You can engage an LLM for role playing - as a professor, fictional character, historical figure, etc. this happens by explicit request. That works fine.

You can engage the LLM as an LLM - "ChatGPT, search the web for me on the topic of "Banjo Ukeleles: Are they more banjos, or more ukelele?". ChatGPT does a massive search and sorts out the results. It's a brilliant super-human assistant. This is the normal web chat interface. And that works fine.

What does not work: an LLM that spontaneously code-switches from assistant into role-playing a person.

If I'm discussing audio tech, talking about a gig, and the LLM replies "...having spent my time behind mixing consoles, that's deeply relatable"

Nope. Pick a lane