If I take it at face value they seem to be blaming LLM hallucinations on questionable training data from published science done badly while focused on p-values.