AI is a slotmachine

Based on experience my colleague Markéta Brožová Hanelová said: “As for interpreting, I don’t think this profession will be easily erased, but similarly to translation, the developments are not driven by the rising AI-translation quality, but rather by clients willing to take risks in exchange for huge savings.”

A recent example of errors by AI is “Ich have Erfahrung mit Didaktik und Sprachlernmaterialien”. “I have” was translated by “Ich have” instead of “Ich habe”.

Some people think the AI does that to mark it as an AI translation. But those markings use special combinations of successive synonyms, not mistakes.
It’s a mystery why it came up with “Ich have”.

A translation with a CAT tool or even with MT wouldn’t do that, because what you have entered correctly in CAT or MT will be reproduced correctly. AI is trained on an enormous amount of source material, and should not come up with “Ich have”. The volume of material would average out mistakes like that. Theoretically, that is.

You could argue that the blunder is easy to notice, but there are a few problems.

First of all, the types of mistakes AI makes are clearly unpredictable. It could be anything, and therefore it demands extra attention. Not only are there seemingly easy to notice problems, but you could also be tackled by mistakes in specialised vocabulary. And that’s tricky business.

Secondly, the user cannot trust AI if one of the two languages is unknown to him. You still have to be good in the source and the target language, because otherwise you’re going to miss errors. And on top of that, you also have to know the specialised vocabulary in both languages, which is especially treacherous because different terms might superficially be the same, whereas they aren’t in reality.

Basically all problems stem from the fact that AI does not reason, but gambles when it chooses words. It’s based on the most frequent occurrences of word combinations, not on the meaning, because it doesn’t understand things.