A landscape appears on screen. You move forward, then ask for the weather to change or a character to be added without restarting the scene. This is the kind of interaction Google is highlighting with Genie 4. The stakes go beyond creating an attractive sequence: the aim is to make an environment evolve in response to the visitor’s actions.
Announced by Google DeepMind on 27 October 2026, the model also generates ambient sounds, music and dialogue alongside the images. Gemini steps in to interpret instructions and guide the scene as it unfolds. These are capabilities presented by the manufacturer; we have not tested them directly.
The change deserves attention. But a plausible image, an enjoyable interaction and a simulation stable enough for professional work are not three ways of saying the same thing.
From sequence to modifiable environment
In a video that has already been made, viewers watch what was decided beforehand. In the experience described for Genie 4, their movements and requests contribute to what happens next. The creation is more like an exploration than watching a finished film.
Google describes trials involving previews for creative teams, prototyping for game developers and scenario generation for research. These avenues are interesting because they relate to the beginning of the process: trying out an idea, comparing an atmosphere, examining a situation before building a finished tool.
It is important to keep this in perspective. A prototype helps you learn something about an idea. It does not automatically have the qualities of a product that can be delivered, maintained and run over a long period. The speed of an initial proposal and control over the final result must be assessed separately.
The decisive test: finding the same object again
The official model page describes several limitations: visual details and objects may drift over time, particularly after complex modifications or prolonged exploration. Interactions between multiple agents remain a research problem; real places are not reproduced with perfect accuracy.
To make this concrete, imagine a test: place a cup on a table, walk around an obstacle, then return. Is the cup still in the same place and the same shape? If you move it, does that change persist afterwards? This is a proposed evaluation, not an account of a test carried out on Genie 4.
A setting can remain pleasant to look at while failing this test. For a brief search for an atmosphere, the flaw may be tolerable. For a game based on precise rules or robot training, it could completely change the value of the experience.
A few minutes do not tell us everything about a use case
Google says the model allows a few minutes of continuous interaction, rather than long sessions. Audio accuracy also remains a work in progress. The presentation page specifies that the sequences shown are shortened.
These points invite us to consider useful duration, not just spectacular duration. How long does the environment retain the characteristics needed for the task? How many edits can be made in succession before you have to start again? What proportion of a session can ultimately be reused?
These questions can be asked without assuming a negative answer in advance. A time-limited tool may have a specific and valuable role. That role simply needs to be described as precisely as its demonstrations.
Why realism cannot replace a protocol
The word “simulation” covers different methods. To put the comparison in context, NVIDIA describes Isaac Sim as a robotics simulation environment in which users assemble scenes, configure robot and sensor models, and enable physical properties. The system is used, among other things, to test software pipelines and produce synthetic data.
This comparison is not a ranking of two products. It shows why professionals need to know what they can control. In a robotics trial, being able to change an identified parameter and measure the consequences is a different requirement from quickly obtaining a plausible scene.
Any future use of generated worlds should therefore specify the quantities being assessed, the errors that can be tolerated and the checks to be carried out in the real world. A convincing appearance on screen cannot, by itself, certify that a learned behaviour will be safe off screen.
What a useful trial should document
A serious evaluation could retain the initial instruction, the requested modifications and the complete sequence. Successful scenes would then have context: the number of attempts, any restarts required, the elements that remained stable and the errors encountered.
For a creative professional, the question might be: does this exploration help with choosing a framing or an artistic direction? For a game designer: does it make it possible to test an intended interaction? For a researcher: does it produce situations whose properties can be controlled? Each use requires its own success criterion.
This framework also prevents a demonstration from being turned into a promise of general intelligence. The ability to generate credible reactions is a result to examine in its own right; it does not settle every question about understanding the world.
Access should not be confused with general availability
As of 4 November 2026, Google’s page still presents Genie 4 as a research preview for trusted testers. It then announces plans to open it to Google AI Ultra subscribers through Project Genie and to early-access partners via Gemini API and AI Studio. This announcement does not mean that everyone can access it immediately.
Genie 4’s interest today, then, lies in offering a new way to sketch out interactive worlds. To find out how far it can go, we will need to look beyond the initial surprise: the continuity of objects, control over interactions and the possibility of reproducing a useful experience.
Illustrative video game photo: a person holds a controller. It does not show Genie 4 and does not indicate that the model is available on consoles. Credit: InclusiveGameLab / Wikimedia Commons / CC BY-SA 4.0.
