What AI Can, Can’t, and Shouldn’t Do for Games

And this morning we're going to take a look at what AI can, can't, and shouldn't do for games. And for what it's worth, generative AI did not make this presentation. So I've been coding and making games since I was 11. So it's over 45 years. I'm 15 in this picture. I'm rocking a double monitor Apple 2 plus setup. And in my 40 plus year career as a technologist, I believe that the advent of AI is the single greatest technological change since the invention of the Internet.

It's caught many by surprise and, mind you, I've been around since, you know, a long time. So, this was pretty much the most amazing thing that a computer could do back then. Hello, world. So what do we mean when we say AI? AI has actually been around since the dawn of computers. But when we're talking about AI today, the stuff that's setting the tech world on fire, what most people mean is generative AI, the type of AI designed to generate new content, text, images, music, or other data types based on the input that it receives. And not just generate it, but generate it at a level that challenges, but not yet passes the Turing test. To even challenge is pretty incredible. AI is now amazing, is capable of really amazing things.

So you know, immediately everyone in finance is like, yeah, let it rain money! And we'll have generative design and code and art and music and Q&A and we'll empty out the offices and we'll keep all the money! Bam, bam, bam, bam! But you know, that didn't go so well, unsurprisingly. So let's examine how game developers are using AI right now. So in January of this year, the Game Developers Conference, which is the largest conference in my field, released their State of the Industry Report. And for the first time, AI was a topic of the survey, which shows you how fast the field has evolved. And they asked participants about their use of generative AI tools, such as GitHub's Copilot, ChatGPT, and Adobe Generative Fill, some real basic ones there. Nearly half of game developers said generative AI was being used in their studio. And 31% of people noted that they were using AI themselves. 18% said that they didn't, but their colleagues did. And 15% said that while they didn't use it, their studio was interested. And the remainder really had no interest or they were unsure. And answers also varied depending on people's roles in the company. 44% of people in business and finance used AI tools, compared with 16% for those in AI-related roles, and only 13% for those in narrative. And what about those who didn't use AI tools? Well, 21% of AAA studios, typically creators of big budget blockbuster games, probably over $75 million to create, they prohibited the use of AI entirely. For smaller developers, this number was 12%. And 7% of developers allowed some tools, while other studios just made it optional.

So what about at Romero Games, my company? Well, we use no generative AI in games at all, whether in art or code or audio or writing or designing, and we're pretty firm about that. AI can sometimes be useful, obviously, as a super Google for research, for parsing thousands of links into a potentially definitive answer. But there's still a lot of work to do, and AI has told me that I was born in Guatemala. And I wasn't. I was raised by my Mexican grandmother. But at least she was Mexican, and that would have been pretty great, though. Evidently, I appear in music videos, which is news to me. And best of all, I'm the son of horror movie maker George Romero, if you didn't know that. Which given the names of the games I make and the movies he makes, well, that's kind of plausible. But the truth is that these things aren't true, and obviously that's a danger, and I'll come back to that later on.

So why haven't game developers like my company, Romero Games, jumped on the generative AI bandwagon? Well, given the cost of game development, which is averaging, you know, $75 million at minimum for the average AAA game, Spider-Man 2 came out not that long ago, $352 million to create that game. So the really big budget, and the idea of having AI create your content at least seems like a no-brainer. However, when surveyed, a full 84% of game developers said that they were very or somewhat concerned about the ethics of AI use. Only 12% said that they were not concerned at all. And if anything, people wanted to see AI automate processes to make their creative work easier, not to take that creative work itself away. That's why we're doing what we do.

So even more than that, we don't want to potentially impinge on the rights of others. Now, this is not to say that game developers don't use AI. We definitely do. Tons and tons of AI, and we really have for decades. If anything, people wanted to see AI automate processes to make the creative work easier.

