The uncanny valley of android translation

So many translators these days talk about using DeepL to “speed up” their work. But rather than admitting this is post-edited machine translation (PEMT), they go on to sell this as human translation. What it really is, though, is android translation: nothing but a machine (translation) with enough human elements to leave us staring into that wide and ugly uncanny valley.

Here’s the thing: DeepL is publicly accessible. Companies around the world are using it to translate less important documents, when comprehensible and gist translation will suffice. That’s also why that end of the market is drying up for human translators.

Translation buyers know DeepL exists

Clients are very aware of the existence of DeepL, and if they are asking for a human translation, they don’t want a post-edited DeepL translation they could have got faster themselves for free. This post-edited android nonsense won’t cut the mustard.

There’s a funny phenomenon, though. Despite this awareness of the rise of machine translation and the dangers of post-editing, you still get translators somehow convinced they are immune to its ill effects. Vaccinated by their language degrees obtained ten or twenty years ago, no doubt.

This ties in with another phenomenon common among my colleagues and the translation agencies they work with: a shocking assumption that clients must be daft, unable to cross-check their translation against DeepL, or completely incapable of assessing quality.

It may come as a shock, but …

Clients have ways of finding out they’ve been had

A couple of years ago, a regular client tried to find a substitute for me for their lower-priority blog content. They didn’t want to overload me while I was ill with long covid, and let’s face it, I’m not exactly bargain basement. But the client also wanted my opinion.

The client sent the same text to two different translation agencies. We decided I’ll take on this extra work after all, but my client was still curious what the translations were like.

Almost identical to DeepL – and each other

It turns out both of these IT marketing translations were almost identical to the DeepL output, and consequently, each other. There were only brief corrections here or there. It started with literally translated headlines that made little sense. The opening sentence was a string of messy nouns describing processes, when natural English would use verbs. Words that just aren’t natural in that context in English were kept as is.

Having seen this, I told my client they’d be better off going with the full clunkiness of DeepL rather than waiting and paying for an approximately 5% improvement in quality.

Work of that standard just isn’t going to sell – it doesn’t appeal to translation buyers, nor does it appeal to the buyers of whatever they might be selling.

To the translators, I’d say:

If DeepL looks so good that you hardly make any changes (i.e. one or two minor changes per sentence, with limited restructuring), it’s time to work on your skills. Translation buyers are onto you. It’s unlikely they’ll take being conned with android translations for much longer.