Last year, AI-generated art finally broke through the mainstream—but not without significant public controversy. The rampant art theft required to build an AI’s dataset and the resulting forgeries eventually led to a class action lawsuit against AI generators. Yet that hasn’t stopped developers from using the technology to generate images, narrative, music and voice acting for their commercial video games. Some game developers see the technology as the future, but caution against over-selling its benefits and present capabilities.
AI has been making headlines lately for the wrong reasons. Netflix Japan was blasted by professional artists for using AI to make background art—while leaving the human painter uncredited. Around mid-February, gaming and anime voice actors spoke out about the “pirate” websites that hosted AI versions of their voices without their consent. AI seems to be everywhere. One procedurally generated game has already sold millions of copies.
A few years ago, Ubisoft Toronto, known for games like Far Cry 6 and Tom Clancy’s Splinter Cell, was not only using AI in its development process—it created an entire design system that heavily relied on procedural generation. “In the future — potentially as soon as 2032 — the process of making digital nouns beautifully will be fully automated,” Ubisoft director Clint Hocking wrote in a Polygon op-ed that claimed that within a decade, players would use AI prompts to build their games. Think “a side-scroller where I am an ostrich in a tuxedo trying to escape a robot uprising,” as Hocking put it. This futuristic vision of games would work in the same way you might tell AI image generator Midjourney to produce new images based on text descriptions.
Despite the eyebrow-raising boldness of his claim, the industry has already seen some strides. Watch Dogs: Legion, Ubisoft’s open world action-adventure game, seemed impressive for what it was: A blockbuster title that randomly generated NPCs in every playthrough and promised to allow players to “play as anyone.” While reviewers did encounter “repetitive loops” in the quest system, Legion seemed like a solid first step in the future of procedurally generated gameplay.
“10 years [to create an AI game] is insane, as it takes 5 to 10 years to make a standard AAA game,” said Raj Patel, former product manager on Watch Dogs: Legion. He was wary of how designing non-linear games incurred an additional layer of labor-intensive complexity. He told Kotaku over messages that he didn’t think that AI games could be “wholly original, bespoke, [and] from scratch with the same quality” as existing AAA games. “There is certainly potential [in machine generated games], but Star Citizen has been in development for 10 years so far,” he said of a space sim MMO that boasts of procedurally generated planets. The game has raised nearly $400 million, but has not been released since it was first announced in 2010.
If Ubisoft’s forays into NFTs and web3 are any indication, the company has been quick to jump on trends that sound buzzy to investors. But that didn’t mean that they were necessarily pushing the technology forward.
Game designer and AI researcher Younès Rabii felt that integrating AI with these expensive processes was more about “hype” than a technological inevitability. “There’s always a 15 to 20 year gap between what academia has produced in terms of [AI] advances and what the industry actually uses,” Rabii told Kotaku over Zoom. They had strong feelings about how Watch Dogs: Legion seemed to fall short in being the public face of what AI games could be. “This is because it’s way too long to train [developers] to use [advanced AI]. It’s not worth the risk. It doesn’t bring enough money to the table.” Ubisoft told investors that the game’s predecessors have sold around ten million each, but never publicly released the sales data for Legion beyond its launch period. They felt that Ubisoft had taken the risk with Legion as a marketing hook. “It’s not that interesting… they have a series of simple nouns and properties, and they behave according to it.”
Reviewers seemed to agree with him. One critic noted that “there’s not much of a human element” to the Londoners in the game, and that they “don’t meaningfully interact with each other.” Another struggled with “repetitive” missions. Kotaku panned the campaign for being “empty and soulless,” but praised the more interesting DLC for ditching the procedurally generated recruitment altogether.
Hocking himself admitted in a Washington Post interview that “reinventing open world design” during Legion’s development had been “uncertain,” “difficult,” and “scary. Being able to play as any character in the game was an idea that Ubisoft had never experimented with before.” Human designers had to manually account for every single possibility that the players could choose—it wasn’t a computer that could understand how human players would emotionally respond to randomly generated scenarios. Hocking had been much less optimistic about the possibility of creating a gameplay experience that didn’t feel entirely samey. “There isn’t infinite diversity,” Hocking said in the interview. “You’re still going to encounter, ‘Oh, yeah. I recognize that voice. I recognize that person. Or, this is one of the people who has the technician fighting style. They fight in a certain way, [similar to] that other person.’ But it still blurs the lines quite a bit.”
