Report on the Current State of Societal Biases in Slovak AI

This report serves as a concluding report of our project Societal Biases in Slovak AI running from November 2022 to October 2023. The goal of our project was to understand how gender biases in particular affect Slovak AI systems, but also to increase the awareness of this issue in both the expert community and the general public, as this is an issue that affects virtually everyone using modern communication technologies. Our aim was to approach this issue from an interdisciplinary point of view. The team included people with technical expertise in AI and natural language processing; but also AI ethics experts, social scientists, translators, and gender experts.

The main contribution of our project is the evaluation of multiple types of AI systems, namely English-to-Slovak machine translation systems, Slovak masked language models, and Slovak speech-to-text systems. We have proposed and implemented an evaluation methodology and observed whether these systems exhibit biased behavior with respect to gender. Each experiment was aimed at a specific type of biased behavior, such as male as norm behavior, stereotypical thinking, or non-equitable performance. Worryingly, we were able to discover some sort of problematic behavior in all the models taken into consideration

Contents of the report

• Chapter 1 introduces the topic, our project and the role of this report.
• Chapter 2 provides an executive summary of this report and presents the main findings and conclusions of the project. It can be read to ascertain what the main outcomes of our project are at glance.
• Chapter 3 introduces the issue of fairness in AI and provides a brief theoretical background for our research. It introduces and discusses terms such as fairness, bias, equality, algorithmic harms and others.
• Chapter 4 is an invited chapter written by gender experts Mariana Szapuová and Janka Kottulová. The chapter discusses what gender stereotypes are, how they impact our society and language.
• Chapter 5 describes the experimental research we conducted. It documents how we evaluated the presence of gender biases in various Slovak AI systems, introduces the results, and discusses the implications of what we have found.
• Chapter 6 documents a data ethics assessment performed during the project to ensure that our research on sensitive issues is as ethical as possible, taking into consideration the needs and points of view of all the stakeholders.

The project is co-funded by the U.S. Embassy Bratislava