Images are a key feature of the communication occurring on social media and other digital platforms. Commercial image recognition services are widely used in science and industry, offering the welcome promise to learn about image content at scale, but their biases remain understudied. In this talk an evaluation of several image recognition systems will be presented using the example of Congressional digital behavior as a case study. Results uncover considerable gender biases in the data these image recognition services produced. Most notably, images of women receive three times more labels related to physical traits in comparison to images of men. It will be discussed how these encoded cultural biases limit the validity of insights we can gather from image recognition systems and how they reinforce harmful stereotypes.
Carsten is a postdoctoral research associate at CITP (2019-2020). He holds a bachelor and master degree in sociology and received his Ph.D. from the University of Bamberg in Germany. He also was previously a research associate at the University of Bamberg. His dissertation focused on the application of computational methods for the study of ethnic minorities. Carsten is also interested in social media communication, algorithmic bias, natural language and image processing, data mining and software development. He co-organized a partner site for the Summer Institute in Computational Social Science and gave courses on methods of political sociology and computational social science at University of Bamberg, University of Constance and Humboldt University of Berlin.
In an effort to support sustainability at our events, attendees are encouraged to bring reusable items for their personal use.