If AI was able to reliably detect AI output, why not use that to correct itself to be less AI-like?

You just described adversarial training, something I've been hearing about for at least a few years now.