As AI models are increasingly being developed to simulate environments that people can interact with rather than simply generate images or video clips, Google DeepMind remains at the forefront.
The AI research laboratory has now taken this approach further with Genie 4, a new world model that creates interactive environments in real time and allows users to change them using natural language prompts.
Announced on October 27, Genie 4 can generate photorealistic environments with movement, physics, sound, and interactive characters. Users can navigate these worlds with standard keyboard controls while using text prompts to change elements such as the weather, camera position, people, or objects as the simulation continues.
Genie 4 Generates Worlds That Respond to Users
Genie 4 is designed to simulate what happens next as users interact with an environment. According to Google DeepMind, the model can handle physical events such as collisions and aerodynamics, while generated characters can move and respond more naturally to their surroundings.
The model also allows users to modify a scene without restarting the simulation. A person could change the weather, introduce a new character or object, or move the camera, with Gemini models helping Genie 4 adapt the scene to the new instruction.
This makes Genie 4 different from conventional AI video generation, where the output is generally a finished sequence. Genie 4 instead generates an interactive environment that continues responding as the user moves through it and provides new instructions.
Native Audio Adds Another Layer
Audio is one of the major additions in Genie 4, with Google DeepMind describing it as its first real-time world model capable of generating audio natively.
The model can produce ambient sounds such as wind and waves, as well as music and interactive dialogue. Characters can also respond to events and user input while conversations take place within the generated environment.
Google says the model is also built with Gemini capabilities – Gemini Omni helps anchor the simulation in real-world appearance and behaviour, while Gemini 3.8 Flash acts as a guiding model that observes the generated environment, breaks goals into steps and checks whether those steps have been completed.
Genie 4 Is Being Tested for Robotics and Creative Work
Google DeepMind is already testing Genie 4 for several applications. In robotics research, Google boasts that Gemini 4 Argon can break a task into multiple steps and pass those instructions to Genie 4 while the simulation is running. The company demonstrated this approach with a simulated humanoid robot playing checkers.
Trusted testers in filmmaking, journalism, education, gaming, and engineering are also using the model. Filmmakers can test different versions of a scene, while game developers can explore interactive prototypes before writing engine code.
However, Genie 4 remains an experimental research model. Google DeepMind says visual details and object permanence can drift during longer simulations or after multiple complex edits. As such, rendering readable text, maintaining spatial accuracy, handling complex physical interactions, and producing accurate real-time audio also remain areas of ongoing research.
Other limitations also include the fact that Genie 4 currently supports only a few minutes of continuous interaction rather than extended sessions, and it cannot accurately reproduce real-world locations with perfect accuracy.
Genie 4 is currently available as a research preview to a limited group of trusted testers. Google plans to make it available to Google AI Ultra subscribers through Project Genie and to early access partners using the Gemini API and Google AI Studio.