So let's take a look at what AI can do for games. So in 1993, you probably know what this game is. In Doom, cacodemons used path-finding AI to traverse the level to find their way to the player or another enemy that caused them damage. And we used algorithms like ASTAR to find the shortest path from one place to another. So in Red Dead Redemption 2, NPCs used finite state machines to manage enemy and NPC behavior, whether they patrol, or they identify the player, they attack, they flee, draw a card, or drink beer. All of that is up to the AI, and they transition based on certain conditions or events. For instance, if they hear a gunshot, they might duck for cover. And in complex strategy games, decision trees are commonly used. Decision trees help agents in the game make decisions based on specific conditions in the environment. They use hierarchical structures that evaluate multiple conditions to choose an action. And what resources are they gonna use? You know, what are their short-term combat tactics? What's the long-term combat strategy against the enemy? And decision trees in games like Age of Empires 4 can be substantial, especially when you have probably like 12 different factions all fighting each other and having to decide how to do that. Behavior trees allow us to create more flexible and modular AI behaviors. And in 2004, Halo 2 used behavior trees to manage complex behaviors of its enemies. Basically, behavior trees are composed of nodes that define various behaviors like move to cover or throw grenades or, you know, engaging the player in combat. And these nodes can be reused across a variety of enemies, making it easier to implement diverse behaviors without redundant coding and design. And though less in use now, utility AI is another method of decision-making AI for enemies and NPCs. So agents possess a defined set of actions, each of which is evaluated and they have a score. And according to its current usefulness in the existing state of the world, there's basically a formula that determines which way the AI should go. Racing games or games that require realistic movement use steering AI to control the direction and speed of NPCs to achieve goals like fleeing or seeking or avoiding obstacles.

It's often used in simulation in racing games like Forza for natural and fluid movement. And while not yet in use in traditional commercial games, applications like Google's DeepMind and AlphaGo illustrate the challenging potential of AI versus the human brain. In games like Minecraft use procedural generation to create never-ending worlds. And the game basically has its own crafted algorithms to generate content, like in this case you can see trees and there's dungeons and animals and flora and fauna, all based on random seeds or predefined rules.

Games like Baldur's Gate 3, Game of the Year for last year, and pretty much every game with NPCs or missions uses scripted AI. Scripted AI allows us to control specific sequences and events and there are prewritten scripts that dictate NPC actions and events based on predefined triggers. And one of the more interesting uses of AI is design-mediated AI, where design works in conjunction with AI and these can have particularly big payoffs for gameplay. So, as an example, in 1992's Wolfenstein 3D we had what we referred to as sound zones. Now normally the AI would not decide that when I open a door in a room full of enemies that it should also activate an enemy on the other side of the entire map. But that's not something AI would ever think of doing, obviously. But for designers, it provides a tremendous sense of suspense for the player. So hearing that something is coming towards them when doors are opening and closing and they're getting closer and closer to them, and that's a design decision that uses AI to make it happen. But AI didn't make it happen on its own.

And another interesting example of this is in the Civilization series, as Sid Meier, the game's designer, told the story. A tester was playing the game and he had a 33% chance of winning against a particular group of enemies. And he attacked once and he lost the skirmish. He attacked again and he lost the skirmish again. Attacked a third time, lost a third time. To this tester, this was unfair and it was a bug. Because if he has a 33% chance, he thought that he should have won at least one of those battles. Of course, that's not how random number generation works. But Sid didn't want to disappoint players. And so he reasoned that even if the AI did not choose to do that for the player to win, he could override the AI and improve the player experience and make them happy. So he forces a win on the third try. But the difference between AI and game developers, the AI that game developers use and generative AI, is that we built or we licensed this AI according to using our tools from our creativity. And middleware companies like Havoc have made a complete business around licensing AI and other tools for game developers. Among other things, game developers can license all their navigation AI from Havoc, for instance. And Havoc is just one company among many offering such solutions. And while AI can do a great many things, there are definitely things it can't do.

And while AI can do a great many things, there are definitely things it can't do. So let's go back to 1993. There were five of us discussing our next game. Me and Adrien, John, Tom, and Kevin. And we knew only one thing. We wanted to make a game unlike anything that we had ever played. And so on January 1st of 1993, while people were still probably recovering from their New Year's hangover, we sent out a press release, ID Software to Unleash Doom on the PC. Okay? So we listed all of the game's features. And you can find the press release online if you want.

