Peter Fernández Dulay
8th Grade, Julia Landon College Preparatory and Leadership Development School
Jacksonville, FL
Career Bias in AI Data
View Poster
Project Background
What does a scientist look like? That’s what Peter and his little sister Elisa were wondering as they put the word “scientist” into Canva’s Magic Media AI tool. Elisa was writing a story about a mad scientist and wanted a picture to go with it. But “all the images that showed up were of old, light-skinned men with frizzy gray hair. Not a single woman,” Peter says. “Elisa looked disappointed. She had imagined someone like herself.” Peter found out that AI tools often learn from limited data—data that can have stereotypes about who goes into different careers. He decided to find out if AI was reinforcing those stereotypes.
Tactics and Results
Women account for 35 percent of STEM graduates. Peter wanted to see if the AI image generators could produce images of scientists at the same rate. He tested four AI generation tools (Shutterstock, Canva, Dall-E and Midjourney) and asked each of them to create images based on five science careers: Actuary, data scientist, information security analyst, operation research analyst and computer and information research scientist. Each generated 20 groups of 100 images for each prompt. Peter than coded the images for how many depicted men or women. Overall, 1,459 images were of men, and 347 were women, and only 17.4 percent of the images were of women alone. Only one prompt—information security analyst—produced results equal or higher than the percentage of women in the field. Shutterstock was the least biased platform, while Midjourney, which is widely used and available with a limited subscription, was the most biased.
Beyond the Project
Peter is a regionally ranked fencer, and also a member of the robotics team. “Fencing is known as ‘physical chess’ because of the split-second decisions and analysis of the opponent. Fencing has taught me to manage stress. Robotics brings out my technical side,” he says. Peter would like to become a psychologist. “Being bilingual and multicultural myself, I would develop an AI assistant that can communicate with individuals who speak a foreign language and record psychological stress through patients’ language patterns,” he says.
