They realized AI did fine at providing on-point images if asked to show either Black African doctors or white suffering children. It was the combination of those two requests that was problematic.

A postdoctoral researcher asked AI to generate images depicting Black doctors from Africa treating white children, but it kept showing the opposite.