In February 2025, Microsoft unveiled an AI model called ‘Muse’, which can generate gaming visuals and controller actions as you continue to play the game. Microsoft stated that Muse is a World and Human Action Model (WHAM) that leverages the power of AI to generate entirely new sequences of gameplay. It uses a Transformer-based model.
After scaling the Muse model on 1 million training updates, Muse learnt basic movements, geometry, correct interaction, flying mechanics, and more. Now, finally, Microsoft has launched Copilot Gaming Experiences, as part of the Copilot Labs program, where the company demonstrates experimental AI demos and experiences.
Microsoft has taken the classic Quake II game and used the Muse AI model to generate gameplay in real time. It dynamically generates the gameplay sequences, and after every user input, a new AI-generated moment is created instantly. Best of all, it’s available freely for anyone to play the AI-generated version of Quake II in a web browser by clicking this link.
I played Microsoft’s AI version of Quake II, and it was decent as an early proof of concept. You can move around with WASD keyboard controls and perform actions as well. However, there is no mouse support, which is a bummer. The graphics quality is not great right now, but this is an early demo, and with future improvements, Generative AI may deliver great results in gaming, as we have seen with Nvidia’s multi-frame generation in DLSS 4.
Having said that, there is a short time limit (about 2 minutes) that prevents users from fully exploring the AI-generated game. During my gameplay, I couldn’t find any goals or objectives, and most interestingly, objects and enemies kept disappearing. So yes, there is no consistency in the AI version of Quake II.
Google’s AI Model for Gaming
For what it’s worth, Google DeepMind also demonstrated a similar AI technology for gaming in 2024. GameNGen is Google’s game engine, which is powered by an AI neural model (Diffusion-based) that allows real-time interaction in a complex environment. Google showed that GameNGen can simulate the classic DOOM game at 20 FPS on a single TPU.
Next, Google’s Genie 2 is a large foundation world model (Diffusion-based) that can generate diverse gaming environments and can respond to keyboard controls. It can also handle gravity, reflections, smoke, lighting — nearly all Physics concepts. Now, with the early advent of AI in gaming, we need to see what the future holds for the gaming industry.
Source: Beebom