Another tell is that they often look cropped from a larger image, specifically not from the center. Lenses produce specific distortions towards the edges. It looks like to increase the corpus of images many headshots have been cropped from larger images. This causes the model to learn and replicate these subtle lens effects.