'AI Habsburgs': Scientists warn AI models are degrading

Scientists from Oxford and Cambridge universities are raising alarms about a phenomenon known as "model collapse," which is leading to a degradation in the quality of artificial intelligence responses. This occurs when successive generations of AI systems are trained on synthetic content generated by their predecessors, a process experts liken to digital inbreeding.

“

Errors that are initially marginal begin to become entrenched and permeate into subsequent versions of the systems as part of their picture of reality.

"Errors that are initially marginal begin to become entrenched and permeate into subsequent versions of the systems as part of their picture of reality," explained Patryk Oleszczuk, a technology expert at SaldeoSMART. Other industry terms for this include "AI cannibalism" or "AI Habsburgs." The core issue is not that AI is becoming "dumber" in a colloquial sense, but rather that recursively trained models may gradually lose their ability to accurately reflect real-world data distributions. Rare and atypical cases, often crucial for system quality and safety, are the first to disappear.

“

Other industry terms are 'AI cannibalism' or 'AI Habsburgs.'

Jacek Treder, head of AI at Digitree Group, noted that the proliferation of low-quality internet content isn't solely due to generative algorithms but also human laziness and the pursuit of shortcuts. "The biggest sin of modern creators has become the complete delegation of the thought process to machines," he said, lamenting the trend of expecting finished, brilliant texts from simple prompts, resulting in content lacking authentic perspective.

“

The biggest sin of modern creators has become the complete delegation of the thought process to machines.

In response to these challenges, the trend is shifting away from massive, universal AI models toward specialized solutions. "In industries with zero tolerance for errors, such as finance, law, or accounting, huge general-purpose models can pose an additional source of risk," said Oleszczuk. Smaller, tailored systems trained on verified datasets and designed for specific business processes are gaining an advantage, offering greater reliability and accuracy.

“

In industries with zero tolerance for errors, such as finance, law, or accounting, huge general-purpose models can pose an additional source of risk.