A saxophone player performs on a small platform surrounded by a quiet lake. The camera can move freely around the scene, the environment responds to changes, and the instrument's sound follows the unfolding action. Rather than playing back a prerecorded clip, Google's Genie 4 attempts to generate and maintain the entire experience as users interact with it.

The demonstration captures what separates world models from ordinary AI video generators. Video generation produces a sequence of images; interactive simulation must account for changing inputs, camera movement and actions that were not necessarily specified when the scene began.

Genie 4 is Google's latest attempt to solve that problem. It builds upon the earlier Genie research while introducing native sound, more realistic human motion and a greater ability to steer simulations through text.

Changing an environment without starting over

Users can navigate Genie 4's simulated environments through conventional keyboard controls. Natural-language instructions provide another level of interaction: changing the weather, introducing new objects, adding characters or shifting the camera while the simulation continues.

Google says its model can respond to these changes contextually rather than simply generating an unrelated new scene. That requires the system to maintain a representation of previous events while updating the visual environment.

The company also claims improvements in physical behavior, including aerodynamics, collisions and human interactions. Characters are intended to move more naturally and react to their surroundings rather than behaving like isolated animated elements.

Genie 4 can reportedly maintain a reasonably consistent environment for approximately two minutes under the demonstrated conditions. That doubles the duration associated with Genie 3, though consistency becomes less reliable as interactions accumulate.

Native audio changes the scope of world simulation

Previous Genie experiences focused primarily on visual interaction. Genie 4 introduces audio generation as part of the real-time simulation process.

Ambient sounds, music and interactive dialogue can now emerge alongside the changing environment. A character may respond to a user's actions or participate in a conversation that develops according to events within the simulated scene.

This makes audio synchronization a particularly demanding technical challenge. The system needs to associate sounds with relevant actions and maintain conversational context while the world continues to change.

Google's demonstrations include realistic outdoor environments as well as highly stylized virtual spaces, such as an explorable landscape made entirely of stained glass. Genie 4 is intended to support different visual representations without abandoning real-time interaction.

The company acknowledges that its audio generation is still imperfect. Sound accuracy and synchronization remain active research problems.

Gemini gives the simulation a planning layer

An important architectural change comes from closer integration with Google's Gemini models. Gemini Omni helps anchor the generated worlds in realistic appearance and behavior, while other Gemini components provide reasoning and planning.

One demonstrated configuration uses Gemini 3.8 Flash as an agent that observes the scene, interprets objectives, breaks them into steps and checks whether those steps have succeeded.

When the simulation changes unexpectedly, that supervisory process can adapt its plan. Genie 4 remains responsible for producing the immediate visual and sensory experience, while Gemini helps maintain the sequence of events.

Google has also experimented with Gemini 4 Argon to coordinate a simulation involving Apptronik's Apollo humanoid robot playing checkers. Argon plans the task and supplies instructions to Genie as the environment evolves.

This is an early research demonstration rather than evidence that the model can already reproduce all the physical and mechanical complexities required for reliable robotics training.

Game prototyping may be a more realistic near-term opportunity

The gaming industry offers an obvious potential application. A developer could describe an environment, generate an explorable version and experiment with visual styles or interactive ideas without immediately constructing an entire level through conventional tools.

That does not make Genie 4 a replacement for established game engines. Reliable gameplay systems require precise rules, predictable state management, optimized assets, persistence and consistent behavior across many hours of interaction.

Genie 4 currently operates on much shorter timescales, with environmental continuity still subject to drift.

Filmmaking is another area under investigation. Creative teams are experimenting with real-time previsualization, including adjustments to camera placement, lighting and character behavior. Journalists and educators are also exploring interactive interpretations of historical environments.

Robotics research could eventually benefit from generated training scenarios covering rare or hazardous events. However, such applications require more than convincing imagery. The underlying simulation must reproduce physical constraints accurately enough for lessons learned in virtual environments to remain useful in the real world.

The problem of persistent worlds remains

Despite the advances, Google's own documentation identifies several unresolved weaknesses. Visual details can drift, objects may lose their consistency and complex physical interactions do not always behave predictably.

Legible text inside generated environments is another challenge. Spatial accuracy and long-term coherence also remain unreliable as users navigate through a scene or repeatedly modify its contents.

These limitations help define the current technology. A convincing two-minute interactive experience is impressive, but a persistent virtual world must preserve meaningful information over much longer periods.

Google says video and audio produced with Genie 4 will contain SynthID watermarking to help identify AI-generated material.

The model has entered a research preview for selected trusted testers. Access is planned for Google AI Ultra subscribers through Project Genie and early partners using the Gemini API and Google AI Studio. A waitlist is available, while general consumer access has not yet been announced.