What is Google Genie 4 — and how is an interactive world model different from AI video?

Google DeepMind’s Genie 4, unveiled October 28, turns a text prompt into a virtual world you can walk through and change while it runs, with Gemini models reasoning about what should happen next and generating matching sound. Google says it is in research preview with trusted testers — here is what it does, how it differs from one-way AI video, and what to check before believing the demos.

Short answer: Genie 4 is Google DeepMind’s newest “world model” — you type a text prompt and it builds a small virtual world you can walk through and change while it is running, instead of a finished video clip you can only watch. Google unveiled it on October 28 in research preview, and says it is being tested with trusted partners first. Why it matters: an AI video generator renders one fixed sequence of frames, so if you turn left, there is nothing there until a new clip is made.

A world model keeps a running simulation — it tracks where things are, applies physics-like rules, and answers your controls moment by moment. Google says Genie 4 pairs that simulation with Gemini reasoning: Gemini models watch the scene, break a goal into steps, and check the frames to keep events coherent, while Genie 4 generates matching sound — wind, waves, music, and dialogue. Those are Google’s claims, from Google’s own demos.1.

Check whether you can actually try it. As announced, access starts with trusted testers, with wider Project Genie and Gemini API access described as next in line — if a random site offers you “Genie 4” today, it is not the real thing.2. Learn the controls it uses.

In Google’s demos, people move with ordinary keyboard controls and steer the world by typing — change the weather, move the camera, or add a person or object mid-scene — and the world adapts instead of restarting.3. Test consistency, not just looks. Google says a scene can stay consistent for about two minutes.

The real test of any world model is whether a place still looks the same when you leave and come back, and whether objects keep behaving the same way.4. Know the likely uses. Google points to games, training simulations, and robotics testing — where a cheap, changeable practice world beats a pretty one-way video.Pro tip: Treat launch demos as advertising until independent testers publish hands-on sessions.

The gap between a curated two-minute demo and a world that holds together for an hour is exactly what reviews will measure.Ad’s take: Video AI shows you a place; a world model lets you walk into it and push on the walls. That is a genuinely different tool — and the pushing is where the honest reviews will come from.