Hit 10,000 AI-generated images reviewed and finally saw the pattern
I do data labeling on the side for a training set. Last Tuesday I hit 10,000 images reviewed, mostly faces and street scenes. At first I thought each one was its own puzzle, but somewhere around image 8,700 I started spotting the tells without even trying. Hands with six fingers, ear shapes that repeat, shadows that don't match the light source. It made me wonder if the people building these models ever sit and stare at the raw output like this. If you label or review AI content daily, how long until the fakes stop being impressive and just look like a checklist of errors to you?