A language model is trained to determine if Sentence B is the direct successor to Sentence A. The training data consists of 'positive' pairs (consecutive sentences from the same document) and 'negative' pairs (two sentences randomly selected from different documents). The model achieves 99% accuracy on this task but fails to perform well on tasks requiring nuanced language understanding. Based on the training setup, what is the most probable reason for this discrepancy? Explain how the model likely achieved its high accuracy without developing a deep understanding of language.