Researcher: AI can make games more personalised if developers let go of control

Illustration: No Man's Sky

As game engines get closer and closer to achieving photorealism in real time, the question of what will be the next goal for the gaming industry arises. Artificial intelligence (AI) entails a great deal of unexplored potential.

Most people associate AI in computer games with how difficult it is to fight the computer-controlled opponent, but with artificial intelligence there is a whole host of innovations that can not only change how we play games, but also how games are created.

According to Julian Togelius, associate professor of Computer Science and Engineering at New York University, the games of the future will have much more AI-created content than we are used to.

“We’re going to see games that figure out what you’re good at, that give you more levels that suit your playstyle, or that challenge you to change your playstyle. A system that looks at what you do in the game and then gives you what you want—even though you may not have known it yourself,” Julian Togelius says.

He has spent most of his working life finding the connection between getting “games to make AI better and AI to make games better,” as he puts it. And he believes that there is a future for more personalized content with, for example, adaptive level design.

“We have a personalised feed when we go on Facebook, Twitter, or Spotify, for example. It’s all shaped by your behaviour. Why is it not like that in games?”

Games that resemble the YouTube rabbit hole

Many people probably know the feeling of going down a YouTube rabbit hole, while the algorithm keeps suggesting new videos that you just have to watch. It is pretty much the same vision that Julian Togelius has for the computer games of the future.

“Imagine that in a couple of years you’ll be playing Grand Theft Auto 7 or something like that,” Julian Togelius says.

“You may have just stolen something in the game and are being chased by the police. So you think, ‘Where should I go?” And then you just drive off the country road until you come to a city with a completely new type of architecture, brand new kinds of cars, a different kind of fashion. In every building you enter, there is something worth exploring. There will be people you can talk with a unique dialogue. There are new plots and thus new missions. And all that content is presented to the player based on what the algorithm expects you to like.”

Illustration: Julian Togelius

Lack of data

It is not due to technical limitations that we are not yet closer to the above scenario, Julian Togelius says.

“One of the big reasons why the gaming industry has not quite gotten there yet is that it has not fully understood how much you can do with the data you have. User behaviour has been studied to find out what types of players will spend money, and it has otherwise been pretty simple computer science. But virtually all games collect data for the game developers,” Julian Togelius says.

There is also a challenge in that it is difficult to access large enough amounts of player data because the game developers consider it a trade secret and keep it confidential.

“There are also some privacy challenges involved. We have seen a lot of research that suggests that you can make some predictions about people based on their game data. For example, you can see how a player interacts with Minecraft and predict how he will react to things in the real world based on that. You can get an insight into their personality, and maybe also age, gender, and so on,” Julian Togelius says.

Creative control

Another big challenge is that game developers fear that the AI will take the creative control away from them.

“A lot of AI researchers tend to see games as a problem to solve, where for game developers, it is more about designing an experience. So you often see that they misunderstand each other,” Julian Togelius says.

“Game developers most often see what they do as an art form—and it certainly is, at least to a certain extent. But they fear they will lose the magic if they go down a data-driven path. They want control over what happens in the game. And if you work with AI, then you also have to let go of some control.”

Isn’t it also a problem if the games always just give you want you want? Shouldn’t the players be challenged?

“Yes, that’s right. But then the AI model is wrong,” Julian Togelius says.

“If the AI model has seen you play the game 10,000 times and it has seen that you like to cook soup in this game, then a stupid model will just give you more ways to make soup. Whereas a good model will recognize that it may be time for you to try a different type of challenge. It will know that players who like to cook soup often also like to dance, for example.”

Aren’t you afraid that games are going to lose the feeling of not being mass produced?

“There are many ways to answer that question. Now I choose the slightly provocative answer: They have already lost their feeling of not being mass produced. It is already the case that there are very small indie games that explore a small topic and then there is Call of Duty. And there is room for it all. Games are more diverse than ever. But in a way, I also think that the ability to create as much as you can with AI will give small studios the ability to make just the games they want. They can outsource all the grunt work and focus on the parts where they can express themselves.”