The best part is, the MTPE workers’ output is 100% going to get fed back into the algorithms, so it’s only a matter of time before the average error rate of the models is good enough that there’s no real reason to pay anyone to look at it.

Not quite, yes and no. It will be ever so slightly off, and the slightly off gets fed back, and language will change. It is already happening since CAT tools introduced segmentation - which means texts were segmented, and thus translated, at sentence level. Where a human translator in front of an unaltered text would have joined some sentences and separated others in the translated text, one tends to stick to the segmentation when working in a CAT tool, and MT- and CAT-translated German sounds almost, but not quite, like human German - remember that beverage machine making tea in the Hitchhiker's Guide? Only we got used to drinking the stuff and reading the texts, what can we do.

And now you have the added feature of MT flavoured translated texts. I for my part really enjoy handmade things these days, fuck AI. I thought AI generated images were funny at first, then understood the computation needs involved, the intellectual property stolen and put to work for corporations, the implications for many digital workers ... it's just more awful than fun honestly and should just fuck off back into the arseholes of the thieving scammers who invented it.