13th November 2024
Artificial intelligence has been widely discussed since the release of ChatGPT, a chatbot based on machine learning, where computers “learn” from examples using large datasets. Beyond chatbots, machine learning can handle various data types—such as images, structured data, text, and video—enabling applications like classification, clustering, and predictive analysis. These applications rely on extensive data from society, capturing useful relationships but also reflecting biases and stereotypes. Recently, several cases have highlighted how such biases in AI systems can lead to unintended discrimination, including in language models.
The project BIAS — Mitigating Diversity Biases of AI in the Labour Market (2022–2026) is a European research project under the Horizon Europe program, which is co-financed by the State Secretariat for Education, Research and Innovation SERI. It addresses the question of how such stereotypes in language models and AI applications affect the labour market.
Mascha Kurpicz-Briki, professor for Data Engineering at BIAS partner BFH and leader of the technical work package, has recently written an article about this topic for a magazine of the Federal Commission for Women’s Issues of Switzerland. The article describes how bias in AI and in particular in language models can happen, and what the consequences are. As many of the problems related to this are open search questions, the article outlines the current challenges of this field of research, in line with the goals of the BIAS project. The article also describes the possibilities on how the large public can engage with the BIAS project, for example by joining the National Labs of the participating countries.
The issue of stereotypes in language models and other AI applications is a major technical challenge that has not yet been conclusively solved by research. Nevertheless, AI is used in a wide range of different applications. Taking this concern into account when planning, developing, purchasing and using such tools is essential, but requires critical awareness of the risk. Depending on the type of application or use case, the potential difficulties of discrimination can be more or less relevant.
Read this full article in German: Mascha Kurpicz-Briki (2024) Gesellschaftliche Stereotype in Sprachmodellen: Herausforderungen und Stand der Forschung. In: Frauenfragen, Eidgenössische Kommission für Frauenfragen EKF.
