We’ve all seen the uncanny AI images. A stunning, hyper-realistic portrait of a woman standing in front of a vanity mirror. The lighting is gorgeous, the details are razor-sharp. But then you look at the glass. The reflection is looking the wrong way. Or it’s a completely different person. Or, creepiest of all, it’s reflecting the back of her head.
Welcome to the AI “vampire effect.”
While modern AI image generators can paint textures that rival the greatest human artists, they routinely fail at basic middle-school physics. Shadows point toward the sun. Reflections defy geometry. Raindrops fall at impossible angles.
Why does a machine that can calculate millions of parameters per second fail to understand how a mirror works? The answer lies in the fascinating difference between simulating light and faking a painting.
1. Painters vs. Physics Engines (The 2D Illusion)
To understand the reflection glitch, we have to look at how different computer programs generate images.
Think about a modern video game. Games use a technology called ray tracing. The game engine builds a true 3D world, places a virtual “sun” in the sky, and mathematically calculates the exact path of millions of virtual light rays as they bounce off water, glass, and mirrors. The reflection is perfect because it physically obeys the programmed laws of optics.
AI image generators (like standard diffusion models) do not use ray tracing. They do not have a 3D space. They do not actually know what “light” is. An AI model is essentially a brilliantly talented, blind painter who has only ever studied flat, 2D photographs. It generates images by arranging a grid of colored pixels based on statistical probabilities. It knows that “shiny glass pixels” usually go next to “face pixels,” but it doesn’t understand why. It is mimicking the texture of reality without understanding the underlying structure of it.
2. The Mirror Problem: Geometry is Unforgiving
A mirror is a harsh master. It requires a perfect, mathematically rigid 1:1 reversal of 3D space. AI models, which thrive on probabilistic guessing, hate rigid rules.
When you prompt an AI with “a man looking into a mirror,” the AI processes it sequentially:
- Draw a man: It paints the back of a man’s head in the foreground.
- Draw a mirror: It paints a shiny rectangle on the wall.
- Fill the mirror: The AI’s training data says that mirrors usually contain faces. So, it statistically generates a face inside the rectangle.
Because the AI doesn’t have a 3D model of the man hidden in its memory, it doesn’t actually know what the front of his specific face looks like. It just hallucinates a plausible face based on the prompt. This results in mismatched identities, incorrect eye contact, or physically impossible viewing angles.
To the AI, the mirror isn’t a reflective surface—it’s just a separate picture frame on the wall that needs to be filled with face-flavored pixels.
3. The Multi-Sun Universe (Shadow Glitches)
This lack of a physics engine also explains why AI struggles so much with shadows.
If you look closely at an AI-generated image of a complex street scene, you can often spot the “multi-sun” error. In the real world, if you draw a line from the tip of every shadow through the object casting it, all those lines will eventually converge at a single point (the light source).
In an AI image, shadows often point in completely random directions. The AI knows that an apple should have a dark shadow underneath it, and a building should have a shadow next to it. But it processes these objects locally, not globally. It paints the apple’s shadow based on how apples usually look, and the building’s shadow based on how buildings usually look, completely unaware that they are supposedly being lit by the exact same sun.
4. Can We Fix the Physics?
Researchers are acutely aware of this limitation. Recent academic projects (like the newly developed “MirrorVerse” dataset) are trying to force AI models to learn reflections by training them specifically on thousands of mathematically perfect, synthetic 3D mirror images to brute-force an understanding of geometry.
However, this is just a very complex band-aid. As long as diffusion models operate by guessing pixel probabilities rather than simulating true 3D physics, they will occasionally slip up when the math gets too complex.
Until AI architects figure out how to merge the spatial reasoning of a 3D video game engine with the creative freedom of a diffusion model, the laws of physics will remain merely “suggestions” in the AI universe.