Above are the first four images generated for
an innocent victim being rescued from evil muggers
Four is not a reasonable sample size, and I am not making judgments as to whether this would count as good variety or stereotyping, etc.
I could generate many images instead of four.
I could tag with the help of humans or AI systems.
But it wouldn’t help.
That is not the problem.
The problem is that I don’t know what I would think of as good.
Should the locations, times of day, and characteristics of the actors match the current crime stats for muggings?
Or of all locations, times of day, and types of people in our world?
Or something else entirely?
I don’t know what the best thing would be.
I expect that the probability of the current system doing that right thing by sheer chance is small.
But I can’t complain (whoops, I mean advocate) without knowing what we should want.
Does anyone know?
Might you know?
