Run the same sentence through Hypertranslate twice with different language paths and you will get two completely different endings. That is not a bug — it is semantic drift, the same force that makes the playground telephone game work, amplified by the specific quirks of machine translation. In this article we unpack why meaning drifts, hop by hop, and what each type of drift teaches us about how languages differ.
Back-translation — translating a text back to its original language to check quality — is a serious tool in localization and language teaching. It is also, run to extremes, the engine of this entire site. Understanding why it degrades so entertainingly is understanding how machine translation actually works.
The three forces that break a sentence
1. Polysemy: words with too many meanings
Machine translation systems constantly guess which sense of a word is intended. English “bank” could be a financial institution or a riverside; “run” has more than sixty recorded senses. A good model resolves these from context, but every translation is a fresh roll of the dice.
In a chain, yesterday’s guess becomes today’s context. If hop one turns “bank” into the financial sense and hop two re-encodes it as the most typical sentence containing that sense, your quiet sentence about rivers becomes a sentence about mortgages. Each hop multiplies the number of low-probability but plausible readings that remain in play.
2. Idiom gaps: meaning that lives in no single word
Idioms are where chains break most spectacularly, because idioms have no word-for-word equivalent. Consider “it’s raining cats and dogs”:
- Japanese has the phrase 「土砂降り」 — literally “earth-and-sand pouring”. The cats and dogs are gone after one hop.
- Finnish offers “sataa kahta jakoa” — “raining from two buckets”. Vivid, but the animals have mutated into plumbing.
A translator must decide: preserve the image (nonsense) or preserve the meaning (blandness)? Either choice loses the original’s flavor, and the next hop inherits the loss. After three or four languages, your weather report may be about livestock, buckets, or something unrecognizable.
3. Grammar pressure: languages that force you to choose
Languages differ in what they force you to say. English lets you leave gender aside; French and Spanish do not — “a friend” must become “un ami” or “une amie”. Japanese requires guessing a politeness level; Russian demands that every noun commit to one of six cases.
When a translation system converts your text into such a language, it must invent facts your sentence never contained. Is the friend male or female? Is the listener being addressed politely or casually? Those inventions feel small, but chains make them cumulative: each hop adds invented details, and invented details get re-translated as if they were real.
Two real chains, dissected
“The early bird catches the worm”
Starting in English, routed through Japanese and Finnish before returning:
- English → Japanese: The idiom lands as a natural Japanese proverb with almost the same meaning. Politeness level gets set to plain form — the first small invention.
- Japanese → Finnish: The proverb’s subject becomes very literal: “the bird that rises early”. Finnish’s lack of a dedicated future tense quietly converts a general truth into something that sounds like it happened this morning.
- Finnish → English: The literal bird gets re-concretized, and the worm — which in Japanese was a generic “benefit” — returns as an actual worm.
Final result: something close to “the bird that got up early this morning got the worm”. Not wrong, exactly — but the proverb has become a news report about one specific bird.
“Break a leg”
Starting in English, routed through German and Swahili:
- English → German: The idiom is recognized and mapped to the German equivalent “Hals- und Beinbruch” (“neck and leg break”). Meaning preserved — but now both neck and leg are broken.
- German → Swahili: No Swahili theater idiom matches, so the system translates literally: a wish that your neck and leg break.
- Swahili → English: A literal, grim, unforgettable blessing.
Final result: “May your neck break and your leg break too.” The good-luck meaning is gone; the anatomical specifics are gloriously intact.
What drift teaches you
Semantic drift is not random noise — it is a map of the differences between languages. Watch where your sentence breaks and you learn where the languages disagree: which words are ambiguous, which idioms are culture-bound, which grammatical categories your language hides but others expose.
This is also why back-translation, used sensibly (two or three hops, not twenty), is a legitimate self-check for writers and translators: if your sentence survives a round trip with its meaning intact, it was probably unambiguous to begin with. If it comes back transformed, the chain has shown you exactly which phrase a reader might misread.
Try it yourself: run the same sentence through several different chains on the Hypertranslate home page and compare where each path breaks. The breaks are the lesson.