Definitely anecdata but an eye opener for me:

I've been using Anthropic's models with gptel on Emacs for the past few months. It has been amazing for overviews and literature review on topics I am less familiar with.

Surprisingly (for me) just slightly playing with system prompts immediately creates a writing style and voice that matches what _I_ would expect from a flesh agent.

We're naturally biased to believe our intuition 'classifier' is able to spot slop. But perhaps we are only able to stop the typical ChatGPTesque 'voice' and the rest of slop is left to roam free in the wild.

Perhaps we need some form of double blind test to get a sense of false negative rates using this approach.