Story
Oct 29, 2026
Google’s Genie 4 Makes AI Worlds Feel More Real—Just as the Stakes Rise
DeepMind’s new world model promises controllable, audiovisual simulations for creators and robots, but its own limits remain visible. The launch also lands amid stark warnings that the race toward more capable AI needs slowing down.
Google DeepMind unveiled Genie 4 on Tuesday as a research preview, pitching a world model that can generate and alter explorable scenes in real time. The company says it combines photorealistic visuals with native sound, interactive dialogue and Gemini-guided reasoning—an advance it frames as important to understanding and simulating the world on the path to AGI.1
Sundar Pichai called it Google’s “most realistic and interactive world model,” highlighting the ability to steer events, camera angles and actions while a scene is running, with audio generated to match.
2 DeepMind chief Demis Hassabis similarly described a system bringing together “real-time physics, reasoning, and audio,” and hinted at the larger implications of simulations nested within simulations.
3
The practical case is already clear in the company’s rollout. Google says filmmakers and other creatives are testing Genie 4 as a rapid previsualization tool, while robotics researchers can use it to rehearse tasks in imagined environments before attempting them in the physical world. The first users are trusted testers; Google plans to expand access through Project Genie for AI Ultra subscribers and selected Gemini API and AI Studio partners.1
Yet the launch is not a declaration of solved simulation. Google acknowledges that visual details and object permanence can drift during longer sessions, particularly after complex edits; spatial precision, readable text and accurate real-time audio remain works in progress.1 Outside reporting puts the consistency window at roughly two minutes—an improvement on Genie 3, but hardly an infinite world.4
That caveat matters beyond product polish. Nate Soares, president of the Machine Intelligence Research Institute, argues that increasingly capable general-purpose AI presents an existential rather than merely technical risk. He says governments should halt training of such systems and concentrate on narrower applications, summing up his position bluntly: “The solution … is to not build the thing that would kill everybody.”5
For DeepMind, Genie 4 is a guarded research milestone with safety evaluations and SynthID watermarking. For its critics, the very ambition behind it—more coherent, controllable AI worlds—sharpens the question of how fast the frontier should move.