This is in fact precisely what happens. LLM output becomes increasingly incoherent with each subsequent generation trained off of previously AI generated data.
This is in fact precisely what happens. LLM output becomes increasingly incoherent with each subsequent generation trained off of previously AI generated data.