This is just how assisted machine learning works. The model needs to be updated with corrected annotated data that its gotten wrong in the past. It needs to have enough of this data until the model accounts for it properly so that it begins to prioritize the corrected label instead of the incorrect one.

You're not going to be able to go in and hand-fix the model so that it doesn't cause these errors. This is one of the biggest ongoing issues with convolutional neural networks - you cannot debug them to determine why something went wrong.