What's the Hardest Language to Translate? (We Tested Them All)

Updated July 2026 — with original round-trip test data.

Short answer: by structure, Mandarin Chinese. By our test data, Japanese. And the language everyone expects to win — Finnish, with its fifteen grammatical cases — came through our torture test without a scratch.

That last result surprised us too. Here's the whole experiment.

How We Tested It

Everyone ranks the "hardest languages" by counting characters and grammatical cases. That tells you how hard a language is for a human student. We wanted to measure something different: how hard is a language for a machine — how much meaning actually gets destroyed when a translation passes through it?

Translation Mixer is, conveniently, a machine for destroying meaning. So we ran a controlled experiment:

1. Take one English sentence loaded with translation traps — an idiom, an ambiguous pronoun, a conditional past tense, and the word "singlehandedly":

   "If you had told me yesterday that my grandmother would beat the entire chess club singlehandedly, I would have laughed it off — yet here we are, eating our words."

2. Round-trip it through a single candidate language and back to English (English → Japanese → English).

3. Feed the result back in and do it again. Three round trips per language — the photocopy-of-a-photocopy method. Damage compounds, and the language that mangles meaning fastest reveals itself.

We tested the six languages that top every "hardest to translate" list: Mandarin Chinese, Japanese, Korean, Arabic, Finnish, and Hungarian.

The Damage Ranking

From most destructive to least:

1st place: Japanese never stops moving

After three round trips through Japanese, our sentence had become:

"If someone had told me yesterday that my grandmother could beat every member of the chess club by herself, I would have laughed it off. But now I take that idea back."

Every pass changed something. "You" became "someone" (Japanese drops subjects, so the return trip has to guess). "Would beat" softened to "could beat." "The entire chess club" became "every member of the chess club." And the ending drifted from we are retracting our words to we take that back to I take that idea back — the group of embarrassed speakers gradually collapsing into one person quietly changing their mind.

None of these are errors, exactly. They're guesses forced by a language that routinely omits the subject, marks politeness instead of plurality, and trusts context to carry what English spells out. Round-trip it, and the guesses stack.

2nd place: Korean, where "we" becomes "I"

Korean showed the same disease: by round three, "we are taking our words back" had become "I think I have to take that back." Along the way my grandmother briefly became Grandma — Korean kinship terms don't carry the possessive, so the machine had to decide whose grandma she was, and it decided she was everyone's.

3rd place: Chinese commits, hard

Chinese did its damage in one pass and then stopped:

"If someone had told me yesterday that my grandmother could single-handedly defeat an entire chess club, I would have laughed it off — but now we have to admit we were wrong."

Notice what happened to "eating our words." Chinese has no patience for the idiom — it resolved it straight into its meaning: we have to admit we were wrong. Accurate. Bloodless. The joke is gone and the information remains, which is Chinese-to-English machine translation in a single sentence.

Why Is Chinese So Hard to Translate?

Chinese earns its reputation, but not for the reason people assume. The 80,000-character writing system and the four tones are human problems — a machine handles them fine. What actually breaks translations is that Chinese refuses to encode things English considers mandatory:

• No tense. Verbs don't conjugate. 打败 means beat / beats / beat (past) / will beat, and context decides. On the return trip to English, the machine must invent a tense — in our test, "would beat" came back as "could defeat," a confident prediction downgraded to a mere possibility. We pushed harder with a deliberately tense-stacked sentence: "I had been meaning to call you before you left, but by the time I remembered, you would already have gone." One round trip through Chinese returned: "I was planning to call you before you left, but by the time I remembered, you were already gone." Past perfect progressive and conditional perfect, both flattened to simple past. The meaning mostly survives; the grammar came back in a smaller box. (Run this one yourself →)

• No plurals. "The entire chess club" vs. "an entire chess club" — Chinese has no articles either, so both drift freely.

• No gendered spoken pronouns, and dropped subjects everywhere. Whoever "she" is has to be re-guessed on the way out.

• Kinship precision English lacks. Here's our favorite. We ran this deliberately pronoun-soaked sentence through Chinese and back:

  Original: She told her sister that her daughter had lost her keys, but luckily she found them before she got home.

