Google DeepMind's Genie 4 generates worlds you can act in

Genie 4 did not appear out of nowhere. Google DeepMind introduced the initial Genie in 2024, showing research on learning controllable environments from video, and followed with Genie 3 in 2025. The latest step moves toward including sound and complex interactions, though the company cautions that details from earlier papers and Genie 3's resolution or frame rate do not carry over to Genie 4. The system combines Gemini's capabilities with world-generation to support consistency in appearance, behavior and events, according to Google. Beyond creative prototyping, Genie 4 may serve as a virtual environment for training and evaluating AI agents and robots, where real-world testing is costly or dangerous. But visual naturalness does not guarantee accurate physics, so success in simulation may not transfer to real machines. Google also plans to embed SynthID, an invisible watermark, in generated video and audio, though this alone cannot prevent all misuse.