Also not a solid tell for AI even though it struggles with uniform lighting and shadows. Humans tend to be lazy with shadows as well.

That's not an inherent thing with AI tho. Humans can do blurry hair and actually my experience with SD is hair tends to come out pretty sharp when the generation is good. Especially upscaling.

71% vs 75% has to be margin of error, 20 is probably just fine.

We were doing AI art Turing tests with SD 1.4

Most people got a little above average, but the artists tended to guess correctly the human artists and miss some of the AI. It comes down to experience really, you can see brush techniques (especially digital brushes) and pick up on things like how some aspects of AI will be inconsistent in terms of skill level across the same image or how human art will take shortcuts reusing parts of the image. The test images were carefully picked so you couldn't determine it was AI by obviously bad anatomy, text etc