I agree that AI sycophancy is an issue, but as is typical in these studies, it's missing a realistic human baseline. They say "instead of reaching out to other people", but they don't appear to have done that themselves, and instead compared the AI responses to curated datasets of relatively good advice. Seeing how many people around me (and my past self included) stay in bad situations for years, I'd want to know how AI does compared to actually talking to one's friends.