Core argument: Yes — world models hallucinate. A world model is an AI system with an internal picture of how the world behaves, used to predict and simulate what happens next. Its hallucinations do not look like a chatbot’s invented facts. They look like invented physics: video that stays smooth and convincing while drifting away from how reality works. Research published in June 2026 argues these failures are predictable before they occur, and rooted in one cause: gaps in what the model’s training data covered.
This is the tenth chapter of an ongoing series on world models. Previous chapters: why LLMs hallucinate, world model vs LLM, how world models work, who is building them, what artificial general intelligence really means, when it might arrive, the forces shaping AI’s decade, the Sora question, and spatial intelligence. The series hub holds the full map.
1. The short answer
Yes, world models hallucinate — and this is no longer an impression from weird clips on social media. It is the title finding of a research paper, OpenAI admits it in its own technical report, and Google DeepMind lists it in its own limitations section.
First, the vocabulary, since it matters here. A world model is a machine’s internal representation of how the world behaves: give it a situation and an action, and it predicts the outcome. The modern versions are generative — they produce their prediction as images or video, frame by frame — and one generated sequence is called a rollout: an imagined future you can watch.
Hallucination, in AI, means a model produces something wrong with complete confidence. For a chatbot, that is an invented citation. For a world model, it is stranger to watch: the rollout stays visually fluent while quietly departing from real dynamics. Nothing looks broken. That is exactly the problem — a false sentence can be caught by a skeptical reader, but false physics renders as smoothly as the truth.
The cause is shared by both kinds of model. Each is trained on examples, and when a model meets a situation its training never covered, it fills the gap with a plausible guess.
2. The flight simulator nobody programmed
A traditional flight simulator is a machine that pretends to be the sky. Engineers built it from known physics — lift, drag, engine response — and every behavior inside it can be traced to code someone deliberately wrote. If the simulator is wrong about something, the wrongness has an address.
A learned world model is a flight simulator nobody programmed. Nobody wrote its rules. It inferred them from watching enormous amounts of recorded footage. Point it at a situation, press the stick, and it renders what it believes comes next. Most of the time its belief is excellent, because most situations are ordinary and well covered by that recorded experience.
The analogy breaks right where it matters for hallucination. In a hand-built simulator, a wrong behavior is a bug in someone’s code — findable, testable, fixable. In a learned world model, the error has no address. It lives in the gap between the situations the training data covered and the situation you just requested. You cannot open the model and locate the wrong physics, because there is no physics stored in there, only a statistical memory of what usually happens next. The simulator can be audited against its spec; the learned model is its own spec.
3. Four ways an imagined world drifts from the real one
Training coverage has holes. A 2026 study by Nicklas Hansen and Xiaolong Wang at UC San Diego puts the mechanism in one sentence: hallucination concentrates in low-coverage regions of the state-action space — the combined space of every situation a model can be in and every action it can take. Where training data covered a region, the model has effectively seen the outcome. Where it did not, the model extrapolates, and an extrapolation rendered at 24 frames per second looks exactly as confident as a memory.
Plausibility is not physics. Video generators are trained to produce frames that look right to a viewer, and “looks right” is a weaker condition than “is physically correct.” OpenAI’s February 2024 report introducing Sora conceded this: the model “does not accurately model the physics of many basic interactions, like glass shattering,” and the report lists “spontaneous appearances of objects” among its common failure modes. A glass that fails to shatter is a hallucination with no visual glitch, which is why humans miss it.
Memory fades. A world you can walk around in must remember what it already generated, or it will contradict itself. Genie 3, DeepMind’s interactive world model, keeps its generated world consistent “for a few minutes” and its visual memory extends “as far back as one minute ago,” by the company’s own description. Leave a room for longer than that window, come back, and what greets you is a reconstruction, not a recollection — close, usually, but not guaranteed to be the room you left.
Errors compound. These models generate autoregressively — each new frame is produced from the model’s own previous frames, which means yesterday’s small error becomes today’s input. DeepMind states the consequence plainly in its Genie 3 post: “inaccuracies tend to accumulate over time.” A drift too small to see in any single frame becomes, a few minutes later, a different world.
