How can they be sure that the results are actual words from the scroll and not just hallucinations of the neural network?
If the scrolls are in such bad condition that the data is almost only noise then what stops their high-tech deep learning model from just.. making it all up?