Panel 3: Metadata and Human Stereotypes in AI Systems

“偏见”是一个技术概念在机器学习,referring to situations where the training data used is not representative of the real world, leading to systematically skewed patterns or models. In contrast, "fairness" is a social concept that holds significant implications for users. According to Hannes Hapke et al., fairness is defined as the ability to identify when certain groups of people experience problems or differences compared to others. To illustrate this, consider a scenario where fairness becomes an issue in predicting credit extension for loans. If an AI model aims to determine who should be granted credit, fairness demands that those who don't pay back loans should have a different experience, i.e., their credit should not be extended. However, a problem arises if the AI model incorrectly denies loans to only individuals of a certain race. The digitalization of cultural heritage (CH) objects initially began for preservation purposes but has since enabled the application of AI technology to extract knowledge, enhancing user experiences and becoming a valuable resource for GLAM (Galleries, Libraries, Archives, and Museums). When generating digitalized data for AI use, it needs to be annotated meaningfully and relevantly for future ML tasks. Iconclass serves as a great example of such annotations. These annotations form the foundation on which AI models are built, making any embedded ideas and concepts in these structures apparent in the final AI model, thereby potentially reflecting the original biases present. The presentation will shed light on how metadata can unintentionally contain prejudices against the LGBT community, perpetuate gender inequality, or reinforce colonization stereotypes. We will demonstrate how these biases can permeate AI systems and underscore the importance of fairness in AI in general and for GLAM especially.