Characters in games and other digital experiences tend to be rather static, working from a set of lines and responses written long ago. But the future of games could be more responsive, generative, and of course AI-powered — something Inworld AI is attempting to enable with a newly available beta tool that lets developers create a rich, interactive characters as simply as they might tell another AI to draw a bird.
Inworld’s claims, which it has been putting about over the last year, are that it is able to quickly create NPCs and such like characters with just a few sentences of description and twiddled dials, and that once created they will instantly be deeper and more interesting to interact with than ordinary scripted characters.
Now, there are obvious limitations to these claims — you couldn’t, for example, outdo the cryptic utterances of characters in Elden Ring, since those are highly crafted scripts intended to be encountered in a specific way. But what about the lady who runs the weapon shop in a fantasy world? Normally all she’d say is “Buy somethin’, will ya!” or the like — and Inworld wants to make her a more real presence, someone you can ask about the world, her family, the problems of the area, the weapons themselves, and get intelligent, meaningful answers.
Be sure that I went into the demo more than a little skeptical. In the first place, who wants to ask the shop lady about her favorite color or the weather? And in the second, how would you be able to make such a fully fleshed-out character with such a simple process?
The second part first. If you think about how generative AIs like GPT-3 work, they are trained on huge datasets of language, and when you give them a short prompt, they extrapolate from it based on that larger body of knowledge. So “A skeleton goes to the store” instantly turns into a 3,000-word story. Inworld appears to extrapolate in the same way, but focused on how a character with certain knowledge and other aspects would react to various questions and situations.
Building a character is an extremely simple process, much more so than the traditional way of setting up networks of dialogue and story triggers. “Our interface is very simple — you create a character with natural language, no code at all,” said Kylan Gibbs, Inworld’s Chief Product Officer. (To be clear, the visuals are done elsewhere, so nothing here establishes or limits the graphics.)
First there’s simply a “core description” that defines the identity: a few sentences, like “Asha is a weaponsmith and merchant in the town of Rolheim. She comes from the far north, where her family is. She wants to convince people to buy her weapons so she can become a master smith in the great southern city of Ekomit.”
“The reason this is important is that none of this is scripted, so this tells her how she should interact,” said Gibbs, as he scrolled through the aspects of a character they’d put together for a demo. “It’s biasing a character towards a personality.”
There are dozens of optional fields that inform the character of their stage of life, their motivations, their tendency towards sadness, politeness, etc. Specific rote responses for specific questions in case you need to those. And a field where you can put common knowledge like the general geography of the world, who lives in it and where, how to get to the inn in town, and so on. Ideally the character should be able to interact like a “real” denizen of that world. (You can also easily blacklist words or topics, and there are a set of standard safety filters.)
It runs online, meaning the game would be querying a large language model constantly for dialogue — though there would be graceful fallbacks — essentially to the original scripted style. You add the character to your environment as you would another asset.
In the demo they showed me, a character equipped with a core identity and a general sense of the plot interacted with one of the developers, who asked a variety of questions and elicited as many reactions. The responses nearly all had the feeling of natural conversation or scripted responses, but none of it was coded in. She speculated on the identity and behavior of a villain, confirmed the source of her beliefs about the relationship between robots and humans, and generally was responsive to relevant queries.
I’m sure you could break the spell by asking “What Tiktok cooking creators do you follow?” but you can also make a game world look unrealistic by crouching and looking at the ground up close. Why would you? What’s more important here is that when you ask “Who’s the boss around here” they don’t respond “President Biden is the boss” or something wild like that — always a potential issue with generative AIs. The game-relevant responses the character gave were genuinely impressive, though of course it was a limited and staged demo.
The company today published a video showing off some of the generative dialogue in an upcoming game or experience they plan to publish:
All right, so it’s a fancy chat bot. What’s the point? You spend more time pointing your gun, or sword, or gunsword, at people in games than talking with them. Why equip them with more intelligence than could possibly be practical? One thinks of Edge’s review of Doom in 1994asking “If only you could talk to these creatures, then perhaps you could try and make friends with them, form alliances…”
Of course it makes little sense in Doom, and therein lies the joke. But what if you could converse with the characters of Monkey Island to find clues, or ask your RPG party members about their hometowns when you sit around a fire, ask your strategist for ideas in battle, or interrogate a witness outside of a handful of dialogue choices?
The truth is our games have been designed around the concepts of scripted NPCs and dialogue trees because that’s all that was ever possible. What games might be possible if you work around the idea that the characters in it have some knowledge and agency, can be effectively threatened or convinced depending on what you actually say?
I don’t say that Inworld AI in itself enables all these things. The demo I saw was impressive, and their Studio product appears powerful and effective, and integrates into common game development environments like Unreal and Unity. But at this point it’s still only an experiment. But experiments are how many great games start out.
Furthermore, having recently raised $10M and counting angel investors from Riot Games, The Sandbox, Roblox, Disney, Animoca Brands, Twitch, and Oculus, it’s clear that there are plenty of people optimism about the possibilities.
Inworld plans to publish a short game that demonstrates the ability of its character generators to tell a compelling story, but it’s still a ways off. For now you can watch the “sizzle reel” above knowing that it’s what you might consider alpha footage — and if you want to give it a shot, apply for early access to the private beta.