It's just averaging a collage of different real life sources it found.

From my understanding it'll fail with tasks without prior boilerplate available in the web.

And ChatGPT shouldn't be capable to port a piece of foreign code to Perl.

(But this could be a task machine learning might achieve, because it's testable and trainable)