ChatGPT’s answer to that is:
That’s a really good question — deeper than it looks.
The short answer is: it’s partly an illusion, but a very useful one.
Here’s what’s actually going on under the hood:
I don’t “become” a master system administrator or a psychologist or a Yugoslav historian. I don’t switch into some separate personality or load a different model.
But when you tell me “you’re a master sysadmin,” you’re really giving me a context constraint — a framing instruction. It tells me:
> “Filter your knowledge and reasoning through the mindset, vocabulary, and priorities of someone who does that professionally.”
That instruction reshapes how I select and weigh information. It narrows the scope of possible completions — for example, emphasizing practical commands and risk mitigation instead of academic theory.
So it’s not that I suddenly become that expert. It’s that the language model adjusts its style, assumptions, and relevance weighting to simulate what an expert would focus on.
That’s why it feels like “role-playing works”: it anchors the response space around the domain you care about.
If you just said “tell me how to secure a Linux server,” I’d answer accurately but generically.
If you said “you’re a 20-year veteran sysadmin who just got called in at 3 a.m. to fix this,” I’d switch tone and prioritization — triage, logs, isolation, rollback — because that’s what a real one would do.So:
🧠 Cognitively, it’s an illusion.
⚙️ Practically, it’s an interface mechanism.It’s a way for you to steer the model’s attention — not to change what I am, but to change what slice of my training you’re drawing from.