The Recursive Loop of AI-Generated Content

Abstract Authors Available Metrics References How to Cite

Abstract (en)

 In recent years, generative artificial intelligence (AI) based on large language models (LLMs) has started to produce a lot of online content, but can AI models be trained on data produced by other AI models? In this recursive loop, algorithms consume their own outputs, blurring the line between human-generated knowledge and machine-generated text. Training a new model based on AI-generated data can cause irreversible defects—this phenomenon is called model collapse. As a consequence, the model might begin to forget the nuances of genuine human language as it is trained on its own outputs. In other words, the model loses diversity and converges to a narrow and repetitive state. Unless new human-generated data are introduced to break the cycle, each generation of AI may perform worse. This recursive loop not only threatens the performance of future AI systems but also raises alarms about the integrity of the results produced by AI.

Abstract (es)

NA