I think some of your points are myopic — the same thing was said for language models, but RLHF and RLAF [0] clearly show a trajectory of the models getting better at doing this self-improvement process with less “humans are special”ness.
A tool like Copilot can more or less automatically improve via telemetry, I’d expect the same thing as image models catch-up. I’d also expect the human signals to get further and further downstream of the actual creation process (e.g. Gameplay tester reports visual bug versus artist manually edits character)