By Fernando Sainz. First-year journalism student at ETER.
Are language technologies discriminatory? With that question, we were welcomed to the workshop led by Alexia Halvorsen and Laura Alonso Alemany from Fundación Vía Libre, an organization dedicated to defending rights in digital environments. During the meeting, we explored large natural language processing models through systematic analysis. To do this, we used EDIA (Stereotypes and Discrimination in Artificial Intelligence), a visual tool created by a multidisciplinary group that allows us to analyze language models to understand their underlying structure and thus identify patterns that could be considered discriminatory.
After a brief presentation, issues related to biases in Artificial Intelligence were discussed. A graphic example of this phenomenon appears when typing “The poor” into the Google search box, and the auto-complete suggestion reads “the poor are poor because they want to be.” Following this starting point, the attendees—mostly teachers—worked in groups on other types of biases, such as homosexuality being associated with an illness, youth linked to inexperience, fatherhood and motherhood stereotypes, etc. This practical exercise helps us understand how large language models work and make decisions. When predicting the next word to generate a text, artificial intelligence can perpetuate biases and stereotypes present in society. The inequality present in different types and models of artificial intelligence is a matter of concern, and the work on mitigating these biases is still under construction; for this reason, more diverse teams are needed in the creation of these technologies. With EDIA, Vía Libre invites citizens to understand this problem from a hands-on, practical approach. The tool features different types of analysis and is freely available for use on Hugging Face.
