We are in an era where “believing your own eyes” has become debatable. With each new version of the AI model, the anomalies previously revealed by AI images – such as the famous six fingers or irregular rows of teeth – slowly disappear. However, the latest study by data analysis experts from the University of Hull (University of Hull) suggests that artificial intelligence still cannot cheat the laws of physics, specifically reflections in the human eye.
Problem: A perfect picture with a physical flaw
Generative AI models are masters of statistics, but not of physics. They “learn” how images should look based on billions of examples, but they don’t understand why light falls in a certain way. This leads to a phenomenon where the picture looks perfect at first glance, but is physically impossible.
The researchers applied methods commonly used in astronomy to measure the brightness of stars and galaxies to analyze it corneal reflections.
Detective work in the pupil of the eye
The main “fora” for the detection of AI images is hidden in the consistency of the reflection between the left and right eyes. In a real photograph of a person looking at a light source, the reflections in both eyes must be geometrically consistent, i.e. the position of the light source must correspond to the perspective of both eyes, but also physically coincide, The depth and intensity of the reflection must be almost identical (with minimal differences due to the angle).
By analyzing AI-generated portraits, the researchers found that the models often generated reflections that were completely different in the left and right eyes. For example, in the left eye, the reflection can be round (like from a studio light), while in the right eye it is blurred or has a completely different shape. Our brain often overlooks these details, but mathematical analysis recognizes them as impossible.
Why does AI make mistakes?
The reason is simple: AI does not render the scene in 3D space with the right light source. It just predicts pixels. Since the eyes in the images are often hundreds of pixels away from each other, the model treats them as separate entities, and not as two parts of one physical entity that must react to the same light environment.
Current restrictions and the “arms race”
Although this method is currently very accurate, the researchers caution against two things. Next, it is not infalliblebecause requires high quality images where the eyes are clearly visible. Also, AI is constantly learning, and developers are already working on implementing “physics” into the models. More advanced systems (such as those using ray-tracing AI-friendly techniques) may soon correct these errors.
Back to basics
In a world where digital forensics are trying to track AI with even more complex AI, it’s fascinating that the answer lies in classical optics. While AI deals with pixels, physics deals with truth. The next time you doubt the authenticity of a photo, don’t look at the fingers or the background – look deep into the person’s eyes. There may be proof that the person in the picture does not actually exist, reports TechSpot.