Article contents:
- Introduction: The silent crisis of artificial intelligence
- “Habsburg AI”: When the model collapses
- The root of the problem: Human data shortage
- Visible symptoms: Yellowed images and repetitive answers
- Fragile defenses: Can we stop the collapse?
- A race against time
- Conclusion: Toward a more monotonous future?
Introduction: The silent crisis of artificial intelligence
For several months, a strange phenomenon has been puzzling internet users: AI-generated images take on a yellowish tint. Whether in memes, photos, or videos, this visual bias has become so widespread that numerous tutorials on how to “remove the yellow color” from these images are now circulating online. For specialists, this is about more than just aesthetics. This uniformity could be a sign of a deeper crisis threatening artificial intelligence. The reason? A form of digital “inbreeding” that occurs when AI models are trained on data already produced by other artificial intelligence systems.
“Habsburg AI”: When the model collapses
“Training future generations of software models on previous generations ultimately leads to model collapse,” says Alain Gouday, associate dean in charge of digital technology at Nema Business School. The phenomenon was dubbed “Habsburg AI” by researcher Jathan Sadowski, as it recalls the negative effects of inbreeding within the Habsburg dynasty. And the experiments speak for themselves. In a Nature study, British and Canadian researchers trained an AI model to draw handwritten digits based on a real dataset. They then repeated the process, reusing the AI-generated digits at each stage. After 20 generations, the numbers became blurry, and after 30 generations, they converged into a single, indistinguishable shape.
The root of the problem: Human data shortage
This degeneration stems from a shortage of usable human data. Major models like ChatGPT, Gemini, and Claude have already been trained on most available online content. To keep growing, companies are turning to synthetic data, which is more abundant but also cheaper and royalty-free. The problem is that this data is often of lower quality, amplifying the risk of “inbreeding.” According to Gouday, “Just 0.01% of contaminated data can lead to a sharp drop in performance, whether in images, text, or video.”
Visible symptoms: Yellowed images and repetitive answers
In text, the outcome is no more reassuring: over several iterations, a chatbot tasked with completing the sentence “To cook a Thanksgiving turkey, you…” began reciting endless lists before getting bogged down in absurd repetitions. Gouday explains: “The system tends to focus on the average, and outliers begin to disappear. This is known as early collapse, followed by late collapse, where responses become depleted and sometimes far removed from reality.”
Fragile defenses: Can we stop the collapse?
To avoid this self-degradation, the most obvious solution remains favoring diverse human content. Some companies, like OpenAI and Mistral AI, are already forming partnerships with image banks and news agencies. Others focus on detecting and tagging AI-generated content. But Gouday warns: “Currently, there is no way to guarantee that content was entirely human-generated.” Watermarking, which involves embedding recognizable signals to identify AI-generated content, remains weak and easily bypassed.
A race against time
Other methods are being explored: cleaning model training datasets by detecting and removing “contaminated” content, or assembling the highest quality datasets possible from human sources. Gouday notes: “Major publishers, such as OpenAI working with the Associated Press and Mistral AI with Agence France-Presse, strive to ensure that future generations of models are trained on original data. But it is a race against time, because the phenomenon is exponential.”
Conclusion: Toward a more monotonous future?
Without a course correction, artificial intelligence could enter a new era: one that is bizarrely more biased and monotonous. The “inbreeding” crisis is a stark reminder that the quality of artificial intelligence directly depends on the quality and diversity of the data it feeds on. Without a steady supply of original and fresh human content, artificial intelligence may find itself trapped in a self-made echo chamber, endlessly repeating its own errors and biases.
