Observation operators are modules within weather models that take information about a model state (e.g., estimated atmospheric temperature) and predict observational values (e.g., brightness temperatures observed by satellites). These operators are beneficial for assimilating observations into weather forecasting models; however, classic observation operators are slow and constrained in what they can predict. To address these challenges, we train a variety of machine learning models on different input and output variables. While previous work has shown that machine learning models have much faster inference times than classic models, our work also shows that these models can effectively predict a wide range of variables. In particular, we find that the models are very accurate at predicting the errors between classic observation operator predictions and actual observations. Furthermore, we recognize that most AI forecasting models face computational constraints and therefore use a sparser representation of the model state (i.e., fewer vertical levels) than the dataset provides. We analyze the effect of this sparsification on the performance of our observation operators and find that this does not negatively affect performance until the most extreme cases. We hope that this work will inspire the use of more variations of observation operators within AI weather models.