A hiring manager in Lyon once forwarded a candidate’s CV with a single line highlighted in yellow: “Célibataire des Arts.” The candidate had run their own resume through a free translation tool before applying, and “Bachelor of Arts” had come back as “single in the arts.” Nobody caught it until a recruiter did, and by then the candidate had already wondered, understandably, why nobody had replied.
The tool wasn’t broken. It did exactly what it was built to do: match words to words. The problem is that “bachelor” has two unrelated meanings in English, and a system translating word by word has no way to know which one a French reader needs until it understands the sentence around it. That’s the whole story of why literal translation produces awkward, sometimes wrong, English. It optimizes for the wrong unit.
Word-for-word is not the same as meaning-for-meaning
Literal translation swaps each word or phrase for its closest equivalent and preserves the original sentence structure as much as possible. It’s fast, and for short, unambiguous text, it’s often fine. “The meeting is at 3pm” survives literal translation into most languages without incident.
Idioms, figures of speech, and culturally loaded phrases don’t survive it. “Break the ice,” translated word for word, describes a strange scene involving actual ice in most other languages. The meaning, “ease social tension,” has nothing to do with the literal words, which is exactly why a translator, human or machine, has to translate the idea rather than the sentence.
The same source sentence can branch into a literal reading or a meaning-based one — only one holds up.
This is where things get genuinely interesting, because even sophisticated AI translation systems don’t always agree on what the idea is. A recent test that ran the same French idiom through multiple GPT and DeepSeek models found something worth paying attention to: on a well-documented phrase, every model converged on the identical English equivalent, but on a less rigidly documented one, the models split cleanly along family lines, each producing a different, equally valid English phrase. None of the outputs were literal. All of them were translating meaning rather than words. They just weighted the available meanings differently.
That’s the real lesson buried in literal translation failures: the fix isn’t simply “translate meaning, not words,” because even meaning-based translation has more than one correct answer. The fix is knowing which meaning your context calls for, and no translation engine, human or artificial, can know that without more information than a sentence usually provides on its own.
Where awkward English shows up in real writing
Once you know what to look for, literal-translation awkwardness turns up everywhere: instruction manuals with sentences that are grammatically correct but nobody would ever actually say, marketing copy that reads flat because a play on words got translated into words instead of into a new play on words, and business emails where a polite hedge in the source language becomes a blunt statement in English because the softening phrase had no direct equivalent.
Job postings and recruitment materials are a particularly clean example, because the damage is measurable. A recruitment content review from Ongig found that global talent acquisition teams who tested unsupervised auto-translate features on job descriptions kept hitting the same failure: the translated text looked clean on screen, but candidates flagged errors anyway, because the source English was full of idioms and internal jargon that were never going to translate cleanly no matter which engine ran them. “We move fast and wear many hats” reads as energetic in English. Translated literally, it can read as chaotic, or simply confusing.
CVs carry the same risk in the other direction. Beyond “Bachelor of Arts,” recruiters regularly see job titles and qualifications that shift meaning entirely once translated word for word, turning a mid-level manager into something that reads like an executive title, or a freelancer into something closer to “wanderer.” None of these are the translator’s fault in the way a typo is a writer’s fault. They’re a structural consequence of translating words instead of context.
What actually works better
Four adjustments consistently reduce the awkwardness, and none of them require abandoning machine translation.
Write the source text to survive translation. Idioms, sports metaphors, and regional slang should get flagged and rewritten in plain language before translation starts, not fixed afterward. A sentence that’s clear and literal in the source language is far more likely to translate cleanly, because there’s less ambiguity for any engine, human or AI, to resolve. Before sending anything for translation, it’s worth running the source text through a plain-language readability check to catch the kind of idiomatic phrasing that reads fine to a native speaker and translates into nonsense.
Compare outputs instead of trusting one. Industry testing backs this up at scale. Slator’s coverage of a recent translation automation study found that baseline machine translation systems averaged 10 to 15 errors per text, while requirements-based, customized approaches cut that to as few as zero to two. The gap wasn’t about which single engine was “best.” It was about whether the process accounted for more than one possible reading of the source text before settling on an output. When a phrase genuinely has more than one valid translation, seeing that spread before publishing is more useful than getting a single confident answer.
Source: Intento, State of Translation Automation 2025 (via Slator).
Get a second pass from someone who didn’t write the original. Grammatically correct and unambiguous are not the same thing, and a sentence that’s already a little unclear in the source language will only get less clear once it’s translated. A quick grammar and clarity review of the source text, done by someone other than the original writer, catches the kind of ambiguity a fluent writer stops noticing in their own draft.
Run a back-translation spot check. Translating a finished passage back into the source language and comparing it to the original is a blunt tool, but it’s a fast one. It won’t catch every register mismatch, and it can flag false positives on phrases that were correctly adapted rather than literally rendered. What it’s good at is surfacing the sentences where something clearly went sideways: a negation that flipped, a name that got translated as if it were a common noun, a number that shifted. Treat it as a smoke detector, not a proofreader. It tells you where to look closer, not what to fix.
The takeaway
Literal translation isn’t wrong because it’s careless. It’s wrong because it treats language as a substitution problem when it’s actually a comprehension problem. Every awkward sentence in translated English, from a mistranslated CV line to a job posting that reads as chaotic instead of energetic, traces back to the same root cause: words were matched before meaning was understood. Fixing it doesn’t require perfect translation. It requires writing source text that leaves less room for ambiguity, and building in a check, human or comparative, before anything goes out the door in a language its author doesn’t read.
