It actually can be if they are using fuzzy algorithms such as neural networks. If they are using trained neural networks for text recognition on the ballots, which is highly possible, their model (processing program) can be trained on incomplete training samples data and would fail in the real world application. When the real data comes in, accuracy of a model drops significantly almost instantly, and that require manual relabeling of erroneous data, feeding it back into training loop and updating all the voting machine software on-the-fly as models updates are published. It's a non-reliable technology that extremely depends on peculiarities of training data. It may fail when you are using different pen color, print template differs even slightly from the training data or even lighting was a little bit different in scanning area and in training data.

YOU NEVER EVER SHOULD USE THIS KIND OF TECHNOLOGY WHEN ABSOLUTE ACCURACY IS REQUIRED. Who the hell cares if your model accuracy is 95%. It means 5% of ballots are lost, that would be unacceptable. You would not want to have a optional surgery that fails 5% of a time? That would be too much of a risk.