Forum

Genie 3 and the Designer of Worlds That Respond

Layered mountain ridges in silhouette under a soft orange and blue dusk sky

Most generative tools so far have produced objects: an image, a clip, a block of text. On 5 August, Google DeepMind showed something closer to a place. Genie 3, its latest “world model”, generates interactive environments from a text prompt and lets a user move through them in real time.

According to DeepMind, the system produces environments at 720p and 24 frames per second that stay consistent for several minutes, a big step from the short sequences its predecessor could manage. It remembers what it has already generated, so a wall you walked past is still there when you turn around. It also supports what the company calls promptable world events: you can type a change, new weather or a new object, and the world updates while you are inside it.

Genie 3 is a research preview, available only to a small group of testers, and DeepMind presents it partly as a step toward training AI agents. But designers who work with space, from architects to game designers to exhibition makers, should pay attention now.

From drawing a space to directing one

Spatial design has always relied on representation. Plans, sections, renderings and walkthrough animations are all ways of imagining a space before it exists. Each one freezes choices: the camera path, the light, the moment.

A responsive world model changes the medium. Instead of producing a fixed view, the designer sets conditions and lets the viewer explore. The craft shifts from drawing what something looks like to defining how it behaves: what persists, what changes, what the visitor can affect.

That is closer to how game designers already think. It may become how many more designers think.

What it is good for, and what it is not

It would be a mistake to treat a generated world as a design. Genie 3’s environments are plausible, not specified. They do not come with dimensions, structural logic, materials or codes. A world model can suggest an atmosphere in seconds; it cannot tell you whether a stair meets regulations or whether a gallery can carry a load.

The useful applications are earlier and softer:

  • Testing the feeling of a sequence of spaces before committing to geometry.
  • Exploring how light, weather or crowds might change an experience.
  • Letting clients or communities walk through options instead of reading drawings.
  • Prototyping interactive narratives for exhibitions, theme spaces and games.

Each of these uses depends on a designer who knows what question is being asked and how far the answer can be trusted.

New questions of authorship

World models also complicate credit. If a designer writes the prompt, adjusts events and curates the moments worth showing, how much of the resulting experience is theirs? The same debate that surrounds generated images will return in three dimensions, with the added complication that every visitor sees something slightly different.

One sensible response is to treat the prompt and the rules of the world as the authored work, and to document them carefully. A world that can be regenerated from a clear specification is easier to discuss, critique and credit than a recorded walkthrough that happened once.

Relevance for award categories

Competitions that recognize spatial and interactive work will meet these tools soon. The categories listed on the International AI Award Rules page already include AI-driven architectural design, interactive AI experiences and AI in entertainment and gaming. Entries built with world models will need to explain which parts were generated, which were specified by the designer and how the result was tested against real constraints. The 2025 winners, announced at the end of July, give a sense of how wide the field already is.

The longer view

Genie 3 is not a tool most designers can use today, and its limits are significant. But it marks a shift worth naming. Generative AI is moving from making pictures of places toward making places you can enter. When that arrives in everyday software, the most valuable skill will not be producing a convincing world. It will be deciding what a world is for, and keeping people, not prompts, at the center of that decision.