How can you tell if a world or video is AI-generated?

Test its memory and consistency over time, because that's where generated worlds fray. Turn away from something distinctive and back — a generated world often regenerates it differently. Stay in a scene and watch it slowly drift. Retrace your steps and see if the path is the same. Push toward edges where the model has to improvise, and check physics like reflections and shadows. A real recording holds a stable world; a world model dreams each moment from the last, so permanence, spatial memory, and physics are where it gives out.

What is a world model and why does it drift?

A world model is an AI that generates an explorable place — you can seem to move through a scene it invents, often from a single image. It drifts because it largely generates each new moment from the previous one rather than maintaining a fixed, stored world. Without a persistent memory of what it already showed, details it can't currently see tend to change, so the world slowly morphs the longer you explore and forgets things you looked away from.

Will these tests keep working as AI video improves?

The specific tells will fade as memory and consistency improve — that's an arms race you'll eventually lose, the same way single-image tells for AI photos are fading. What lasts is the habit underneath: don't rely on spotting the fake, rely on provenance and source. For anything that matters, ask where a clip came from and whether its origin can be verified, rather than trusting your eye to catch a seam that may not be there next year.

Toolkit · World models

How to Spot a Fake: A Walk-In Diagnostic for AI-Generated Worlds & Video (2026)

Published July 4, 2026 · Vita Indarra

Short answer: AI world models can generate a place you seem to walk through — but they hold it together only so far, because they dream each moment from the last instead of storing a stable world. So you don't test the picture, you test the memory: look away and back, stay and watch it drift, retrace your steps, push at the edges, check the physics. Five quick tests, below, that reveal where a generated world frays — and the one habit that will still work after the tells fade.

Why generated worlds fray

A world model invents an explorable scene, often from a single image, and lets you move through it. The magic is real; so is the limit. Most of these systems generate each new moment largely from the previous one, without a persistent memory of the world they've already shown. That means the things you can't currently see are free to change — and the longer you explore, the more the world quietly rewrites itself. The tells below all probe that same weakness: no lasting memory, no fixed ground truth, just a beautiful dream being spun frame by frame.

The five-test walk-in diagnostic

CLIP / WORLD under test: ____________________

[ ] 1. LOOK-AWAY   — turn from something distinctive, then back.
       real: unchanged.   generated: often regenerates it differently (no memory).

[ ] 2. DRIFT CLOCK — stay in a scene and watch over time.
       real: stable.   generated: colours / layout / details slowly morph.

[ ] 3. ABOUT-FACE  — go somewhere, then return the way you came.
       real: same path back.   generated: often invents a new one (forgot the route).

[ ] 4. SEAMS       — push toward edges and actions it wasn't trained on.
       where it must improvise: odd geometry, melting boundaries, impossible spaces.

[ ] 5. PHYSICS     — watch reflections, shadows, how objects interact.
       generated: the gist right, the physics subtly wrong (mismatched reflection,
       shadow the wrong way, an object passing through another).

TELL PATTERN: permanence, spatial memory, and physics are where a dreamed world gives out.

The habit that outlasts the tells

Be honest about the arms race: every one of these tells will fade as memory and consistency improve, the same way the single-image giveaways for AI photos are fading. So the five tests are for today, and the habit is for always — don't rely on spotting the fake; rely on provenance. For anything that matters, ask where a clip came from and whether its origin can be verified, rather than trusting your eye to catch a seam that may not exist next year. It's the same principle behind verifiable publishing: in a world filling with generated media, checkable origin beats confident detection.

Frequently asked

Is a world model the same as an AI video generator?

Closely related but not identical — a world model lets you move through and interact with a scene, not just watch a fixed clip. The consistency tests apply to both; interactivity just gives you more ways to probe the memory.

Are world models useless because they drift?

Not at all — they're a genuine frontier with real uses. Knowing where they fray is how you use them well and aren't fooled by them, not a reason to dismiss them.

What about spotting AI photos, specifically?

Different tells, same arc — see how to tell if a photo is AI-generated. Both end at the same place: provenance over detection.

Go deeper

The field guide behind this diagnostic

These tests come from The World in the Machine — world models explained plainly: the AI that dreams up explorable places from a single photo, what it unlocks, and exactly where the dream frays. What the technology really does, why the field is betting on it, and how to see its limits with your own eyes. Live on Amazon.

The World in the Machine · $4.99 Which book should I read first?

← More field notes