Creating a video the game requires hard and repetitive work. How could he not? Developers are in the world-building business, so it’s easy to see why the games industry would be excited about generative AI. With computers doing the boring stuff, a small team could whip up a map the size of San Andreas. Crunch becomes a thing of the past; release of the games in a finished state. A new era is dawning.
There are, at the very least, two interrelated problems with this narrative. First, there’s the logic of the hype itself – reminiscent of the frenzied gold rush on crypto/Web3/metaverse – which, consciously or not, seems to view automating the work of artists as a form of progress.
Second, there is the gap between these statements and reality. In November, when DALL-E was seemingly everywhere, venture capital firm Andreessen Horowitz published a lengthy analysis on its website touting a “generative AI revolution in games” that would do everything from reduce the development time until changing the types of titles being created. The following month, Jonathan Lai, partner of Andreessen, published a Twitter feed exponent on a “cyberpunk where much of the world/text has been generated, allowing developers to move from asset production to higher-order tasks like storytelling and innovation” and theorizing that AI could enable creation “good + fast + affordable” games. Eventually, Lai’s mentions filled with so many irate replies that he posted a second thread acknowledging “there are certainly a lot of challenges to be resolved.”
“I’ve seen some frankly ridiculous claims about things that are supposed to be imminent,” says Patrick Mills, acting head of franchise content strategy at CD Projekt Red, the developer of Cyberpunk 2077. “I’ve seen people suggesting that AI would be able to build Night City, for example. I think we’re a long way from that.
Even those who advocate generative AI in video games believe that much of the enthusiastic discussion about machine learning in the industry is spiraling out of control. It’s “ridiculous,” says Julian Togelius, co-director of the NYU Game Innovation Lab, who has authored dozens of articles on the subject. “Sometimes it feels like the worst kind of crypto brethren left the crypto ship as it was sinking and then they came here and said, ‘Generative AI: start the hype machine.'”
It’s not that generative AI can’t or shouldn’t be used in game development, Togelius explains. It’s that people aren’t realistic about what it could do. Sure, the AI could design generic weapons or write dialogue, but compared to text or image generation, level design is diabolical. You can forgive generators that produce a face with wobbly ears or a few lines of gibberish text. But a broken game level, no matter how magical, is useless. “That’s bullshit,” he said, “You have to throw it away or fix it manually.”
Basically, and Togelius had this conversation with several developers, nobody wants level generators that work less than 100% of the time. They render games unplayable, destroying entire titles. “That’s why it’s so difficult to take generative AI that’s so difficult to control and integrate it,” he says.