That's assuming the training set is all of the media that mainstream Google Translate was trained on. The paper talks about more specialized training sets so it

That's assuming the training set is all of the media that mainstream Google Translate was trained on. The paper talks about more specialized training sets so it's quite possible something like onigiri could end up under-represented (though I suspect it's common enough in manga that it would be something more obscure). Naturally the "jelly donut" part would only slip in if they used untrustworthy data like forum posts or 4kids localizations, but that's exactly the kind of data that Google Translate and similar algos scrape from the internet for training.