Interesting! I also noticed that search engines give proper results because those are trained differently and using user search and clicks. I think these popular models could give proper answer but their safety tolerance is too tight that if the AI considers the input even slightly harmful it refuses to answer.