We declared with a shocking level of confidence that in 1993, we fully expect to be the number one cause of decreased productivity in businesses around the world. And it's nuts that we said that, but at the same time, we hadn't even started working on the game yet. But at the same time, we had our own AI, our actual intelligence. A database of games and gameplay in our heads from years. We were hardcore gamers and we knew first and foremost that what we were creating was a game design that no one had ever seen before. We knew no one could deliver the tech to deliver that design. And we knew that when we pulled it off, we couldn't imagine playing anything better at all because we couldn't. And that game became Doom.

The point I'm getting at here is that AI, built on the shoulders of all our existing knowledge, is unlikely to create something great from nothing right now. AI is by its very nature forced to be derivative. It can do amazing things. It can find new uses for medicine based on known interactions and benefits. It can parse data faster than humans and suggest interesting things. It can deliver in seconds the combined knowledge of 10,000 or more links and just save a ton of time. But right now, I don't believe that it can innovate, really and truly innovate. Find that light in the darkness, so to speak. Every game designer, every programmer, every artist, everyone knows that something like Doom was really a once in a lifetime event. The right people, the right place, the right time, the right design, the right art and tech. I'm incredibly fortunate to be there, but I just don't think that AI can do that yet.

So I asked chat GPT, I need you to create a pitch for an FPS game that's truly innovative and can compete with titles like Battlefield, Fortnite, PUBG, and so on. And with the world at its digital fingertips, it had this to suggest.

Echoes of Eternity is a groundbreaking first person shooter that merges intense combat with a deeply immersive narrative. Procedural world building and advanced AI technology that aims to push the boundaries of the FPS games by integrating a dynamic storyline that evolves based on player decisions. A living, breathing world that adapts to the actions of its inhabitants. And a cooperative multiplayer mode that emphasizes teamwork and strategic planning.

So it takes more than saying groundbreaking to be groundbreaking. Pushing the boundaries, that's easy to say, it's not really easy to do. And the reality is that basically all of this has been done. It's been done together in a single game. It's been done in an FPS. And even looking at the detail of its feature list, it just wasn't hitting any kind of high mark or innovative mark. So I tried again. I said, hi AI, I hope you're well. I was polite. I co-created the first person shooter. I'm not impressed with this answer. It's not really innovative at all. All of this has been done before. As usual, JattGBT apologizes. Okay, let's take another approach. And push the boundaries further. How about combining unique gameplay mechanics and technological enhancements that truly set your FPS apart from existing titles?

I can't wait. Okay, so here we go. Echoes of Eternity is a revolutionary first person shooter that defines the genre through groundbreaking integration of quantum mechanics, real time emotional AI, and immersive multi sensory feedback. This game creates an unparalleled experience where players manipulate time, interact with emotionally intelligent NPCs and feel the physical sensations of the battlefield.

Okay. All right. Yes, it's true that no one has done this before. This is actually innovative. But does anyone want it? Is it technically feasible? How many people have a suit, I'm guessing? So that they can feel the physical sensations of the battlefield. How many people? I don't know. So what do you mean by quantum mechanics? Yeah, are we going to split time and space? Because that's already been done in a game before. In fact, even the title has been done before. It's published by Games Workshop. It's also a heavy metal band, a book of religious contemplation, and an actual game that was up for pre-order four years ago.

Yeah, I don't want to put down all the amazing things that AI is capable of, but for me, I'm betting on the right people at the right time, at the right place, with the right idea. But let's just say that, for the sake of argument, that lightning did strike, and AI did, you know, found an amazing standout creative idea. And that brings me to what AI should not do. The challenge with AI, at least in creative fields, is that it's being generated based on aggregated data.

