AI Can Make the Screenshot. The Game Is Somewhere Behind It.

I recently started looking at visually impressive 3D games, the kind where you stop moving the character and rotate the camera for no practical reason. *Hellblade II*, *Death Stranding 2*, *Alan Wake 2*, *Black Myth: Wukong*. They do not merely look detailed. They feel like places someone has thought through for years.

The obvious question is whether AI can already make something like this.

The slightly frustrating answer is that it can, provided we are only talking about the moment captured in a screenshot.

Give an image model a sufficiently elaborate description and it can produce a ruined temple, wet volcanic terrain, a biomechanical corridor or a woman standing under unsettling cinematic lighting. A video model might even let the camera move through the scene for several seconds. At first glance, the result can look like a frame from an expensive game that took hundreds of people to produce.

But a game is not really the frame. It is everything the frame is hiding.

The door has to open. The room behind it has to exist. The object placed on the floor needs to remain there after the player walks away. The character needs to turn around without quietly changing clothes or growing another finger. The bridge destroyed earlier in the story should still be destroyed. Every attractive piece of the world needs some relationship with physics, animation, navigation, sound, performance and the thousands of other rules that have already been established.

This is where the difference between generating something plausible and building something persistent becomes difficult to ignore.

Google DeepMind describes Genie 3 as a general-purpose world model that can generate interactive environments which people can explore in real time. This is genuinely strange and impressive. The model is no longer only producing a picture of a world. It is producing the next view of that world in response to what the player does.

Yet this also reveals what kind of world it is. The environment exists partly as a prediction. It does not necessarily exist as a complete, carefully constructed place waiting behind the camera. It keeps answering the question, “What would probably appear next?” A conventional game engine is answering a more rigid question: “Given the exact state of every relevant object and system, what must happen next?”

For a brief walk through an imaginary landscape, the distinction can seem philosophical. For a twenty-hour game, it becomes the whole problem.

World Labs is approaching this from another direction. Its Marble system generates persistent 3D worlds from text, images, video or rough spatial layouts. Those environments can be explored, modified and exported as meshes or Gaussian splats for use in other tools. The important word here is not beautiful. It is persistent. Marble is trying to create a spatial object rather than an endless succession of convincing views.

This feels closer to a usable virtual set. It could dramatically reduce the time required to establish a landscape, room or atmospheric prototype. But a virtual set is still not a game. It is the place where the game might eventually happen.

Roblox’s work makes the remaining gap particularly visible. Its Cube models can generate 3D objects, but Roblox has also been working on CubePart, which allows creators to request named, separate components. Instead of receiving a car-shaped lump, a creator could define wheels, doors, lights and other pieces that need to remain individually controllable. Roblox describes the schema as a contract between the generated asset and the gameplay code.

That sounds like a technical detail until you realise it is the difference between an object that looks like a car and an object that can actually become one.

The wheels need to rotate. The doors need hinges. The headlights need to switch on. The body needs an appropriate collision shape. The player needs to know where to sit. The physics system needs to understand where the weight is. A beautiful fused mesh may be useful in the background, but it becomes strangely useless the moment anyone tries to touch it.

This is probably the central misunderstanding around AI-generated games. We judge the output by appearance because appearance is immediately available to us. Game development is judged by whether an enormous collection of invisible agreements continues to hold.

Microsoft Research’s Muse demonstrates another part of this transition. Muse can generate game visuals, controller actions or both, predicting gameplay frame by frame. Microsoft has positioned it as a tool for gameplay ideation, and its initial research was trained around the existing game *Bleeding Edge*. It is not presented as a machine that autonomously produces complete commercial games. It is closer to a new way of exploring possible behaviour inside the visual language of a game.

That word, ideation, matters. A lot of current AI technology is extremely powerful before the decisions become expensive.

It can show what a swamp city might feel like. It can propose a creature. It can create fifty variations of an abandoned railway station. It can help generate a controller, an inventory, dialogue or a rough combat encounter. It can move a small team into a convincing prototype before the team has fully worked out what it is building.

Then the prototype starts asking questions.

Can the player recognise which surfaces are climbable? Does the architecture still make sense when approached from the opposite direction? Is the enemy frightening after the tenth encounter? Does the landscape guide the player, or is every part demanding equal attention? Does the jump feel good? Does this building belong to the same civilisation as the one seen two hours earlier?

AI can respond to each question. The problem is that someone still needs to notice that the question exists.

This is where games such as *Ghost of Tsushima* are useful examples. Its landscapes are visually attractive, but their beauty is doing work. Wind, grass, particles, silhouettes and colour help orient the player. The environment does not merely contain detail. It controls attention.

An AI-generated landscape can easily become richer than this and still be worse. It might place interesting objects everywhere because interesting objects score well. It may decorate every surface, illuminate every corner and make every building a landmark. The result looks expensive but has no hierarchy. Nothing tells the player what matters.

The same problem appears at a larger scale with art direction. Producing one convincing object in a visual style is no longer especially surprising. Producing hundreds of objects that belong to the same economy, history, climate and belief system is different.

Someone has to ask why the doors are shaped this way. Who manufactured the machinery? Which materials are rare? Which colour belongs to the enemy faction? Which symbol should appear only once? Why has one neighbourhood been repaired while another has been abandoned?

Without those decisions, style becomes a surface effect. The game may resemble *Death Stranding*, but it does not have a world underneath its resemblance.

This does not mean AI is irrelevant to serious game development. The opposite is probably true. A solo creator or tiny studio can already attempt kinds of 3D work that would previously have been financially absurd. The amount of material that can be explored before committing to production is expanding rapidly.

What changes is the shape of the work.

When assets were scarce, making the asset was much of the job. When assets become abundant, someone has to reject most of them, connect the remaining ones and keep the system coherent as it grows.

The scarce skill starts to look less like production and more like judgement.

This is not as dramatic as asking an AI to make *Death Stranding 3* in Estonia. It is also more consequential. The machine may generate the flooded Tallinn harbour, the courier’s raincoat, a strange black creature and several minutes of melancholic landscape. Someone still needs to decide why the player is walking through it, what the walk should feel like, and whether the creature should have been there at all.

Source note

This essay draws on official product and research materials published by Google DeepMind about Genie 3, World Labs about Marble, Roblox about CubePart, and Microsoft Research about Muse. They represent different approaches to generating interactive environments, persistent spatial worlds, controllable 3D assets and gameplay sequences.

Connected reading

• Make Anything
• The Prototype Has Become Too Cheap to Trust
• The Product Needs a Harness
• The Interface Is Now Part of the Answer

AI authorship note

This essay grew from Kris Haamer looking at visually impressive 3D games and asking whether AI can already create something comparable. AI supported source research, drafting, editing, and implementation. Kris selected the analytical direction, chose the examples, and approved the final essay for publication.