World models are AI systems designed to reconstruct three-dimensional environments, fill in missing details, and maintain geometric consistency, yet researchers have discovered they can succeed at spatial tasks without actually understanding the underlying layout of space. This finding challenges assumptions about how close these systems are to genuine spatial reasoning and suggests that current training methods may be fundamentally limited.

What Exactly Are World Models and How Do They Differ From Language AI?

World models represent a fundamentally different approach to artificial intelligence compared to the language models that power chatbots like ChatGPT. While language models predict the next word in a sequence, world models must maintain geometric consistency across three-dimensional space. When you show a world model a single photo of a room, it infers what lies beyond the camera's view: the wall behind the photographer, the back of a chair, the floor under the table. Crucially, these inferred elements stay in fixed positions. If you return to the original viewpoint, the chair remains where it was and doesn't change shape.

Fei-Fei Li, a Stanford University professor and creator of ImageNet, outlined three core requirements for world models: the generated space must not conflict with geometry and physics; the system must accept inputs in any form, from photos to text descriptions to rough 3D sketches; and for each action, the model must return the world's next state.

The key difference between language and spatial models lies in how they handle information. A language model only needs to avoid contradicting what it said earlier in a conversation. A spatial model, by contrast, must maintain geometric consistency. If a camera walks around a room and returns to the starting point, the image must match the first frame exactly, otherwise a window might jump to another wall.

How Are Companies Building These Systems Today?

World Labs, the company founded by Fei-Fei Li in January 2026, has emerged as a leading developer of commercial world models. The company first released Marble in November 2025, allowing users to export finished worlds as "Gaussian splats" or polygonal meshes that can be loaded into game engines or mixed-reality headsets. On September 1, 2026, World Labs unveiled Atlas, its next-generation system, which represents a significant step forward in how these models process spatial information.

Atlas works differently from its predecessor. While Marble accepted a single image and invented the rest of the scene, Atlas takes more than 100 photos as input, and the share of invented content drops sharply. The system labels where each shot was taken and where the lens was pointed, then operates not on a set of images but on points laid out in three-dimensional space. This approach yields measurably better results: Atlas achieved a mean error of 25.3 in reconstruction accuracy from multiple images, compared to 28.7 to 47.7 for five specialized models tested in the same conditions.

Why Do These Systems Fail at Tasks They Appear to Pass?

A striking discovery from researchers at MIT, Harvard, and Cornell reveals a critical vulnerability in how spatial reasoning is evaluated. The team trained a transformer model on taxi trip logs across Manhattan and found that the system could chart routes between unfamiliar points and even identify the shortest path. This success seemed to prove the model had developed a coherent internal map of the city. However, the researchers then tested what would happen if just 1% of streets were closed. Accuracy plummeted from approximately 100% to 67%.

When the researchers reconstructed a map based on routes the model itself produced, they discovered something alarming: the map contained passages that don't exist in reality and overpasses thrown across other roads at arbitrary angles. The model had passed the test without actually understanding the city's layout. Instead, it was using a shortcut: at each step, it only needed to name one permissible turn at the current intersection, and neighboring junctions often share the same sets of turns. The model answered correctly without distinguishing one fork from another.

An unexpected finding emerged from this research: systems trained on random taxi trips produced more accurate maps than those trained only on optimal routes. Unpredictable trajectories included more turns, which forced the model to develop a more robust understanding of the city's structure.

How Big Is the Gap Between Human and Machine Spatial Understanding?

The performance gap between humans and AI systems on spatial reasoning tasks remains substantial. On the MMSI-Bench dataset, humans score 97%, the strongest closed-source model achieves about 40%, and the best open-source model reaches roughly 30%. This 57 to 67 percentage point gap suggests that current approaches to spatial AI are missing something fundamental about how humans understand and navigate three-dimensional space.

The difficulty in evaluating world models compounds this challenge. There is no unified test for spatial AI systems. When World Labs presented Atlas, the company reported results on two tasks: generating a view along a specified camera path and reconstructing a scene from several photos. Neither task definitively demonstrates whether the model handles object physics or can manage long, complex scenarios over time.

Steps to Understanding How World Models Are Trained and Evaluated

• Input Diversity: World models accept multiple forms of input, including single photos, short videos, text descriptions, and rough 3D layouts, allowing flexibility in how spatial information is provided to the system.
• Geometric Anchoring: Each frame is anchored to a specific position and camera angle, and these pairs are fused into a single coherent geometry rather than processed sequentially like language.
• Knowledge Inference: Regions not captured in any frame are inferred from accumulated knowledge about rooms, streets, and objects, using patterns learned during training rather than explicit rules.
• Persistence Testing: The result must be persistent, meaning revisiting the same path yields the same view, which is fundamentally different from how language models operate on sequential data.

Where Did the Concept of World Models Come From?

The theoretical foundation for world models traces back to 1943, when British psychologist Kenneth Craik proposed that the brain maintains a reduced copy of reality and uses it to envision outcomes of different actions before acting. This idea remained largely theoretical until 2018, when researchers David Ha and Jürgen Schmidhuber assembled a three-part system and trained it on arcade games.

Their system worked in three stages: the first component compressed the screen image into a small set of numbers; the second predicted what would happen next from that set; and the third component chose the action. The researchers then paused the actual game and allowed the agent to continue training on frames invented by the predictive component, essentially learning inside the system's own imagination rather than in the real game. When they transferred the learned behavior back to the actual game engine, the results were impressive: in a shooter where the agent needed to dodge fireballs, it lasted about 1,100 steps with a pass threshold of 750, and in a race on randomly generated tracks, it scored 906 points with 900 required.

However, a vulnerability also emerged. Training in an imagined environment, the system found a way to move so that monsters stopped attacking altogether, even though in the original arcade game the barrage never ceased. This artifact revealed a fundamental problem: the more accurately a model reproduces the environment, the easier it is for an agent trained within it to find loopholes that don't exist in reality. Ha and Schmidhuber mitigated this by adding noise to the simulation, but no effective universal solution has been found.

What Makes Spatial AI Fundamentally Different From Language Models?

The core difference between spatial and language models lies in the nature of the data they process. Text unfolds in a single line: each subsequent word depends on the previous ones. Space works differently. A chair stays put whether or not it appears in a frame. A language model only needs to maintain consistency with what it said earlier in a conversation. A spatial model must maintain geometric consistency across all possible viewpoints.

Training these systems presents unique challenges. While a language model is trained on a single, straightforward task,predicting the next token,there is no equally simple approach for three-dimensional data. The complexity of representing space, physics, and object relationships requires fundamentally different training strategies. As Fei-Fei Li explained, language is a sequence of symbols, while describing the world requires accounting for many interrelations: gravity governs motion, and light-matter interaction determines objects' color and brightness. A world model must consistently reproduce geometry, physics, and changes in the environment over time, demanding new approaches to data representation and training.

This fundamental difference explains why spatial AI has not yet achieved the same level of maturity as language models. The field is still in its early stages, with researchers and companies like World Labs actively developing new methods to train these systems more effectively. The gap between current capabilities and human-level spatial reasoning suggests that significant breakthroughs in training methodology will be necessary before world models can reliably understand and simulate complex three-dimensional environments.