It doesn’t reason or compute or understand anything. All it does is pick which series of words or symbols is most likely, given all the examples it was trained on, the prompt it was given, and the words it has already picked. Since it was trained on every piece of text the authors could get their hands on (every book, every webpage, every reddit post, everything), it does a surprisingly good job of reproducing natural language. But since there’s no mind behind it, it cannot recognize truth from fiction, or know when it has made a mistake. It will simply mimic the form of successful writing even when the answers it produces are completely made up.
Because it does not 'understand' what 'backwards' means.