> this shifts when they realise

I think this is pretty much is what they were saying about the provenance being blind. If you are able to deduce the LLM usage through the long-form text then that removes the blindness and objectivity.

I wonder if you view adversarially humanised LLM-generated content in the same light as that which hasn't been obfuscated. I recall seeing some paper with examples posted here a few months ago, where I (and the participants in the study) literally couldn't tell the difference between them and human writing. I have considered using humanisation myself but there is both less control over the output and it (obviously) loses the LLM-like structure that I find so easy to read - and is therefore antithesis to my own objectives.