I've been doing that. It helps that I have a substantial human writing corpus to draw from. Results are still mixed. Models simply can't resist adding some LLMisms. Pangram still detects some of it but not others, and also misdetects some of my actual writing as AI generated.

Honestly, I'd rather the shame and hate just went away instead. It's seriously exhausting and I'm starting to feel tempted to just give up and either start using AI more heavily or stop writing altogether.