Is AI Facing Its Own Collapse—And Hurting the Internet?

Despite growing interest in artificial intelligence, with Google searches hitting 92% of their all-time peak, recent research raises concerns about AI’s future. Researchers from Cambridge and Oxford universities have found troubling signs that AI’s success might be contributing to its own downfall.

A study published in Nature last month by Dr. Ilia Shumailov and his team at the University of Oxford reveals that generative AI tools, when reliant solely on AI-generated content, experience a significant degradation in quality. The study shows that responses from AI models start to deteriorate after just a few prompts, becoming nonsensical by the ninth consecutive query. This phenomenon, termed “model collapse,” describes a cycle where AI-generated content continually degrades, ultimately resulting in distorted and unhelpful outputs.

Shumailov notes, “It is surprising how fast model collapse kicks in and how elusive it can be. Initially, it impacts minority data, leading to reduced diversity and performance degradation, which can be obscured by temporary improvements in majority data.”

This issue is particularly concerning given that approximately 57% of online text is AI-generated or AI-translated, according to a June study by Amazon Web Services researchers. If AI-generated content continues to dominate and degrade, it could mean that AI is not only undermining its own effectiveness but also deteriorating the quality of the internet as a whole.