In recent months, a strange phenomenon has been sweeping the web: images produced by artificial intelligence often seem to be bathed in a yellowish tint.
This yellow filter, which has almost become a signature, is as intriguing as it is annoying. Is it a simple bug, a passing trend, or the sign of a deeper problem in the world of AI?
The Yiaho team delved into this enigma to understand what lies behind this golden wave.
Yellow AI Image: A Tint That Betrays AI?
For many observers, identifying an AI-generated image has become child’s play: simply spot the dominant yellow hue that covers landscapes, portraits, or fantastic scenes. The same goes for AI-generated texts, where it is now easy to humanize an AI text on Yiaho or to detect a text written with ChatGPT.
This yellowish phenomenon for images, particularly noticeable since last spring, coincides with the explosion of certain visual trends, such as images inspired by the dreamlike and colorful aesthetics of Japanese animation studios.
But the causes are multiple and intertwined, blending technology, data, and industrial dynamics.
The Snowball Effect of Synthetic Data
At the heart of the problem is how AI models are trained. These systems, capable of generating stunning images, rely on immense visual databases.
Historically, these databases were composed of human content: photographs, paintings, drawings. But with the growing demand for training data, companies are increasingly turning to synthetic data, meaning images generated by other AIs.
This choice, driven by economic and practical reasons, has unexpected consequences.
According to experts, the repeated use of synthetic data can create a kind of closed loop. If an AI is trained on images that already contain a yellow dominance (perhaps due to an initial bias in the data or a popular stylistic trend, like the Ghibli with ChatGPT fashion), it risks amplifying this trait in its own creations.
Over iterations, this bias intensifies, giving rise to images increasingly marked by this tint. It’s as if the AI, by copying itself, ends up exaggerating its own flaws.
AI ‘Inbreeding’: A Growing Threat to LLMs
This phenomenon has a name: the ‘inbreeding’ of AI models. When a model is trained on data produced by other AIs, it risks losing diversity and quality. This process could be compared to a dynasty where repeated marriages within the same family end up accentuating undesirable traits.
In the case of AI, this results in increasingly homogeneous, sometimes absurd, and often… very yellow results.
This ‘inbreeding’ could even lead to model collapse.
After several generations of training on synthetic data, an AI can begin to produce inconsistent, even delirious results. The images it generates lose realism, become artificial, or get stuck in repetitive patterns, like this famous yellow filter.
A Human Data Crisis? Why do companies rely so heavily on synthetic data? The answer is one word: scarcity.
Human content, though rich and varied, is no longer sufficient to feed the voracity of modern AI models. Creating synthetic data is fast, inexpensive, and bypasses copyright issues. But this shortcut comes at a price. Studies show that models trained massively on synthetic data quickly lose performance, sometimes becoming incapable of producing relevant results.
Also read: Training Data: What is AI Training Data? Example and Definition
Towards the End of AI’s Golden Age?
This yellowing of images may only be the visible tip of a larger problem. Some experts believe that AI, after years of rapid progress, could be reaching a plateau. Current models, though powerful, are showing signs of strain. Elon Musk had already warned about this phenomenon, which is called Peak Data.
The growing dependence on synthetic data, combined with the difficulty of accessing quality human data, could hinder innovation. Tech companies, aware of the financial stakes, often prefer to minimize these challenges. Yet, warning signs are accumulating.
How to Break the Vicious Cycle?
To counter this phenomenon, an obvious solution would be to favor human data, which is more diverse and rooted in reality. But this involves solving complex ethical and logistical problems, such as respecting copyrights or collecting varied content on a large scale. Another approach would be to improve algorithms so that they detect and correct biases, like this obsession with yellow, from the earliest stages of training.
In the meantime, the yellow filter remains a fascinating symptom of a technology in full transformation. Far from being anecdotal, it reminds us that AI, despite its prowess, is still perfectible. This yellowing, both comical and unsettling, could well be the signal of a turning point for artificial intelligence: a future where quality will take precedence over quantity, and where humans will regain a central place in data creation.
