Much better to include the feature in the training, and exclude it from the inference. That actually cancels out the effect, rather than just incentivizing the model to reconstruct it.