We Round-Tripped One Paragraph Through 20 Languages. Here's the Damage Ranking.

Updated 2 August 2026. The original version of this post used damage scores measured against a different translation engine. We re-ran the entire experiment on Google Translate, which is what the mixer serves by default, and replaced every number below. The rankings changed. What that means is discussed at the end, because it turned out to be the most interesting part.

Every list of "hardest languages to translate" you can find online ranks languages by how hard they are for a person to study. Character counts, grammatical cases, tone systems. Useful if you're picking a language class. Useless if you want to know what actually survives a trip through Google Translate.

So we measured it directly. One paragraph, 20 languages, one round trip each, scored for damage.

The Method

We wrote a paragraph designed to carry the things machine translation tends to drop: a specific person, an ambiguous pronoun chain, a couple of tense shifts, and some ordinary sentiment that a machine might flatten.

When my grandmother passed away last spring, she left behind a small wooden box full of old photographs. I hadn't realized how much she meant to us until I sat there flipping through pictures I'd never seen before, each one telling a story I would probably never fully understand. It made me wonder what memories I would leave behind someday, and whether anyone would take the time to look.

Each language got the same treatment: English in, one hop out, one hop back. No chains, no stacking, nothing designed to inflate the numbers.

For scoring we used the same word-overlap calculation the mixer runs on every translation, which compares the set of words in the original against the set in the result. Zero means the text came back identical. One means nothing survived. It's a blunt instrument that misses meaning shifts when the vocabulary happens to match, so we read all 20 outputs by hand as well.

The Damage Ranking

Measured on Google Translate, August 2026:

The top three are the three languages in this set that routinely omit the subject of a sentence. That is not a coincidence, and it is the whole finding.

First Place: Japanese

Japanese did the most damage, and the interesting part is that the output reads perfectly well:

When my grandmother passed away last spring, she left behind a small wooden box filled with old photographs. It wasn't until I sat there flipping through photos I'd never seen before that I realized how important she was to us. Each one told a story I would probably never fully understand. It made me wonder what memories I would leave behind someday, and whether anyone would take the time to look at them.

Nothing there is wrong. The sentence boundaries moved, "meant to us" became "was important to us," and the middle clause was restructured into a cleft. It is a competent paraphrase rather than a translation, which is precisely what happens when a language forces the machine to rebuild rather than convert.

Japanese drops subjects, drops objects, and doesn't mark plural. Going in, all of that gets stripped. Coming back, English requires every one of those, so the machine reassembles the paragraph from the meaning rather than the structure. We took Japanese apart in detail in a separate post, because the failures get much stranger than this once you stop feeding it easy prose.

Korean and Chinese, second and third, did the same thing for the same reason.

What Sits at the Bottom

Icelandic and Finnish barely moved the text at all, at 0.033 and 0.049. Both came back nearly word for word.

That surprises people, because Finnish has fifteen grammatical cases and a reputation for being impossible. Its reputation is about learning difficulty, which turns out to have almost nothing to do with translation difficulty. We wrote up why the agglutinative languages score so low separately.

The useful lesson from the bottom of the table is that structural similarity to English is not what protects text. Hindi placed 10th despite a different script, different word order, and grammatical gender English lacks. What protects text is a language marking the same things English marks.

The Pronoun Problem

Nine of the 20 languages here use a third-person pronoun that doesn't specify gender: Finnish, Hungarian, Turkish, Basque, Georgian, Swahili, Zulu, Mongolian, and Tagalog. When the paragraph passes through one of them, the fact that the grandmother is a woman stops existing, and the return trip has to guess.

Four of the nine guessed wrong.

Hungarian, Basque, and Georgian all demoted her to "it," as in "I didn't realize how much it meant to us." Tagalog changed her to "he." The other five, including Finnish and Swahili, happened to recover the correct pronoun, most likely because "grandmother" appears earlier in the same paragraph and gave the model enough to work with.

Georgian added a second error on top:

When her grandmother died last spring, she left behind a small wooden box full of old photos.

One word, and the paragraph is now about somebody else's family.

A coin-flip failure rate on a detail this basic is worth knowing about. The machine isn't confused, and the Georgian sentence is a correct translation of what it received. The information was destroyed on the way in, and on the way back the machine filled the hole with a guess.

What This Means If You're Building a Chain

If you want a chain that mangles text, reach for the subject-droppers. Japanese, Korean, Chinese, and Mongolian all force the machine to rebuild rather than convert, and two of them in a row compounds, because the second one re-guesses whatever the first one guessed.

If you want a chain that preserves meaning, stay with languages that mark subject, number, and gender explicitly. Icelandic, Finnish, and German will hand your sentence back nearly intact.

One caveat on the numbers. A single round trip is a controlled test, not a worst case. Every language in that table will do considerably more damage inside a longer chain, and some of the low scorers become genuinely destructive once another language has already scrambled the input.

About Those Numbers Changing

The first version of this post ran the same experiment against a different translation engine, and the ranking was noticeably different. Mongolian placed first instead of fifth. Japanese placed fourth instead of first. Turkish sat near the bottom instead of seventh. The pronoun failure rate was eight of nine rather than four of nine.

We are leaving that on the record rather than quietly swapping the table, because it is a genuinely useful result: the damage a language does is a property of the engine as much as the language. Every ranking of this kind you will read, including this one, is measuring a specific model on a specific day.

What survived the engine change is the part worth trusting. Subject-dropping languages did the most damage under both engines. Finnish and Icelandic finished near the bottom under both. The pronoun failure showed up under both, at different rates. Those are properties of the languages. The exact ordering is not.

If you run the chain below and get numbers that don't match ours, that is the experiment working correctly.

Run It Yourself

The paragraph is loaded and pointed at the top four in a single chain, which is much harsher than anything we did above:

→ Run the top four through your own paragraph

Swap in your own writing and see which of your sentences the machine has to guess at. The ones about people usually go first.

---

Related reading: Why Machine Translation Still Struggles With Japanese takes the first-place finisher apart line by line. Finnish and Hungarian Don't Break Translators explains why the languages everyone calls impossible finished at the bottom. And What's the Hardest Language to Translate? covers the iterated version of this experiment.