Large language models validate misinformation, finds study