Translation, and the fluency trap

Three eras in one lifetime

Machine translation has been rebuilt twice, and you may remember all three versions.

The first was rule-based: dictionaries and grammar rules written by linguists. It produced stiff, often comic output, and it needed a team of specialists per language pair.

The second, from the early 2000s, was statistical: align millions of sentences that already exist in two languages (parliamentary records, subtitles, UN documents) and learn which phrases correspond to which. Better, and still visibly assembled from fragments.

The third arrived around 2016, when the major services switched to neural translation, which reads the whole sentence and generates the whole sentence. The jump in quality was large and sudden. This is also the architecture that a couple of years later became the basis of language models generally, so translation is where the current wave actually started.

What it is good at, precisely

Ordinary prose between two languages with an enormous quantity of parallel text (English and Spanish, English and Chinese, English and French) is now handled well enough that a competent reader will often not notice. For gist, for correspondence, for reading a foreign article, it is solved for these pairs.

That qualification is doing all the work. Quality tracks the amount of parallel text that existed for that pair, and the distribution is brutally uneven.

Where it fails, and why it fails silently

Low-resource languages. For most of the world’s roughly seven thousand languages, almost no parallel text exists. Output degrades from accurate, to approximate, to fluent nonsense, and crucially, the fluency does not degrade with the accuracy. A translation into a low-resource language can read beautifully to the system and be wrong.

This is the trap: in every other tool, breakage looks broken. Here it looks finished.

Gender and formality. Translating from a language without grammatical gender into one with it forces a guess. “The doctor arrived” going into a gendered language must pick a gender, and the pick follows the training text: historically male for doctors, female for nurses. Providers have added alternatives for short phrases; the underlying tendency persists in longer text. Honorifics and formal registers, which matter enormously across South Asia and East Asia, are frequently flattened.

Code-mixing. Sentences that switch between languages mid-way (Hinglish, Taglish, Franglais) are how a great many people actually speak, and are badly under-represented in parallel corpora. Systems often translate the wrong half or lose the register entirely.

Idiom, names and numbers. Idioms are translated literally or replaced with an unrelated idiom. Proper nouns get translated when they should not be. And digits are occasionally altered, which is the failure most likely to cause real harm in a medical or contractual document.

Checking a translation without a second language

Three techniques that need no bilingual friend:

1. Back-translate. Put the output back through into your own language, ideally using a different service. Meaning that survives a round trip is usually intact. This will not catch everything (some errors are stable in both directions) but it catches a lot for two minutes’ work.
2. Check the numbers and names by eye. Dates, amounts, dosages and proper nouns should match character for character. This is where the expensive mistakes live.
3. Use two independent systems. Where they agree you are probably fine; where they diverge, something is uncertain and worth a human.

The free path, and the Indian-language case

Google Translate and DeepL are the familiar names, and DeepL is often better for European pairs. But the free and open options here are strong and under-known.

IndicTrans2, from AI4Bharat at IIT Madras, is an open-source model covering all 22 scheduled Indian languages, released with open weights and usable free. Bhashini is the Indian government’s national language platform with public APIs. NLLB-200 from Meta covers 200 languages with open weights. Argos Translate runs entirely offline on a laptop.

For a great many people reading this, that matters more than the commercial services, because these models were built specifically for languages the commercial systems treat as an afterthought, and because you can run them without sending the document anywhere.