Lightning Talk: ChatGPT - Facts and Fiction - Stefan Pabst - ACCU 2023 

All these chat-programs: They do not "understand" anything as they are LLMs and as such they fundamentally just work on pattern-recognition and trying to create a sequence that the will give it a good rating - which means correctness is not even a consideration. This also leads to a really strange phenomenons: When asked to write something in old/popular programming-languages it will produce something that looks very familiar but nearly all the time very wrong - either just out right does not compile or contains some nasty bugs (but in general doing what you want), but when you ask it to do the same in a little-used language you get way better code. And why is that? The reason is very simple: There are just way fewer bad examples. With Java most code you will find is people asking why it does not work, with say Elixier you get "look at how beautiful this looks" - and the chatGPT-code looks just as nice and works.