Google DeepMind is widening access to Project Genie, an experimental AI system that turns text prompts or images into interactive game-like worlds. Starting Thursday, Google AI Ultra subscribers in the U.S. can test the research prototype and send feedback as DeepMind pushes further into world models.
The release is not being framed as a finished consumer product. DeepMind researchers told TechCrunch that Project Genie can be impressive, inconsistent, and occasionally confusing, which is exactly why broader testing matters.
What Project Genie Is Built To Do
Project Genie begins with what DeepMind calls a “world sketch.” A user writes prompts for an environment and a main character, then the system creates an image that becomes the starting point for an explorable world.
The tool is powered by a combination of Google’s latest world model Genie 3, the image-generation model Nano Banana Pro, and Gemini. Nano Banana Pro generates the initial image, and Genie then uses that image to create an interactive space.
Users can move through the generated world in either first- or third-person view. They can also remix existing worlds, build on prompts from other creations, browse curated examples in a gallery, or use a randomizer tool for inspiration.
Once a world has been explored, users can download videos of the experience. For now, DeepMind is limiting world generation and navigation to 60 seconds.
Why World Models Matter
DeepMind’s launch comes five months after the Genie 3 research preview. The timing matters because world models are becoming a major focus across AI labs.
World models are AI systems that form an internal representation of an environment. That representation can help the system predict future outcomes and plan actions. Many AI leaders, including those at DeepMind, see this as an important step toward artificial general intelligence.
The nearer-term path is more practical. DeepMind sees early uses in video games and other entertainment experiences, with possible future applications in training embodied agents, also described as robots, in simulation.
Competition is also increasing. Fei-Fei Li’s World Labs released its first commercial product called Marble late last year. Runway has launched a world model recently. Yann LeCun’s startup AMI Labs will also focus on developing world models.
Where The Prototype Works Best
TechCrunch’s hands-on testing found that Project Genie performed best when asked to create more stylized or imaginative worlds. Artistic prompts using watercolors, anime style, classic cartoon aesthetics, or claymation style produced stronger results than attempts at realism.
One test generated a castle in the clouds made of marshmallows, with a chocolate sauce river and trees made of candy. The result was described as whimsical and playable, with pastel-and-white castle shapes that matched the fantasy prompt well.
The tool also showed signs of interaction. When given a photo of a desk with a stuffed toy, Project Genie animated the toy moving through the space. Other objects sometimes reacted as it passed them.
There were also moments where Genie appeared to remember parts of the space it had already generated. In one test involving a cat exploring a desk, the environment mostly remained consistent when revisited, though the model once generated a second mug.
The Limits Are Still Clear
Project Genie’s weaknesses were just as visible. Realism was a problem. Photorealistic or cinematic prompts often produced worlds that looked more like video games than real people in real settings.
Real-life photos were also unreliable as source material. When given a photo of an office and asked to recreate it exactly, Project Genie produced a world with some similar furnishings, including a wooden desk, plants, and a grey couch, but the layout changed. The result looked sterile and digital rather than lifelike.
Navigation also caused frustration. The tool uses arrows to look around, the spacebar to jump or ascend, and the W-A-S-D keys to move. In TechCrunch’s testing, the keys were sometimes non-responsive or sent the user in the wrong direction.
Interaction with the environment remains limited too. Characters sometimes walked through walls or other solid objects. DeepMind is aware of these shortcomings and is working on better realism, stronger interaction, and more user control over actions and environments.
Why Access Is Limited For Now
The 60-second cap is partly about compute. Genie 3 is an auto-regressive model, which means it requires dedicated compute while a user is exploring a generated world.
Shlomi Fruchter, a research director at DeepMind, told TechCrunch: “The reason we limit it to 60 seconds is because we wanted to bring it to more users.” He added: “Basically when you’re using it, there’s a chip somewhere that’s only yours and it’s being dedicated to your session.”
DeepMind is also using this release to gather feedback and training data. The company is presenting Project Genie as a research prototype, not a daily-use product.
Safety guardrails are already active. TechCrunch could not generate nudity, Disney-like worlds, mermaids exploring underwater fantasy lands, or ice queens in wintery castles. The article notes that in December, Disney hit Google with a cease-and-desist, accusing the firm’s AI models of copyright infringement by training on Disney’s characters and IP and generating unauthorized content, among other things.
For now, Project Genie shows both the promise and the rough edges of AI world generation. It can turn a prompt into a playable place within seconds, but it still struggles with realism, controls, consistency, and physical boundaries. That makes it a revealing early look at where world models are heading, and how far they still have to go.