Short-form video platforms like YouTube Shorts have transformed user engagement through rapid, personalized recommendations. Yet, biases embedded in these algorithms, particularly those related to the physical intensity of video content, remain largely unexplored. In this study, we present ViMET-R, a transformer-based regression model that estimates Metabolic Equivalent of Task (MET) scores to quantify the energy intensity of short videos. Built on VideoMAE embeddings trained with the Kinetics-700 dataset, ViMET-R produces robust and balanced predictions across activity levels. Applying the model to 84,816 YouTube Shorts, we uncover a consistent pattern: the recommendation algorithm converges toward moderate-intensity content, regardless of the user’s initial viewing behavior. This subtle but systematic activity-level bias introduces a new dimension to algorithmic drift, suggesting that platforms may be shaping user behavior in ways that go beyond traditional content personalization. Our work bridges physiology-related metrics with recommender system auditing and offers a novel framework for analyzing the behavioral dynamics of content exposure in short-form media, with implications for fairness and user autonomy.