Every model has a faint personality, but since the personality gets "mass produced" any personality or writing style makes it easier to detect it as AI rather than harder. e.g. em dashes, etc.

But reducing personality doesn't help either because then the writing becomes insipid — slop.

Human writing has more variance, but it's not "temperature" (i.e. token level variance), it's per-human variance. Every writer has their own individual style. While it's certainly possible to achieve a unique writing style with LLMs through fine-tuning it's not cost effective for something like ChatGPT, so the only control is through the system prompt, which is a blunt instrument.