Microsoft Just Accidentally Proved Why Generative AI Isn’t The Future Of Gaming
There’s still nothing impressive about generative AI.
Microsoft and Nintendo have been working together more lately, and Microsoft took the opportunity last week to take the heat off its new pal when it needed it most. Just when online discussions about the worst parts of the Switch 2 were at their peak, Microsoft stepped in to provide an even easier target by revealing a generative AI demo that might be the worst example of an already widely hated technology we’ve seen in some time. Pitched as a version of Quake II “reimagined” by AI, the demo instead feels like an object lesson in exactly why generative AI for game development is at best a joke and at worst a threat to the industry itself.
Microsoft calls the tech demo, which was made with the company’s Muse AI model, an “AI-powered gaming experience,” which “dynamically generates gameplay sequences inspired by the classic game Quake II.” The awkwardly worded descriptions are working around the fact that even Microsoft isn’t bold enough to actually call the experience a game.
Quake II has been ported to more than a dozen platformers and operating systems at this point, but Microsoft’s demo is decidedly not a port. It’s a simulacrum designed to look and act like Quake II, without actually being Quake II. And while at a glance its visuals do in fact seem “inspired” by the classic shooter, it’s quite a stretch to say the demo resembles the game in any meaningful way. Like any product of gen AI, the Copilot Gaming Experience, as Microsoft calls it, is cobbled together by an AI model that attempts to deliver a convincing approximation of what Quake II looks like in every frame. The problem is, like everything else the games industry’s current favorite shiny toy has ever produced, it does a pretty awful job of it.
Rather than building a coherent game, as gen AI’s supporters claim the technology will be able to do one day, the Muse AI model is just taking whatever is on screen at any moment and generating a plausible image to follow it in the next. In videos circulating online, you can see for yourself how that turns out, as the model essentially forgets everything it’s not currently looking at. That means that simply looking at the floor and back up causes the model to regenerate a plausible room based on Quake II footage it’s been fed, resulting in the entire room simply being replaced by a different one more often than not.
The Copilot demo has a lousy frame rate and incredibly blurry textures, but it’s this inconsistently that makes it not meaningfully a game. Unless we’re talking about intentionally rule-breaking experiments, games need to have systems that govern what happens within them, which players can rely on behaving the same every time they’re played. If the world were randomly regenerated every time you jumped in Super Mario Bros., it would become unplayable, and that’s more or less what’s happening in Microsoft’s gen AI demo. The problem isn’t that the Muse model isn’t advanced enough, it’s that making a coherent game is entirely out of the scope of its capabilities.
Even if Microsoft’s demo mostly looks like Quake II, it can’t compete with the real thing.
And as with other generative AI creations, all it takes to make this awful-looking non-game is a tremendous amount of electricity. Game developer Mikołaj Kamiński, also known as Sos Sosowski, who also shared the video above, estimates that it could have taken as much as three megawatts of power to train the model running the demo, around a quarter of what the average American reportedly uses in a whole year. Whatever the exact figures are, AI models consume a vast amount of electricity just to produce applications as thoroughly unimpressive as Microsoft’s new demo.
“Copilot Gaming Experiences represents a practical step forward in exploring how AI can create brand new ways to interact with games while building upon existing ones,” Microsoft claims.
But there’s nothing novel being created by generative AI, which is limited to producing facsimiles of things that have already been fed into the model. As much as companies like Microsoft talk about gen AI pushing gaming forward, the real use case seems to be a hope that gen AI can take work from actual human game developers, and offload it to a machine that conveniently doesn’t collect a paycheck. The future of gaming clearly won’t be shaped by the kind of AI the Muse Quake II demo represents, but that won’t stop companies like Microsoft from trying to use it to replace developers, or from insisting that we don’t believe our own eyes when we’re shown a technology this dysfunctional and told it’s the next big thing.