Florence Smith Nicholls, story tech at the award-winning indie studio behind Mutazione, also had a more muted perspective of AI. They told Kotaku over video call that AI was already being used extensively in AAA development, like in Fortnite. “When people [say] it’s going to completely revolutionize gaming, it feels kind of similar to what we’ve had with discussions around NFTs and the blockchain.” They pointed to the chess playing program Deep Blue as an example of artificial intelligence in gaming.
Mostly, though, we’ve seen a wide range of applications for AI in games when it comes to automation–but how we define such a thing can get confusing for the average person. Because of popular generators such as Midjourney and Chat GPT, most people associate them with neural networks that create text or images based on a dataset that it scrapes from the internet. Researchers have very broad definitions of AI. “If you showed someone Google Maps in 1990 and showed that you could plot a route between any two points on the planet… that would be considered a hard AI problem,” said Cook. “Now people just think of that as something that your phone does. It’s the same thing in games. As [technology] becomes more normal, they no longer look like AI to us.”
“We talk about AI when it doesn’t work,” said Alan Zucconi, director of computer game programs at Goldsmiths, University of London. “When it works, it’s invisible and seamless.” He acknowledged that artists and programmers don’t see eye-to-eye on the technology. “There is friction [with AI], especially for artists… Those same artists are using AI every day, they just don’t call it AI,” said Zucconi. “Tools like the select all regions tool in Photoshop, smudging colors… tools we take for granted are not seen as AI… so I find it very fascinating when people think that these are something new. It’s not.”
“The real utility [of AI] in the short term is helping with more discrete tasks in the process of producing work,” Patel wrote, recounting his experiences with working on Ubisoft games. “In one game, we had AI testing the open world… It would log the framerate and any clipping issues. The machines would be left running moving through the world and note areas where things had issues. That helped us find areas to check without having real people have to do that otherwise tedious work. Real people could focus on checking, verifying, and figuring out details.” Rather than risking whether or not a player might be able to tell if something was AI-generated, “[AI] let our QA staff not do the tedious parts and focus their time more efficiently on problem areas.”
Automated development often sounds incredibly sinister when coming out of the mouth of a gaming executive who doesn’t sound adequately troubled about the plight of crunching developers. But testing has been automated for years, and QA professionals are calling for studios to ditch fully manual testing. Despite the popular image of QA as low-skilled work, AI experience is often a necessary prerequisite to being a games tester, because automated testing is often a key aspect of a studio’s workflow. And it’s not just testing—automation is a shipped feature of AAA video games too.
Mike Cook is an AI researcher and game designer at King’s College London. He told Kotaku over a Zoom call that games such as Minecraft are procedurally generated by AI, and blockbuster games such as Assassin’s Creed makes use of AI for certain mechanics. “When your character places their hands and legs in unusual places to climb up the side of a building, that’s not a handmade animation,” he said. “There’s an AI that’s helping figure out where your body’s limbs should go to make it look normal.” He noted that online matchmaking and improving connectivity were both aspects of games that were supported by AI.
Despite the possibilities, Nicholls said that procedurally generated content was only really useful for “very specific tasks.” They cited examples such as changing the weather or generating foliage in Fortnite. AI would need to be able to handle several different tasks in order to be considered a game-changing force in development.
However, they had concerns about which developers would benefit from extensive automation. They pointed out that in the case of art outsourcing (the practice in which studios pay cheaper studios to create low-level assets), the “main” studios were doing more “intellectual work” such as design. They thought that AI could similarly create an underclass of artists whose work is less valued.
Sneha Deo, an AI ethicist from Microsoft, draws the connection more overtly. “I would say a lot of the undercutting of [tech labor] value that happens today is due to differences in the value of currency.” It’s cheaper to hire developers from a country with a less powerful currency, rather than paying developers from the U.S. or western Europe. She also attributed the devaluation of human labor to the last mile effect. “Humans trick themselves into thinking if a machine can do it, then the [labor] that the humans are adding to it isn’t as valuable because most of it is automated.” So even if AI created new ‘AI design’ jobs, those jobs might not necessarily pay a reasonable amount.