So the idea of who owns the copyright to these outcomes is a matter that's incredibly contentious right now. It's just too risky in any creative field, particularly when you're creating brand new IP. And even Steam, which is the biggest online platform for PC games, looks to be limiting its own liability by forcing developers to disclose that they use AI in anything being uploaded to Steam for sale. And what happens to your data, you know, your questions, is that also feeding into a data set? Will it show up in someone else's search? There's certainly ways to prevent that, but I wonder if that's something everyone knows. And more than that, more than any of this, you know, all of us do what we do because it means a lot to us, and when I was young, I wanted to do nothing more than make games, code all night, listen to heavy metal. Finally, the length and color of my hair has changed. I don't want anything that's going to take away that fun from me, and I look forward to a future where AI is an exceptional tool, not like the alpha and omega of the creative process. And as a technologist, I think what's happening now is the single greatest technological leap since the invention of the Internet. It's not hyperbola, it is groundbreaking, it is revolutionary, it is literally game-changing, and I'm really eager to see where we will take it. I want to discover new medicines, crunch incredible amounts of data, but I don't see it creating the next Minecraft, Fortnite, or Doom. Not yet. Thank you.

Thank you. I do have a quick question of myself, right? One thing which I really loved is the fact that you were talking about how much the different technologies and how they've been uptaken by different teams within companies. For example, you saw the artists were different to maybe some of the business people in the games companies. Do you see any ideas of which parts of companies would have to have different approaches to the way they use AI, especially in the creative side versus maybe the more analytical or just mathematical side?

Yeah, I mean for narratives, you're not going to say, hey, make me create dialogue between two characters and then have it spit it out and then use that. You just wouldn't do that. For narrative, I think with narrative, some of the stuff that we've used it for is to dump a bunch of text in and just say, summarize this for me. Just tell me what this is, TLDR, you know. Basic stuff like that. For code, we don't even use it for code at all. It's just we have really good programmers. They don't need AI to help them with design. I don't see us using it for design. We use it just for, I guess, taking something and just elaborate on this thing, on this concept, or what is this thing, what's the best, has this ever been found? Instead of using Google to find something and search through a million links, we can just say, does this exist, and kind of go from there. I love that actual rubber duck, kind of when you're ideating, which is really cool. Like the super Google, yeah. For sure, for sure.

Okay, we've got some questions that are coming in. I'm going to save the top one for last, but don't worry, I haven't missed it. What's the level of intelligence needed to call something AI in game dev? Because for example, it seems to be different, because we've had game AI for years now, so what sets that apart? Well, AI is, game AI is basically a set of rules that entities have to obey, and the more rules and the more complexity you put in there, the more interesting the AI feels like. But you don't want, in a real-time game, in an FPS or a third-person game, you don't want the enemies to be too smart, because then you can't even fight them. They're too difficult. So if you've noticed, you could have 20 enemies coming towards a player, and we use things called ticketing systems that kind of hand out tickets to the enemies to say, you can attack, and you can attack, and you can attack. Everyone else, act like you're doing something. So you have to purposely dumb down all the AI so players can even aim at them and shoot them. You can't be super intelligent. There was the very first Unreal game that came out in 1997, had enemies that would drop and roll and do cover, and players were like, what do I do here? It was just too complex. And so you have to keep it really simple. But the AI is something that obeys all these rules that we create to make it look like there's a lot of stuff going on, and also intelligently using behaviours against the player and all that stuff, but it's not using anything generative or something from outside itself. Could you tell the Elden Ring developers to dumb it down for me? I know. I'm struggling. Yeah.

All right, let's go on. We do have another question. You said, for example, that you separate, you don't use AI to help write code, and someone's just asking, what's your opinion just generally about code helpers like a Copilot and some of the other tools that are out there? They're great. I think that they're really great to use, especially for junior programmers that are just learning everything. Everyone has used intelligent agents or helpers while programming, even from the late 90s, where it can just look up the function call for you or autocomplete for you and all that. That's all super helpful. Even with Apple's Xcode, they have a thing called Playgrounds where it's actually showing you what you're typing and what it looks like. So if you're making a for loop, it will create a graph as you're typing it. So all these helpers are really great to help you understand what you're doing, and if you don't know the functions to just look it up for you. All of that stuff is great because that just means you're learning more, but you're not depending on AI to do your job.