Project Genie can help you test a scene before you build its rules
An explorable generated scene can support a creative decision. The current prototype and Genie 4 research preview still require different expectations from a shipped game.

An explorable scene lets a designer ask a question a still image cannot answer. Does the approach to a doorway make sense from ground level? Is a landmark visible after a turn? Moving through a generated environment can help a team decide what to explore in a conventional prototype.
Project Genie offers that kind of experiment. Google's January 2026 launch description introduced a web prototype powered by Genie 3 alongside other Google models. It generates the view in response to movement, so the interaction differs from watching a fixed clip. That response can be useful even when the resulting world is unsuitable as a finished product.
The model news has moved since that launch. On October 27, Google DeepMind announced Genie 4 with native audio and more control over the simulation. The announcement described an initial research preview for trusted testers, ahead of wider access. It should not be read as proof that every Project Genie subscriber already has that model.
Keep the model announcement beside the product instructions
Google's current Project Genie help page still describes a 60-second exploration session and identifies Genie 3 in its known-issues guidance. It says eligible Google AI Ultra subscribers aged 18 or older can use the prototype where Ultra is available. This is broader than the US-only launch wording, with access still subject to the account's eligibility.
The distinction between a research preview and an available product affects how a team should plan an experiment. A capability announced for selected testers cannot be included in a production schedule unless the team actually has access to it. The experiment should record the experience available to the account used.
The help page also says regenerating a world from the same prompts can produce a different result. Reusing the text therefore does not guarantee that another reviewer will walk through an identical environment. If a team needs to discuss a particular moment, it should preserve a recording of that session.
That recording supplies evidence about what appeared during the run. It does not turn the generated environment into an editable scene file. Before assigning downstream work, check what the product actually exports and whether that output fits the next stage. A video can support a design discussion while leaving level construction to another tool.
Give a short exploration one question
A useful first test could concern a landmark in a fictional courtyard. Specify the intended relationship: a visitor should see the entrance after moving around a central structure. Explore the scene and record where the entrance appears. If it disappears or changes unexpectedly, the observation limits what can be concluded from the run.
This test has a narrower purpose than asking whether the generated world is good. The team can learn that a certain view or approach is worth reproducing. It can also discover that the generated interaction is too inconsistent to answer the question reliably. Both results help decide what work to do next.
Control responsiveness belongs in the notes. The current help guidance acknowledges latency between inputs and actions. A missed turn may arise from that delay, so an observer should avoid attributing every navigation difficulty to the layout. Record the intended action beside the visible response when that distinction affects the result.
The same caution applies to physical behavior. A simulated object that appears to react plausibly has passed an appearance test in that moment. Claims about a real material or a measured physical process require a different validation method. A generated environment should not silently become evidence for real-world safety or engineering decisions.
Rebuild the decision in a controllable prototype
Suppose the courtyard exploration suggests a promising entrance position. A developer can reproduce that relationship in a game engine with explicit geometry. The next test can then hold the entrance position constant while changing the player's starting point. That controlled setup can establish whether the route works repeatedly.
The generated session has helped choose a candidate. The engine prototype supplies the repeatable conditions needed to evaluate it. Keeping those roles separate also helps a team avoid overcommitting to incidental details that appeared during generation but have no defined behavior.
A similar distinction applies to audio in the newer model announcement. Sound can make a scene more convincing. A shipped experience still needs predictable triggers and an appropriate response when the player repeats an action. Those requirements should be tested in the actual system that will deliver the experience.
For a first Project Genie session, write the design question before generating the world. Save the resulting clip with the account's available model context and a short note about what the run answered. Use that observation to specify the next controllable prototype, with one relationship carried over for further testing.