This is part of a series of columns written by developers and others speaking at or about the Game Developers Conference in March.
Game developers dream of building gameplay that automatically adapts to the player’s skill level, where no player ever leaves because they got lost, frustrated, bored, or harassed by other players. Game writers talk about designing stories where every choice the player makes, however small or large, affects the rest of the narrative so that every play experience is unique. And builders of sandbox games imagine simulations where you could stop any random character on the street, engage them in conversation, and find that they had a rich backstory and inner life.
Getting to that level of responsiveness and depth requires more than any human development team can afford to construct. If games are ever to achieve (or even come close) to those goals, it will be artificial intelligence that makes it possible.
Content Built with Machine Learning
Deeply responsive games need a lot of content. These days, machine learning-based models are increasingly able to generate content that used to be expensive and time-consuming to produce. Generative AI approaches can fabricate photo-realistic images and textures, set up models of objects in a game landscape, or animate characters automatically.
Text to speech systems are getting better as well, turning a line of written dialogue into a voice performance complete with breath sounds and naturalistic intonation. And once games no longer require every line to be voice-acted in advance, character behavior can be far more dynamic, with characters who decide what to say during gameplay. While the performance qualities aren’t yet at a level to compete with the work of professional actors, companies such as Lyrebird are leading the research.
Even in areas where we’re not yet ready to hand off the whole task of generation to an AI, computational creativity tools can suggest possible level designs, animations, or even lines of dialogue for a human user to fine-tune.
Compelling Characters within Games
Games have long had AI-driven opponents in combat and strategy situations. Storylines involving characters, though, have mostly continued to be heavily pre-authored, every rivalry and romance written in advance.
In recent years, we’ve started to see AI that can track relationships with characters and tweak their behavior to match, or alter the direction of the plot depending on details of player history. Researchers at UC Santa Cruz are exploring dynamic narrative systems that recast scenarios with different characters, or play out different gameplay consequences, depending on exactly what the player has done so far. And Spirit AI is developing Character Engine, a tool to make characters who remember past interactions, react differently depending on their personalities, and actively advance the game’s narrative.
Games That Understand Speech
Alexa and Facebook Messenger have become platforms for interactive stories the player can experience through voice or chat. Currently, those games tend to be simple, and to look for just a handful of possibilities in the user’s input.
But the technology that drives those systems is advancing rapidly: speech to text understanding is more accurate, and our ability to extract meaning from sentences constantly improves. The better we get at understanding the meaning and context of language, the more freely we can create characters who can understand player questions about the game world, or obey the player’s verbal commands.
Recognizing Emotions and Frustration
Language isn’t the only kind of input getting a boost from AI. Affective computing systems let us guess emotion from whether the user shouts or whispers, smiles or frowns while speaking. Using that data has challenges, especially because different cultures express emotion in different ways and with different intensity — so we have to be careful about the assumptions we make when we deploy these systems. But the existence of the tech opens the possibility of characters who show greater empathy.
Other AI projects look at how players are acting during the course of normal gameplay, just from the frequency and speed with which they’re moving through the game space. With these cues, the AI can estimate whether the player appears to be frustrated, stuck, or lost. That allows the system to adjust difficulty levels, or add visual cues in the environment to lead the player to their next destination.
Moderating Game Communities
Another area for AI involvement in the game industry is in moderating communities of gamers. Like other areas of social media, the chat systems surrounding games are subject to bad actors: con artists looking to defraud their fellow players; hate speech, bullying, and sexual harassment; and bots designed to spam other gamers with advertisements. It only takes a small number of poorly behaved players to make an entire community feel less friendly or welcoming.
Many game companies employ teams of moderators to try to remove bad content from their communities. But it’s difficult to keep up with the volume of chat on popular game systems, which means that a lot of material goes unnoticed and uncorrected. Meanwhile, the moderator’s job is to spend all day looking at the worst interactions in their communities. The job is at best tedious, and at worst traumatic.
AI can help here as well, by detecting tendencies in community data. Spirit AI’s Ally tool is designed to discover patterns of bad interaction in the target community, and triage the results, so that moderators can see immediately which members of the community are responsible for the most incidents of fraud and harassment. Ally can also find and surface positive information about the community — like which players are most thanked by other players.
Offering moderators a view of both the best and the worst factors in their community makes their work more effective and more pleasant.
Most of the major AI advances of recent years rely on large supplies of data, and the ability to interpret what that data means. As we’ve seen in other industries, companies have a strong incentive to gather as much data as possible, often from users who don’t realize that they’re giving it away. Meanwhile, the way we interpret that data is subject to the cultural biases of the culture that collected the data set.
Economic justice should also concern us. AI data labeling is often a dull task for which researchers pay low rates to workers on Mechanical Turk or similar platforms. At the same time, the existence of AI solutions may displace other, better-paid workers. For instance, the ability to clone voices for vocal performances is exciting, but we need to consider how we credit and compensate the voice actors who allow us to build these systems.
Is AI going to take all the game developer jobs?
For all the advances in this space, AI is unlikely to drive human game designers, writers, or community managers out of their jobs in the near future. We still very much need human innovation, supervision, and curation, as well as human insight into what we should be building.
Emily Short is the chief product officer at Spirit AI, and one of the advisors to the AI Summit at GDC. She has been working with game narrative and AI-driven characters since 2000, and has written for dozens of games, most recently Failbetter Games’ “Sunless Skies.”