Digital twins have revolutionized the industry in recent years by providing virtual representations of physical assets, systems, or processes, and relying on real-time data for effective functioning. These twins enable real-time monitoring, analysis, and simulation of real-world entities through the extensive use of various digital models, including design models, scientific models, and data models that capture the status and behavior of the corresponding physical entities. However, as the real world evolves, these models must adapt to maintain consistency with their physical counterparts. This adaptation process can lead to semantic drift, a misalignment between the digital representation and the physical reality over time. In this paper, we propose a classification and formalization of different types of semantic drift and review how this concept is understood in the literature on model-driven digital twins. We further illustrate the scenarios associated with each type of semantic drift using an urban mobility use case, explicitly highlighting the practical implications.