. You can have these kinds of stories if you, the player, are engaging in behaviors that suggest that that is where you're going."

It's an ambitious pitch, to be sure, if not entirely novel; many games have attempted something similar (Bethesda's Radiant quest system springs to mind) and many developers have spoken at length about how AI and procedural generation techniques might be applied to video game storytelling. Most recently, game AI expert Dave Mark suggested to Gamasutra that it's reasonable to expect in the near future we'll see AI that can autonomously collect data on how a player likes to play and create new content to fit their playstyle.

However, the scenario Mark outlined involved AI remixing modular pieces of game content to create fresh experiences within designer-authored quest templates. By contrast, Evodant seems to be designing Toska to work without templates, or at least to be able to tweak and edit them to such a degree as to render the distinction meaningless.

In practice, this means putting players through a character creation phase and then plopping them down into a procedurally-generated world and spinning up characters, quests and entire campaigns just a few steps ahead of the player by watching how they behave and dynamically adjusting the game accordingly.

A (very early, pre-alpha) build of Gyre shows Toska's monitoring and actions described in the upper-left

"We have an observation period and an execution period. Toska will observe what you are doing until it has enough certitude that it can make some conclusions on what kind of gameplay you are looking for," says Fitzgerald, noting that the engine (which is being designed as an Unreal plugin) will generate challenges for the player if they don't strike out on their own. "If nothing's happening for a while, something will happen that forces you to do a rock-paper-scissors type choice. That doesn't dictate the story, but it sets the ball rolling so that the player can engage in other activities and continue with the observation period."

From counter-intelligence expert to game designer

Fitzgerald has some experience in observing people and predicting how they'll react. Before he was a game developer, he served in the Canadian Air Force and worked in counter-intelligence. He became the anti-terrorism advisor to the deputy commander commander of the North American Aerospace Defense Command (NORAD), but eventually got out of the service because, as he says, "my wife and I both agreed that we liked her more than my job."

So he took a break for a while, then took a shot at getting into film and TV work. He wound up going through the Canadian Film Centre's interactive media program, and landed in alternate reality game design.

"Between the anti-terrorism stuff and the military police stuff and the interactive work, it allowed me to get into what was then the heyday of alternate reality game design," says Fitzgerald. "I did a lot of transmedia stuff before it was called transmedia, and the goal of transmedia, and in many ways marketing, is very similar to counter-intelligence. Which is, I want a large group of people to all think what I want them to think, and I want them to think it was their idea."

Fitzgerald worked in marketing and transmedia game design for awhile, but grew tired of selling the work of others and wound up at Evodant. He's not alone, either.

"A lot of people have gotten out of alternate reality gaming and transmedia because they're tired of marketing somebody else's product," he says. "So I can trot it out, because I know the tactics and the techniques, but I really enjoy just being creative on my own product."

So how, exactly, does that experience affect his work on Toska? Fitzgerald says he spent a heck of a lot of time just observing people, and studying how they initiate and respond to certain behaviors in public spaces, like airports.

"I'd sit in a boarding lounge with my journal out, just writing conversations verbatim, while there was a large group of people right behind me," Fitzgerald says. "All I was doing is just listening. It was an enormous preparation for designing the [Toska] AI, how we may anticipate what players may do in response to AI behavior. We have this library of gameplay moments, where the only reason the AI is going to have an NPC say or do something is because we want to see what happens if I push that button."

"We need to have smarter ways to do these things"

Here, Evodant creative director and CTO Dwayne Rudy is quick to jump in and point out that "We don't want to give the illusion that there's one magic algorithm that pulls everything together." Toska's final form still seems to be sort of up in the air, as the team nails down details in the course of Gyre's development. Right now, Rudy describes Toska as multiple systems, each with subsystems, that all talk to each other and influence things like NPC AI behavior patterns (goal states, drives, even dialogue), the game's stage manager (which can do things like dynamically change camera distance to, say, ratchet in close during a scene intended to evoke suspense) and a sort of dynamic narrative planner that can create new quest lines for the player based on their actions.

"The narrative planner is a very different beast than you might find in a typical RPG," says Rudy. "Our narrative planner acts differently; it goes to a certain depth and says 'Okay I can kind of see where we're going," and gives the player more of that path. But it continues to watch, and continues to adapt itself, so it's like 'Do I want more dramatic moments? Oh wait, they murdered someone. Okay, we're switching gears now.'"

From the same (very early, pre-alpha) build of Gyre, one example of how a player might construct sentences, which Toska can (theoretically) parse and respond to

The folks at Evodant say they're also interested in potentially licensing the tech out to other devs, so if Toska proves itself capable of procedurally generating satisfying narrative arcs for players, the game industry may shift accordingly.

Procedural storytelling, a realm already being explored by games like Moon Hunters, may become more robust and commonplace. Narrative designers may feel their jobs are under fire, though Fitzgerald is quick to point out that's not Evodant's intent.

"I don't want to put writers out of work." said Fitzgerald. "I'm a writer. We're already one step up from homeless and drunk. The last thing I want to do is tank the careers of my people further than they already are."

As you might expect, Fitzgerald pitches an optimistic view of this tech: if it works well, it could expand the scope of what small game design teams (or even solo developers) can accomplish, in much the same way that increasingly capable proc-gen asset systems have allowed devs large and small to make remarkably complex and detailed games.

"We need to have smarter ways to do these things," says Fitzgerald. "Nobody really wants to have a content farm of artists and developers working away in cubicles, constantly creating stuff that takes them six months to create and a fan 48 hours to burn through."