A scene from an AI-generated facsimile of the computer game Doom
Imagine playing a version of the classic computer game, Doom, that’s completely AI-generated—without any computer code or graphics! A team of researchers has achieved this impressive feat, suggesting a future where artificial intelligence could create games from scratch, just as it currently generates text and images.
The groundbreaking model behind this innovation, named GameNGen, was developed by Dani Valevski and his team at Google Research. Although they chose not to speak to New Scientist, their research paper offers fascinating details. Remarkably, GameNGen can be played for up to 20 seconds while maintaining all of Doom’s original features—like scores, ammunition levels, and map layouts. Players can still attack enemies, open doors, and interact with the environment, just like in the classic game. However, after this brief period, the model starts to run out of memory, and the simulation begins to break down.
Since its original release in 1993, Doom has inspired numerous computer science projects, such as running the game on unconventional hardware like toasters, treadmills, and even espresso machines. But those attempts still relied on the game’s original code. GameNGen, however, is different: it uses a neural network, a type of AI that learned how to replicate the game by observation—without accessing any of its code.
The researchers first trained an AI model to interact with Doom in a manner similar to a human player. Then, another AI model, based on the Stable Diffusion image generator, was introduced to study how millions of game inputs affected the game’s state. This second AI effectively became a virtual replica of the game, with all the rules, instructions, and elements encoded in its network of artificial neurons. In testing, human players found it challenging to distinguish between short clips of the AI simulation and the original game.
GameNGen’s creators describe this as a proof-of-concept, showing that games could be designed by neural networks rather than traditional coding. They envision a future where games could be created from simple text descriptions or concept art, making game production more cost-effective and less dependent on human programmers.
Andrew Rogoyski, from the University of Surrey in the UK, views this as a significant step forward in AI, likening it to getting a neural network to “hallucinate” a game environment and the player’s interactions within it. However, he doesn’t believe AI will replace human game designers anytime soon. “Game studios possess the creativity and skills needed to craft these worlds, understand gameplay, and create engaging stories,” he explains. “There’s something inherently human about designing experiences that resonate with people—something that, for now and the foreseeable future, will mostly come from humans.”
As exciting as AI’s role in game development is, the heart and soul of gaming still lie in the imagination and creativity of human designers. This revolutionary AI model, GameNGen, marks the beginning of a new era, but the artistry behind game creation remains a uniquely human endeavor.
GameNGen represents a fascinating step forward in the use of AI for game development, showcasing the potential for creating games without traditional coding. While this technology could streamline production and reduce costs, it is unlikely to replace the creative minds behind today’s most engaging and imaginative games. Human game designers bring a unique understanding of storytelling, gameplay, and emotional connection—qualities that AI, for all its innovation, has yet to fully replicate. The future of game development may see AI playing a supporting role, but the heart of game creation will continue to beat with human creativity and passion.
Video – https://youtu.be/kvJ2jiN9SEI
