The thing is, llm token frequency was derived from human writings like yours, and rlhf for good writing practices, like the emdash. So getting ’detected’ on good writing it’s unfortunately to be expected. My broken ESL english is much safer for now
That tracks with reality, as the majority of people don't have English as their first language. Depending on data sources used for training that could well reflect into AI detection tools.
I would also imagine the majority of people used to RLHF on AI could have English not as a first language. I have no expertise in this area, but I feel like the tuning at that phase would also affect the "voice", "tone", and "style" that gets put out.
