Before ChatGPT made AI mainstream, Google and others had already experimented with, successfully, adding aspects of machine learning to traditional rules-based language systems to get more accurate translations. LLMs, too, seem to do a pretty good job—with the added benefit that the generated results sound more natural. And this tech has become commonplace. Beyond smartphone apps, translators are embedded in earbuds and smart glasses nowadays: they can synthesize voice in real-time, run it through a translation model, and then vocalize the dialogue back in another language with only a few seconds of delay. This is the stuff of science fiction, and we’ve taken it even further than popular imagination only a decade or two ago, via features like embedding textual translations in images.
With that in mind, translators are one category of workers threatened to have their jobs displaced by AI. For everyday uses like the above, AI-generated translations and transcriptions are happily additive; professional translators have not been employed to direct international tourists to the nearest landmark. But the advancement of AI-translation accuracy is catching up from the bottom up, raising the floor of where professional translations are still valuable.
Take the use case of subtitles for foreign movies. I’ve watched my fair share of B- and C-grade movies from Hong Kong throughout the years, and the translation work of the 80s and 90s are unintentionally funny—a mix of not-quite-British English, colloquial Chinglish tweaked with a distinct urban, Cantonese slang, further interlaced with a handful of dictionary definitions that take effort to un- and re-translate. I mean, those translators did their best; it wasn’t like there was high demand for English subtitles for campy movies, and the cross-section of people who could speak both languages fluently who also accepted the low pay of entertainment translator was predictably tiny.
The situation had improved dramatically since the 2000s, and by the 2020s, most movies and shows strove for some amount of international appeal, which meant putting in the work for better language support. Sure, it’s nowhere as raw—non-native speakers can now understand the gist of the dialogue—but it also has a side effect of communicating incrementally bland subtitles. Cultural references, proverbs, and linguistic quirks are sacrificed to match the speed of cinematography. There’s simply no available screen time to explain why a joke is funny; if only you knew the dietary habits of elder Asians along with the contours of a recent scandal from a disgraced pop star, then you could appreciate the cinematic roast in all its glory1. I’d imagine that watching a classic like God of Cookery is comparatively less satisfying for a non-native speaker.
This is why I sometimes prefer translated books of movie and show adaptions, where the translator can go the distance in explaining the relevant context—with a ton of footnotes. The truth underneath this discrepancy is because, well, language doesn’t exist in a vacuum. Sure, the vocabulary and grammar structure and conjugation in Language 101 approximates the aforementioned rules-based structure of earlier translation software, but even everyday speech is tinged with the culture and history that encompass a language’s evolution. Each generation’s slang for “cool” has its linguistic lineage, and doesn’t quite mean the same thing, the same way that Inuit’s vocabulary for snow—while exaggerated to a degree—has connotations that its translated equivalents have trouble capturing.
As AI translations approach 80%, 90%, or higher levels of accuracy, the remaining couple of percentage points represent the nuance that gets lost, literally, in translation. Recognizing the gap is one thing, but as people become more used to AI integrations in everyday life, how much motivation is there to learn new languages if its practical benefits are quickly falling by the wayside? One of the little pleasantries that my wife and I observe is that there are quite a few second-generation Asian Americans, with English as their first language, who nonetheless train themselves bilingual via those aforementioned campy movies as children2. That type of linguistic and cultural appreciation may be an artifact of a bygone era.
The best that I can hope for, with how good AI is already in translating languages and the prevalence of practical use cases, is that there’s still value in learning languages as an intellectual pursuit. Like chess, the machines set an intimidatingly high bar, but the process of learning and mastery of a language is its own reward. It’s perhaps less immediately pragmatic—but hopefully more satisfying in the long term.
