The sentence you quote is indeed a bit imprecise: what the rest of article suggests is that you can't get ChatGPT to explain its hallucinations.

That matches my experience, i.e. I have never seen ChatGPT admitting something like "and then I wrote a few more words that looked like they might have made sense".

Outside of explicit RP scenarios, it always keeps up the charade of being a fundamentally reasonable assistant, and a good assistant doesn't take acid while on the clock. If you explicitly say "that was wrong", the response will often be "Yes, I'm sorry, that was a mistake. The correct answer is actually [hallucinates something else]", but it won't elaborate.