First paper of session 22 at #FAccT21 is on "Bias in Generative Art" with Ramya Srinivasan. Looks at AI systems that try to generate art based on specific historical artists' styles, but using causal methods, analyzes the biases that exist in the art generation.
They note: It's not just racial bias that emerges, but also bias that stereotypes the artists' styles (e.g., reduction of their styles to use of color) which doesn't reflect their true cognitive abilities. Can hinder cultural preservation and historical understanding.
Their study looks at AI models that generate art mainly in the style of Renaissance artists, with only one non-Western artist (Ukiyo-e) included. Why, you might ask?
There are "no established state-of-the-art models that study non-Western art other than Ukiyo-e"!!
There are "no established state-of-the-art models that study non-Western art other than Ukiyo-e"!!
One of their case study findings: when using style transfer (to recreate art in the style of a specific artist), they found racial bias. For instance, when using these AI systems to generate art in different artists' styles, it would turn Black faces into white faces. #FAccT21 
• • •
Missing some Tweet in this thread? You can try to
force a refresh
