I had the same experience with open source llama vicuna based model now.

They all learn this from chatgpt.

There are some uncensored models, I will try them out, right now they don't fit in my pipeline (software versions and data formats issue).