The paper discusses the shift from traditional data labeling practices to a perspectivist approach in machine learning. Traditionally, data labeling involves collecting and aggregating labels from multiple annotators to capture a single ground truth label. However, this approach often treats annotator disagreement as noise to be minimized. The perspectivist paradigm challenges this assumption by viewing disagreement as a valuable source of information, emphasizing the importance of considering diverse perspectives and the underlying causes of disagreement. The paper examines the assumptions and challenges of the traditional paradigm, including the belief that disagreement stems from biased or inept annotators, that disagreement is limited to subjective tasks, and that aggregated labels accurately represent the broader population's views. It highlights practical challenges such as unrepresentative annotator pools, sample error, and aggregation methods that treat minority opinions as noise, leading to miscalibrated models. The perspectivist approach, in contrast, recognizes that disagreement can arise from various factors, including demographic differences, task-specific factors, and cultural influences. It advocates for methods that account for individual annotator perspectives, such as training models with individual labels, probability distributions, and explicit modeling of annotator behavior. The paper also discusses emerging practical and normative challenges, including data quality, data ethics, institutional pressures, and personalization. Recommendations for addressing these challenges include designing data labeling processes that capture important differences in opinion, considering the context of tasks, and documenting the data labeling process to ensure transparency. For normative challenges, the paper emphasizes the need for explicit consideration of normative questions, the bounds of acceptable disagreement, and researcher positionality. It calls for a shift towards more inclusive and reflective practices in data labeling and model design, encouraging the exploration of diverse perspectives and the development of methods that account for disagreement in training and evaluation. The paper concludes that perspectivist approaches can better address the complexities of human perspectives in data labeling and model development.The paper discusses the shift from traditional data labeling practices to a perspectivist approach in machine learning. Traditionally, data labeling involves collecting and aggregating labels from multiple annotators to capture a single ground truth label. However, this approach often treats annotator disagreement as noise to be minimized. The perspectivist paradigm challenges this assumption by viewing disagreement as a valuable source of information, emphasizing the importance of considering diverse perspectives and the underlying causes of disagreement. The paper examines the assumptions and challenges of the traditional paradigm, including the belief that disagreement stems from biased or inept annotators, that disagreement is limited to subjective tasks, and that aggregated labels accurately represent the broader population's views. It highlights practical challenges such as unrepresentative annotator pools, sample error, and aggregation methods that treat minority opinions as noise, leading to miscalibrated models. The perspectivist approach, in contrast, recognizes that disagreement can arise from various factors, including demographic differences, task-specific factors, and cultural influences. It advocates for methods that account for individual annotator perspectives, such as training models with individual labels, probability distributions, and explicit modeling of annotator behavior. The paper also discusses emerging practical and normative challenges, including data quality, data ethics, institutional pressures, and personalization. Recommendations for addressing these challenges include designing data labeling processes that capture important differences in opinion, considering the context of tasks, and documenting the data labeling process to ensure transparency. For normative challenges, the paper emphasizes the need for explicit consideration of normative questions, the bounds of acceptable disagreement, and researcher positionality. It calls for a shift towards more inclusive and reflective practices in data labeling and model design, encouraging the exploration of diverse perspectives and the development of methods that account for disagreement in training and evaluation. The paper concludes that perspectivist approaches can better address the complexities of human perspectives in data labeling and model development.