The paper challenges common assumptions in the unsupervised learning of disentangled representations, which are believed to be crucial for better representation learning. The authors theoretically show that unsupervised learning of disentangled representations is fundamentally impossible without inductive biases on both the models and the data. They then conduct a large-scale empirical study involving over 12,000 models trained on seven different datasets using six prominent methods and evaluation metrics. The results indicate that while the methods successfully enforce properties encouraged by the corresponding losses, well-disentangled models cannot be identified without supervision. Additionally, increased disentanglement does not lead to a decreased sample complexity for downstream tasks. The paper suggests that future work should explicitly address the role of inductive biases and (implicit) supervision, investigate concrete benefits of disentanglement, and consider a reproducible experimental setup covering multiple datasets.The paper challenges common assumptions in the unsupervised learning of disentangled representations, which are believed to be crucial for better representation learning. The authors theoretically show that unsupervised learning of disentangled representations is fundamentally impossible without inductive biases on both the models and the data. They then conduct a large-scale empirical study involving over 12,000 models trained on seven different datasets using six prominent methods and evaluation metrics. The results indicate that while the methods successfully enforce properties encouraged by the corresponding losses, well-disentangled models cannot be identified without supervision. Additionally, increased disentanglement does not lead to a decreased sample complexity for downstream tasks. The paper suggests that future work should explicitly address the role of inductive biases and (implicit) supervision, investigate concrete benefits of disentanglement, and consider a reproducible experimental setup covering multiple datasets.