The risk of collapse for generative artificial intelligence remains ever-present. The very contents they generate can “suffocate” them. The materials they contain are able to seriously “contaminate” the database used to train them.

The study led by Ilia Shumailov, of the University of Oxford and published in the journal “Nature”, raises the problem of a dangerous spiral effect, where AI teaches us new information using the content it produces itself in a vicious circle.

“It is a study that verifies in fact, in a very precise way, the old concerns, that is, that have existed since the birth of generative models”, emphasizes Viviana Patti, computer science expert at the University of Turin.

Generative AI is now more and more widespread, starting with ChatGpt, a dozen other very popular models have arrived quickly, capable of producing all kinds of content in just a few moments, from texts to photos and videos with the internet that is rapidly being populated with content created precisely by algorithms.

Until a few years ago, the Internet could be considered a “cauldron” of content of all kinds, reliable and otherwise, but all characterized by a common element: they were produced by human beings.

Now this is no longer the case and may be less and less, causing a change that can lead to a significant problem.

In fact, AIs learn based on the materials, texts or photos, that are made available to them and on which they are trained, and until now these were generally human-produced content, for example Wikipedia pages or groups animal photos, but now that the internet is being populated with AI-generated content, it could change the quality of what AI itself learns.