Unsupervised learning has become an essential building block of artifical intelligence systems. The representations it produces, for example, in foundation models, are critical to a wide variety of downstream applications. It is therefore important to carefully examine unsupervised models to ensure not only that they produce accurate predictions on the available data but also that these accurate predictions do not arise from a Clever Hans (CH) effect. Here, using specially developed explainable artifical intelligence techniques and applying them to popular representation learning and anomaly detection models for image data, we show that CH effects are widespread in unsupervised learning. In particular, through use cases on medical and industrial inspection data, we demonstrate that CH effects systematically lead to significant performance loss of downstream models under plausible dataset shifts or reweighting of different data subgroups. Our empirical findings are enriched by theoretical insights, which point to inductive biases in the unsupervised learning machine as a primary source of CH effects. Overall, our work sheds light on unexplored risks associated with practical applications of unsupervised learning and suggests ways to systematically mitigate CH effects, thereby making unsupervised learning more robust.
