A year ago, running the same prompt in Midjourney yielded different results than what is seen today. AI programs can now generate photo-realistic portraits, so I wanted to explore any potential social biases in the outputs. As one might expect, the results reflect the disparities in society; most engineers are male, while most nurses are female.
These models are trained on images scraped from the Internet, and have been able to increase racial diversity in some of the outputs. However, gender scenes may be more difficult to depict accurately. In the U.S., occupations such as bartenders or optometrists are gender balanced, but the outputs of these models are mostly masculine. Even for jobs where women vastly outnumber men, such as teachers, some of the models still reflect this bias.
To move past job-specific words, I tried more general descriptors, to see if any different results were produced.