Large language models (LLMs) risk creating a self-perpetuating loop of producing and referencing their own content online