HarperCollins will use AI to translate books

[ZodWallop]

Surprisingly difficult to find articles about this. Maybe a failure of my Google-Fu.

Harper Collins through their Harlequin imprint fired their French translators and it's replacing them with AI. They likely assumed that the 'disposable' romance novels are a good test bed for AI. But if it isn't a disaster, don't be surprised to see it encroach elsewhere.

YouTuber CriminOlly did a pretty good story on it and I think his observations are spot on.

[Sirtel]

Didn't watch the video (too long, haven't the patience), but yech.

[rantanplan]

Publishers confusing translation with localisation, sigh.

[rowe]

It's very common now for translators to heavily use AI.

At work even though I have access to writers and translators, I often use AI because it's faster and better. Just sending things for checking or if I'm not happy with the results I'm able to get. In general there's fewer defects and higher quality from the AI though.

[rantanplan]

@rowe How does your use of AI consider cultural differences, if at all? Because it’s one thing if you want to translate, say a manual. But a romance novel may be subject to cultural differences and may require more than just translation.

[JSWolf]

What I think needs to be done is for AI to translate to French and back to English to see how it compares to the original English. That would be a great way to show how bad the AI translation is.

[rantanplan]

I think you’re under the same misconception that the translation should somehow be identical to the original. That’s not the point of a localisation.

[rowe]

Originally Posted by rantanplan

@rowe How does your use of AI consider cultural differences, if at all? Because it’s one thing if you want to translate, say a manual. But a romance novel may be subject to cultural differences and may require more than just translation.

What I do is technical. And it's a massive advantage that all the major LLMs include our style guide and design system in their training data. It's actually much better at following these than most of the professionals.

But when I've used it for more literary things, it's actually been pretty good at that too.

--

There's a ministerial commitment for everything we write to also be available in Welsh. The AI is much better at this than the human Welsh speaking translators (it's always checked by a 3rd party and the AI gets things right more often because it has a much better understanding of the technical context, not just the two languages)

[slm]

Originally Posted by JSWolf

What I think needs to be done is for AI to translate to French and back to English to see how it compares to the original English. That would be a great way to show how bad the AI translation is.

Mark Twain once had his story "The Celebrated Jumping Frog of Calaveras County" translated into French (by a human being of course) and then he did a retranslation back and published it as "The Jumping Frog". The result is available on the Internet Achive

[Starker]

AI translation will bring with it all kinds of problems that are likely to fly under the radar. For example, if there's a particular phrase in the first chapter and in the last chapter there's a callback to it, AI has no reason to recognise this and translate it with the same wording, as AI is generative and therefore has variance built into it.

Even a human proofreader (cause it's unlikely a company wanting to cheap out on the translation will hire an actual experienced editor, if they aren't willing to hire a translator) will not necessarily catch these kinds of mistakes, as it's their job to only check whether the translation makes sense and is linguistically correct, not whether a phrase needs to translated the same way it has been translated before.

And of course AI doesn't understand whether a particular phrase has symbolism in it or whether it is culturally significant or whether it needs to use a particular register due to the background of the character or whether it's a joke that needs some creative translation or anything about the actual context of what it is translating. And fixing these kinds of things is not the job of a minimum wage proofreader, it's the job that a translator was supposed to be doing in the first place.

[jackm8]

Originally Posted by rantanplan

@rowe How does your use of AI consider cultural differences, if at all? Because it’s one thing if you want to translate, say a manual. But a romance novel may be subject to cultural differences and may require more than just translation.

I did plenty of hobby subtitle translation from English to Slovene in the past. Pulp and regular movies required no adjustments at all. I guess that it's the same with simple novels. Machine translations will do just fine in these cases.

[Gormagon]

Originally Posted by JSWolf

That would be a great way to show how bad the AI translation is.

Obviously you never used DeepL to translate a text, which was already very good years ago.

Originally Posted by Starker

"And of course AI doesn't understand whether a particular phrase has symbolism in it or whether it is culturally significant (...)"

I doubt that most professional translators will get this if the text to be translated is from a specific subculture the translators are not a part of. Of course they can use the internet just as AI can use it to put certain phrases into context.

Let's face it translators (as we know them) will be redundant in a few years, apart from a few exceptions.

[Starker]

Of course there are things that translators miss. It's impossible for any one human to know everything. This is also why a good editor can make a huge difference in a project. But humans are capable of picking up a lot of things an AI text generator simply cannot because they have lived experience and therefore are able to make leaps where it is not directly indicated in the text. They understand whether something is meant also to work as a metaphor, what tone a particular sentence is meant to convey, how humour works (or doesn't).

One thing that's crucial to understand about LLM-based AI is that it does not have a mind of its own. It can not use the Internet to put phrases in context, because it does not understand what context is. It can only generate the most likely next token in the sequence based on the textual data and training parameters it has.

[rowe]

Next token is a bit of a reductive view of how modern thinking models actually function. While the underlying architecture is generative, we are seeing emergent context handling that is often more robust than a human's. Humans, especially when tired or working for low fees, are often the ones just thinking about the next word.

If you look at the Welsh translation example I mentioned, the AI is consistently more accurate because it can hold a massive technical context and a complex style guide in its active attention simultaneously. It doesn't get fatigued or lose the thread of a long-range callback the way a human translator might.

The idea that humans have a monopoly on context or symbolism is becoming harder to defend. These models can often identify and maintain complex structural relationships more reliably than a proofreader on a budget. We have moved past simple word for word replacement - these systems can often effectively parse the structural intent of the text, despite this not being rooted in lived experience.

[Starker]

Yes, machines can work more consistently than humans, but that's also where they are more limited. Translation is creative work more than it is a technical task when it comes to fiction. There are reasons why writers use certain words over others or phrase things in a particular way. Things like whether it's necessary to preserve ambiguity when making translation choices or how to best convey a particular emotion are creative choices. A translator is constantly making trade-offs, weighing different options against each other without there necessarily being any one correct answer.