While he’s normally exuberant about the possibilities of machine learning, Zucconi seemed uncomfortable when asked about whether or not AI would devalue the labor of voice actors. When directly pressed about the possibility of paying actors for using their voices in AI (as Hocking raises in his op-ed), he said: “Licensing voices is probably going to happen. We’re very close to having that technology… I’m hopeful that this is a good future because it means that people can have more work opportunities.” The ability to commercially profit from one’s own “likeness” is enshrined in state publicity laws. Celebrities have been licensing their likeness to third parties for years—the most famous recent example being Donald Trump’s embarrassing foray into NFTs.
Despite his optimism, it seemed that professional voice actors felt differently. Voice actors for popular franchises such as Cowboy Bebop and Mass Effect both spoke out against AI versions of their voices being falsified and used without consent. Some bad actors had even used AI-generated voices to dox people. It’s reminiscent of how decades ago, Jet Li turned down a role for The Matrix because he was concerned about Warner Bros. reusing his motion-captured movements after he collected his last check.
“I think what matters is not any specific deal,” Cook said in regards to compensation and AI-generated art. “I don’t know if licenses are better than labor. What does matter is that the people who are actually doing this job are the ones that get to decide what should be happening,” he said. “And the problem is that in most of these creative jobs, the power dynamic isn’t there to allow people to have that voice.” He also noted that it was easy for artists to accidentally sign away their rights in perpetuity.
Unlike blockchain technology, developers can see clear benefits to adopting automation more broadly in game development. One indie developer told Games Industry that AI development could help smaller studios stay competitive. Failure rates are incredibly high, especially for developers who don’t have massive AAA-sized budgets. No Man’s Sky used machine-generated content to create expansive worlds, only to have a disastrous launch–and it took five years for the game to eventually become a success story.
Deo saw AI as one method of bridging the resource gap between the global north and south. “What’s the rightness or wrongness around using these models to generate art or narrative or text if that’s not your strength? I think about game design as this collaborative process that favors people who already have strong networks,” she said over Zoom video. “[These people] can tap their friends or their networks to come in and do that manual work, [which] is democratized by the replacement of human labor by AI art.”
Deo acknowledged that AI art could undercut junior artists who were trying to break into the industry, but thought that it wasn’t an ethical quandary that should rest on independent creators. “It’s not a black and white thing. I think at larger studios, that’s a place where there’s an ethical issue of: ‘How does this undercut labor that’s already undervalued?”
It was a convenient way to think about AI in a positive light. But AAA games like Fortnite have already taken “inspiration” from indie games such as Among Us. That was just for a game mode. It didn’t feel like a logical leap to think that big studios could borrow development methods too.
And there’s another major stakeholder that’s critical to the success of AI games: the players. Right now, the average person still thinks that “human” and “machine” generated art have inherent differences. “There’s a sense of difficulty in knowing the authorship of certain artwork,” said Nicholls. While games are often attributed to leads in more public-facing roles, they are products of entire teams–and AI only complicates the idea of authorship. Especially when generators such as Midjourney are raising legal and ethical questions on who owns the art that the machine produces. “I wonder if now there’s more unease around AI because people fear that they won’t be able to tell if something is AI generated or not.” Before AI became a prominent image-making tool, it would be reasonable to assume that any painting had some kind of human element. Now, even Bungie community moderators struggle to differentiate between AI and human art.
But Cook thinks that these machines we call “video games” contain a complexity that can only be built by humans. “Maybe it’s possible for AI to generate games but the games that left an impact on us… they’re boundary breaking. Concept breaking. Those are things we can’t necessarily predict with enough data or computer power… If we wanted infinite Grand Theft Auto campaigns or Star Trek episodes, then they would start to feel samey.”
Nevertheless, games such as Minecraft and No Man’s Sky are immensely popular. Although the popular image of artificial intelligence is associated with perfection, that’s not what Cook thinks that gamers necessarily want.
“Players like to be surprised. They actually like it when the AI breaks…Some of the most memorable things that people pull out of these AI systems is when they’ve gone wrong a bit. But I think something that’s really important is that they like to be able to share and talk about these things,” he said. “Although Minecraft or Spelunky 2 has an infinite number of levels and worlds in it, that infinity isn’t really important. What’s important is the one world that you have, or the one thing that you shared with other people. So in the Valheim world, the Valheim world generator is not important. What’s important is the server that you built with your friends.“
Correction at 2/17/2023 at 9:46 A.M. E.T.: Corrected a name and job title.