Deep neural network models have become the dominant approach to a large variety of tasks within music information retrieval (MIR). These models generally require large amounts of (annotated) training data to achieve high accuracy. Because not all applications in MIR have sufficient quantities of training data, it is becoming increasingly common to transfer models across domains. This approach allows representations derived for one task to be applied to another, and can result in high accuracy with less stringent training data requirements for the downstream task. However, the properties of pre-trained audio embeddings are not fully understood. Specifically, and unlike traditionally engineered features, the representations extracted from pre-trained deep networks may embed and propagate biases from the model's training regime. This work investigates the phenomenon of bias propagation in the context of pre-trained audio representations for the task of instrument recognition. We first demonstrate that three different pre-trained representations (VGGish, OpenL3, and YAMNet) exhibit comparable performance when constrained to a single dataset, but differ in their ability to generalize across datasets (OpenMIC and IRMAS). We then investigate dataset identity and genre distribution as potential sources of bias. Finally, we propose and evaluate post-processing countermeasures to mitigate the effects of bias, and improve generalization across datasets.