4. Measured, not just observed
What separates the current moment from “AI videos look weird sometimes” is that the failure now has structure.
Hansen and Wang’s paper, “Hallucination in World Models is Predictable and Preventable,” released in June 2026, describes a 427-hour, 210-task benchmark with ground-truth physics and a 350-million-parameter world model trained on it. The authors identify three distinct hallucination modes — perceptual (the model misreads its own state), action-marginalized (rare actions it barely saw in training), and scene-diverging (the imagined world drifts from the real one) — and, for each, a lightweight signal that predicts the failure before it happens.
Their remedy matches their diagnosis. During training, they concentrate data where coverage is thin. After training, they use the hallucination predictors as curiosity rewards, steering the model toward exactly the situations it would fail in — adapting it to entirely new environments with as few as 50 real-world trajectories. Their conclusion is the strongest sentence in this literature: hallucination in world models is “inherently a data coverage issue.”
5. My take
Three judgments, in descending order of confidence.
First, high confidence: this question has moved from debate to measurement, and that is what progress looks like. The June 2026 paper treats hallucination not as a vague vibe but as a property with names, structure, and early-warning signals. I find that more convincing than any demonstration reel, in either direction.
Second, high confidence: “it’s a coverage problem” is simultaneously the good news and the expensive news. Good, because coverage can be measured, targeted, and bought. Expensive, because the physical world’s long tail is effectively unbounded — text data to cover human claims already existed on the internet, but nobody has pre-recorded every glass shattering at every angle. The data bill for world models is paid in the physical world’s own currency.
Third, medium confidence: a world model’s hallucination will matter more than a chatbot’s, because of who consumes the output. A false sentence gets audited by a reader; a false simulation gets trusted silently by everything trained inside it — including the robots that physical-AI companies plan to teach in imagined worlds. The debates about artificial general intelligence tend to argue about what these systems can do; this chapter is about the quieter question of when their confidence outruns their knowledge. My expectation: predictability of failure, not raw visual quality, becomes the metric the field is judged on.
6. Questions people actually ask
Is a world model hallucinating the same thing as an LLM hallucinating? Same family, different surface. A 2025 paper from Carnegie Mellon University and collaborators argues hallucination should be defined through the world model itself — a model hallucinates when its internal picture of the world is wrong. Under that definition, a chatbot inventing a citation and a video model inventing physics are the same error wearing different clothes. The practical difference: you can fact-check the citation. For the deeper mechanics on the text side, see why LLMs hallucinate.
Can it be fixed? Partially, and the caveat is scale. The 2026 research shows targeted data collection closes specific, detected gaps — 50 trajectories sufficed for one new environment in their setup. “Preventable” in a controlled benchmark is not “solved” for open-ended worlds, because the coverage required is unbounded.
Does this mean Sora and Genie are useless? No. Inside the regions their training covered, these models are useful — and chapter 8 takes up the separate, harder question of whether generated video counts as understanding the world. A model can be both useful and hallucinating; the failures live at the edges.
How can I spot a world model hallucinating? Watch for consistency failures rather than visual glitches: an object that is different when you look back, or physics that drifts the longer a clip runs. Those are the coverage gaps and the one-minute memory window showing through.
Is hallucination ever useful? In one form, yes. The Hansen and Wang paper uses its hallucination predictors as curiosity rewards — a flagged hallucination is a map of where to gather data next. A model that knows where it will fail knows where to look.
Sources
- Nicklas Hansen and Xiaolong Wang, “Hallucination in World Models is Predictable and Preventable”, arXiv:2606.27326, submitted June 25, 2026.
- OpenAI, “Video generation models as world simulators”, February 15, 2024 — limitations: glass shattering, spontaneous appearances of objects.
- Google DeepMind, “Genie 3: A new frontier for world models”, August 5, 2025 — few-minute consistency, one-minute visual memory, error accumulation.
- Emmy Liu, Varun Gangal, et al., “A Unified Definition of Hallucination: It’s The World Model, Stupid!”, arXiv:2512.21577, December 2025.
Next in this series: world models versus game engines — what separates the two, and why it matters. Or start from the series hub.