Information shapes citizens' political decision-making. This process is amply studied by social scientists, for whom human annotation is a crucial instrument in their toolkit. Due to the democratization of data and the advances in NLP more data can be analyzed or classified, making these benchmarks more important than ever: An algorithm trained on biased data will reproduce and often exacerbate bias. Currently, disagreement and bias are often conflated – ignoring the possibility of annotator sample bias and valid disagreement. In a preregistered experiment in the Netherlands and a close replication in the U.S. we show that personal characteristics of annotators, like political ideology or knowledge, can interfere with annotators' judgment of political stances. Our results show that to improve annotated data for automated text analyses, and for stance detection models in particular, we need to critically evaluate how we create our gold standards.
Velden et al. (2025) studied this question.