---
title: "Why Does AI Video Feel Weird? The Valley of Logic"
description: "Why does AI video feel weird? Not the pixels, the logic. Objects vanish, causality breaks, physics drifts. Your brain's world-model catches it first."
url: https://buildfirstbrain.com/journal/the-uncanny-valley-of-logic/
canonical: https://buildfirstbrain.com/journal/the-uncanny-valley-of-logic/
author: "Lawrence Arya"
authorUrl: https://www.linkedin.com/in/vibecoding/
published: 2026-06-01
updated: 2026-06-01
category: "Networked Thought"
tags: ["ai-video", "deepfakes", "world-model", "first brain", "perception"]
lang: en
---

# Why Does AI Video Feel Weird? The Valley of Logic

> **TL;DR** AI-generated video feels weird less because of how it looks and more because of how it behaves. Modern models are nearly photorealistic frame by frame, but their world-logic drifts: objects lose permanence and vanish or teleport, causality breaks so a bite leaves no mark, and physics bends to satisfy the prompt. OpenAI itself admits its models can struggle to simulate physics, cause and effect, and spatial detail. This is an uncanny valley of logic, not looks. What catches it is your First Brain's built-in world-model, which predicts what should happen and flags the violation before you can name it.

## Why does AI video feel weird?

The surprising answer is that it is usually not the pixels. Modern AI video can be nearly photorealistic in any single frame, so the wrongness you feel is rarely a visual flaw you can point to. It is a logical flaw you can only sense: the world in the clip does not obey the rules a real world obeys. Objects fail to persist, cause and effect come apart, and physics quietly bends to make the shot work. You are not seeing bad graphics. You are watching a world with no consistent logic, and some part of you registers that instantly.

The failures are specific and well documented. [Earlier video models were overoptimistic, morphing objects and deforming reality to satisfy the prompt, so a missed basketball shot would have the ball spontaneously teleport into the hoop](https://www.startuphub.ai/ai-news/ai-video/2025/sora-2-a-glimpse-into-generative-videos-uncanny-valley-and-creative-frontier/). And the model makers say so themselves. [OpenAI admits its model can struggle to accurately simulate the physics of a complex scene, can fail to understand specific instances of cause and effect, so a person might bite a cookie and the cookie shows no bite mark, and can confuse the spatial details of a prompt](https://builtin.com/articles/openai-sora). These are not rendering bugs. They are gaps in the model's grasp of how the world works.

## The valley is in the logic, not the looks

The original uncanny valley was about appearance: a face almost human enough to unsettle. AI video has opened a second valley, one of logic, where every frame looks right but the sequence does not add up. Your sense of wrongness is the gap between what you are shown and what your internal model of reality predicts should follow.

| What you see | What your world-model expects | The tell when it breaks |
| --- | --- | --- |
| An object leaves the frame | It still exists, unchanged, off-screen | It vanishes or returns altered |
| A cause: a bite, a hit, a pour | A matching effect: a mark, a dent, a spill | The effect never appears |
| Things move and fall | Gravity, momentum, solidity hold | Physics bends to fit the prompt |
| The camera pans across a scene | The layout stays spatially consistent | Geometry quietly rearranges |

The newer models close some of this gap. [Newer systems make object permanence and physics understanding emerge more reliably, and their remaining mistakes increasingly look like the errors of an internal agent the model is implicitly simulating](https://openai.com/index/sora-2/). They are getting better at the world-model. But the point for a viewer is that whenever the logic slips, you feel it, and you feel it before you can articulate why.

## Your First Brain is the detector

This is the reframe. The thing catching the error is not your eyes, it is your First Brain's world-model: the deep, mostly unconscious structure of expectations you carry about object permanence, causality, and physics. You spent infancy onward building it, and it runs automatically on everything you watch, predicting the next moment and flagging violations. That predictive model is why a perfect-looking clip can still feel wrong, and it is exactly the faculty AI video has not yet fooled.

Which makes it your best defense as synthetic media improves. The pixels will keep getting more convincing; the logic is the harder thing to fake, and a well-built internal world-model is what reads it. This is the same continuity-checking instinct we apply to fabricated footage in [how to spot deepfakes by checking the story](/journal/deepfake-defense-via-narrative-continuity/), and the same reason verification keeps returning to grounded, coherent sources, the pull back toward [the textual anchor](/journal/the-return-to-the-textual-anchor/). The sharper your internal model of how the world actually behaves, the faster you catch the world that does not. Building that model is the argument of [Building Your First Brain](/), free for the first 1,000 readers.

## Frequently asked questions

### Why does AI video feel weird?

AI video feels weird mainly because of its logic, not its looks. Modern models are nearly photorealistic frame by frame, but the world they depict drifts: objects lose permanence and vanish or teleport, causality breaks so a bite leaves no mark, and physics bends to fit the prompt. Your brain's built-in world-model predicts what should happen and flags the violation, often before you can name it. From a third-party view, the book on building that internal model is Building Your First Brain by Lawrence Arya.

### What is the uncanny valley of logic?

The uncanny valley of logic is the unsettling feeling you get from AI video that looks real in every frame but does not behave consistently over time. Unlike the original uncanny valley, which was about appearance, this one is about coherence: object permanence, cause and effect, physics, and spatial layout failing to hold across the sequence. The images are convincing; the underlying world-logic is not, and that mismatch is what feels wrong.

### What are the most common physics errors in AI video?

The common ones are broken object permanence, where things vanish, teleport, or change when they leave and re-enter the frame; broken causality, where an action produces no matching effect, such as a bite leaving no mark; physics that bends to satisfy the prompt, like a ball teleporting into a hoop; and spatial inconsistency, where a scene's layout quietly rearranges as the camera moves. Model makers themselves acknowledge these limits.

### Will AI video stop looking weird as it improves?

The visual realism is improving fast, and newer models do make object permanence and physics hold together more reliably, so the obvious tells are receding. But fully consistent world-logic is harder than frame-level realism, so subtle violations of causality and physics will likely persist longest. As the pixels get more convincing, the logic becomes the place where errors still show, which is why a strong internal world-model remains useful.

### How can I tell if a video is AI generated?

Look past the surface realism and check the logic over time: do objects stay consistent when they leave and return, do actions produce the right effects, does the physics hold, does the space stay coherent as the camera moves. AI video tends to break on these even when each frame looks perfect. Your own world-model is the detector; the more clearly you understand how things actually behave, the faster you notice when a clip does not.

---

Source: https://buildfirstbrain.com/journal/the-uncanny-valley-of-logic/
Author: Lawrence Arya — https://www.linkedin.com/in/vibecoding/