AI and Literary Translation: A Global Consideration

I’m just back from the 48th annual ALTA conference, in Tucson, AZ. At the panel I participated in (thank you to co-panelists Lisa Bradford and Steve Bradbury!), Amy Stolls, formerly of the NEA, suggested that a short guide to the use of AI in literary translation might be helpful to, as she put it, “people like me,” especially when approached by funders, publishers, and others wondering about its impact and potential uses.

As I think I responded at the time, one needs to be careful with such an endeavor. Listing what AI can or cannot do is bound to be obsolete very quickly. There is also the issue of potentially putting people out of work by highlighting certain kinds of works that AI currently handles fairly well.

On the other hand, as I tried to point out more than once during the gathering, generative AI has already improved a lot since we saw the first Chat GPT version in 2023, it can already do impressive things, and it has begun changing the translation landscape, in training, funding, publishing, and the very practice of translating itself. It is both already putting translators out of work and fundamentally changing the way those who still have work actually do it.

In September I gave an invited talk at Princeton University’s Program in Translation and Intercultural Communication entitled “Judgment, Generative AI, and the Practice of Translation.” There I tried to sketch out some of the changes that AI has brought and their effects, primarily on the way humans translate — I am, to be honest, far less interested by how the machines work. The talk at ALTA enabled me to push this thinking along further, fine tuning some of the earlier formulations. These are coming along, and I hope to write something more substantial on it soon.

For now, here are some thoughts of a practical nature that, in the spirit of Amy’s request, I hope will be useful to those trying to make decisions.

First, a global consideration. As with every other kind of writing, AI creates texts that tend to sound like other texts. So if you (publisher, funder, translator) are hoping for something that cuts through the noise, does original things with language, or generally takes the tops of readers’ heads off, human translators are going to be your best bet. Basically, if you’re hoping for art, you’ll need an artist.

A corollary to the global consideration. If you commit to hiring an artist, you are unlikely to improve on that person’s practice by using AI. In other words, artists have their own methods. If those methods don’t already involve AI, sticking it into their practice, or insisting that they do so (e.g., by means of project management software with embedded AI functionality, or some sort of CAT tools), is very likely to end up costing everyone more time and money, not to mention creating a result, an artistic work, that moves no one.

This consideration helps to divide up AI-translated kinds of work by genre and purpose. In terms of genre, as Tom Gally at Tokyo University has discussed in some of his online videos (try this one, for example), some kinds of popular genre fiction can surely be translated by AI with the guidance and oversight of (a) someone who can check the result against the source and (b) a skilled editor. Person (a) and person (b) might be one and the same, of course. In terms of purpose, if you’re just looking to get the gist of something, AI can do that, and it can do it very fast. Both these scenarios assume that the work in question is largely exhausted by, let’s say, plot and/or theme, that is, by the question, “What is it about?”

Please also note the necessity of both guidance and oversight in this process. By guidance, I mean devising and revising prompts. By oversight, I mean checking carefully against the source text or texts to verify that the AI hasn’t mistranslated, made things up, summarized instead of translating, or skipped things — all known problems in AI translations.

It should go without saying, but to be absolutely clear, relying on AI to give you a sense of style would be rather silly, as you generally have to tell the AI how you want it to translate to begin with. With a re-translation of an oft-translated work, e.g., Tolstoy’s Anna Karenina, this stylistic direction will be less necessary, of course (you’ll get something that sounds like other AK translations the first time you ask), but with something that has never been translated before, e.g., a brand new Korean novel, Arabic short story, or handful of poems from Indonesian, stylistically you’ll get what you guide it create.

Where the AI will tend to make more mistakes, at least at this point in its existence, is in

(a) highly context specific scenarios and

(b) places where what is needed is not a translation but something more inventive, a re-imagining or re-creation.

Under (a) cases, where context conditions language choices in radical ways, e.g., a POW camp, a French salon, etc., one can sometimes use extensive prompts, e.g., “this is a story that takes place in a rural town in southern Dalmatia in the year 1923 when X, Y, and Z are vying for power,” etc. An engineer friend of mine recently said that in order to get his AI to yield a creative result, he needs to provide two to three pages of detailed instructions in the prompt. I can imagine this for a short story more easily than a novel, but who knows.

The (b) scenario is probably best exemplified by humor. When translators take on humor, especially of a verbal variety, they very rarely translate in a direct way. Instead, they take things apart, break them down, figure them out, and then, re-imagine them for another audience. I have not been able to get AI to do this in any consistent manner that yields actual humor. Explaining a joke is just not the same as telling one.

I realize it might look like I’ve narrowed things down in the “here’s what you can do with it” category to such a tiny window that no light can penetrate. My translator colleagues will feel good about that, but I have to go on a bit, I’m afraid.

What I’ve sketched above is all in the vein of what I want to call “the old translation,” a phrase I’m using by analogy to George Steiner’s “old criticism” (from his book Tolstoy or Dostoevsky: An Essay in the Old Criticism). The old translation is also tied to the old publishing. Both have changed a lot in recent years and are continuing to do so, and generative AI will surely push them to change even more.

The role of “prompt engineering proficiency” will be key. The better people get at it, the better the products they’ll be able to create. In other words, the question is already not so much what AI can do, it is what I can get AI to do for me. And while I suspect the results will likely still look and feel like products (not art) for a spell, especially to those of us whose sensibilities were formed under the “old translation,” I cannot help but wonder how long it will take for people to lose the ability to tell the difference between what is exceptional and what is adequate.

I put this deliberately in a personal form because I’m a translator, and an editor, and a publisher, so when I see the rapidly changing circumstances all around us, my judgment tells me that trying to simply avoid them is probably dangerous. This doesn’t make me happy when I’m wearing my comfortable “old translation” cap. But if we don’t adjust course, I fear we’re likely to lose not just our caps but our heads with them.

###