This time I wanted to play around with word-to-image generation in order to reveal gender associations. When starting the creative ideation process, I usually predefined a few design qualities through the prior research phase. Often my target groups are not defined by gender, at least not on purpose, but I had the feeling that some of my design qualities are not as gender neutral as I often think they are. With all the late discussions about bias in language I thus started to translate words into image.
Step 1 - preparing the words
I started by collecting words where I felt some kind of underlying association. I relied on descriptions of target group and design qualities, some of my participants used in generative sessions.
Participant 1 was asked to describe the target group and design qualities for the next generation of 'lego mindstorm robotics'. The second participant was asked to do the same for a 'summer dress design'.
Step 2 - Translate text-to-image
This time I used an opensource styleGan called Hypnogram for my experiment. I ran some tests with the different styles it offers, before I found one that worked best for my purpose.
At first the images were so surreal, vague and artsy that I found it difficult to draw any conclusions for my experiment.
Step 3 Analyze
After the first overwhelming attempts I started to focus a bit more on the way I analyze the images. At first I was a bit disappointed, finding out that most of the words I strongly associated with for instance 'female' were depicted more 'male' in the AI-images. These images are obviously very ambiguous, and you might see something completely different in them. However, I noticed that clash of expectation I had, in turn made me question my word-associations and gender biases in language.
As with the previous experiment, I noticed how the process of collecting words, and analyzing the images, itself had an impact on my gender bias awareness. Unlike unconscious associations in objects and color, I this time found a surprising amount of words that strongly relate to gender.
With this awareness gained, I will definetly have to think twice in my next design process, using 'neutral' words to describe my design vision. Setting the right base for a gender-neutral design seems to be key.
I was surprised by the ambiguity of images. I am curious to next explore less stylized ways of transforming word to image. I was furthermore thinking of running the same experiment with names instead of random nouns and adjectives, to see how much this will change the AI's perception of gender.