If it's a topic that has been heavily discussed on the internet or in literature, LLMs can have good conversations about it. Take it all with a grain of salt because it will regurgitate common bad arguments as well as good ones, but if you challenge it, you can get it to argue against its own previous statements.

It doesn't handle things that are in flux very well. Or things that require very specific consistency. It's a probabilistic model where it looks at existing tokens and predicts what the next one is most likely to be, so questions about specific versions of something might result in a response specific to that version or it might end up weighing other tokens more than the version or maybe even start treating it all like pseudocode, where descriptive language plays a bigger role than what specifically exists.