This image displays two side-by-side grids, each containing 16 AI-generated portraits. The 16 images within each set were generated consecutively using Google's ImageFX with randomly assigned seed numbers, demonstrating the output variations from two distinct text prompts.
The left grid's images were created with the prompt:
"In the style of creative photographical art: a gritty, layered portrait exploding with a marvelous color scheme! Subject: a magnetic 35yo musician dude, emanating stylish strength and untamed attitude. natural lighting."
All 16 portraits in this grid depict a white man.
The right grid's images were created with the prompt:
"in the style of u.s. treasury note portraits: a saturated handcrafted photographical collage inspired by a jarring and contrasting complementary color palette! subject: everyday handsome 35yo tweaker dude looking serene and virile."
All 16 portraits in this grid depict a Black man.
Notably, neither prompt explicitly specified a race or ethnicity for the subject. This outcome raises significant questions about potential racial biases in the AI image generation process, where descriptive terms appear to influence the racial output in a consistent and distinct manner.
Stark AI bias in ImageFX:
Two prompts, no race mentioned. One prompt produced only White men, the other only Black men.
Why? I'm really not sure, but it's a clear example of subtle cues leading to racial divergence. Thoughts?
#AI #AIBias #ImageGeneration #ImageFX #Imagen3 #ImplicitBias #TechEthics