Is it the models or the fact that earlier hallucinations are starting to spread, accumulate and contaminate training data?

Performance degradation over time has been a recognized concern for a while now.