Hey Nell! There is a known issue with the current model where it intermittently ignores printed text as Jack explains here:
We’ve spent the last couple of months cleaning up the data and are currently training a new version of the model which we think will no longer have this pathology. So long as all goes well it should be released this week