  Chinese: 她告诉妹妹女儿把钥匙弄丢了，但幸运的是，女儿在回家前找到了钥匙。

  Back to English: She told her sister that her daughter had lost her keys, but luckily, her daughter found them before she got home.

  Two things happened in there. First, Chinese forced a decision English never made — 妹妹 means younger sister specifically, because Chinese has no neutral word for "sister." Second, the original never says who "she" is in "she found them" — the teller? the sister? the daughter? Chinese picked the daughter (女儿), committed, and the ambiguity came home resolved. The round-tripped sentence is more specific than the original. Translation added information — it just had to make it up. (Try it with your own ambiguous pronouns →)

So Chinese-to-English translation is a constant game of the machine filling in blanks that Chinese deliberately leaves open, and English-to-Chinese is the machine deleting distinctions Chinese refuses to store. Either direction, information gets manufactured or destroyed. That's why "why is Chinese so hard to translate" has a real answer beyond "the characters look difficult": the two languages disagree about which facts a sentence is required to contain.

The Upset: Finnish Doesn't Break

Finnish is the language people bring up to sound smart about hard languages. Fifteen grammatical cases! Agglutination! Words like epäjärjestelmällistyttämättömyydellänsäkäänköhän!

Our sentence went through Finnish three times and came back with one trivial change: "singlehandedly" → "by herself." The idiom "eating our words" — which Chinese, Japanese, Korean, and Arabic all destroyed — survived all three round trips intact. Hungarian, Finnish's agglutinative cousin, did nearly as well (it merely turned "laughed it off" into "laughed out loud," making the narrator a bit less dismissive and a bit more of a jerk).

Why? Because grammatical complexity and translation difficulty are different things. Finnish stacks suffixes in ways that torment human learners, but it encodes the same information English does — tense, number, who did what to whom — just with different machinery. The machine converts between machineries fluently. What machines can't do is conjure information a language never wrote down. That's why subject-dropping Japanese beat case-stacking Finnish by a mile.

The languages that are hardest for you to learn are not the languages that are hardest to translate. That's the actual finding, and nobody's flashcard list will tell you it.

What Happens in a Longer Chain

Single round trips are a controlled experiment. Chain several hard languages together and the guesses start compounding into something stranger. Here's David Bowie's Space Oddity through English → Abkhaz → Hindi → Dutch → English:

This is the ground check for Major Tom. You have made a very good assessment. The administration needs to know why you are wearing a shirt. If you mean it, it is time to leave the capsule.

"The administration needs to know why you are wearing a shirt." Mission-control drama, reborn as a bureaucratic dress-code inquiry.

And sometimes a long chain lands somewhere legitimately eerie. Macbeth's "Tomorrow, and tomorrow, and tomorrow" soliloquy, routed through ten languages — Japanese, German, Korean, Hungarian, Finnish, Russian, Icelandic, Traditional Chinese, Arabic, and Thai — came back as:

Tomorrow, tomorrow, tomorrow — these small steps continue into eternity. Our days will eventually end, buried beneath the dust of fools. Go, go, flickering candle! Life is merely a flickering shadow, like a pathetic actor performing for an hour or two on stage. A failed actor, gone forever. It's like a fairy tale told by fools, full of noise and rage, but utterly meaningless.

"Full of sound and fury" became "full of noise and rage." "Signifying nothing" became "utterly meaningless." The machine didn't understand Macbeth. But it found him anyway.

Run the Experiment Yourself

Our test sentence is loaded and waiting — one click sends it through all six contenders in a single chain (Chinese → Japanese → Korean → Arabic → Finnish → Hungarian), which is far crueler than anything we did above:

→ Run the hardest-languages gauntlet on your own sentence

Swap in your own idiom-heavy sentence and see which language breaks it first. If you get something better than "the administration needs to know why you are wearing a shirt," we want to see it.

---

Related reading: We Round-Tripped One Paragraph Through 20 Languages scales this experiment up to twenty languages with damage scores for each. Why Machine Translation Humor Works explains the linguistics of why these failures are funny. Lost in Translation: The Loanwords That Weren't stress-tests borrowed words the same way. And A Brief History of Machine Translation covers how we got machines this good at guessing in the first place.