Machine translation has become good enough that most of the time you can simply read the result and move on. The trouble is the rest of the time: the output still reads smoothly, so nothing warns you that the meaning has changed. This guide lists the errors that show up most often and gives you simple checks that work even if you do not speak the source language.
1. Dropped or flipped negation
Negation is small ("not", "no", "never", a prefix, a particle) and easy for a model to lose, especially in long sentences or languages where negation is split across the sentence, such as French "ne ... pas".
How to spot it: if a translated sentence says the opposite of what the surrounding text implies, or a warning suddenly reads like a recommendation, look at the original for negation words. Even without knowing the language, you can often recognise them after seeing them once or twice.
2. Wrong numbers, dates and units
Numbers are usually copied correctly, but the formatting around them is not always adapted. A European "1.500" means one thousand five hundred; to an English reader it looks like one and a half. Dates like 03/04 flip between day-first and month-first conventions, and units may be translated as words without converting the values.
How to spot it: for anything involving money, dosage, time or measurements, compare the digits directly with the original rather than trusting the translated sentence.
3. The wrong meaning of an ambiguous word
Many words have several meanings: "bank", "light", "charge", "mean". The model chooses based on context, and when the context is thin it chooses the most common sense. A message about "charging" a customer can come out as charging a battery in a language that uses different words for the two.
How to spot it: watch for a single word that does not fit the topic of the paragraph. That odd word is usually the one with several meanings.
4. Pronouns pointing at the wrong person
Languages handle pronouns differently. Some drop the subject entirely ("went home" instead of "she went home"), some have gendered "it", some have none. When translating into English, the model has to invent a pronoun and sometimes invents the wrong one, so "he" and "she" swap or a thing becomes a person.
How to spot it: in any story with two or more people, check who does what. If the actions seem to switch between characters for no reason, the pronouns are probably wrong.
5. Wrong level of politeness
English has one "you". German, French, Spanish, Russian, Japanese, Korean and many others distinguish between formal and informal address, and Japanese and Korean go much further with whole levels of politeness. A translator working into those languages has to choose, and in a business email the wrong choice can come across as rude or oddly distant.
How to spot it: you usually cannot without knowing the language, so prevent it instead. Some translators let you pick formal or informal tone. Otherwise, a short, polite, complete sentence is more likely to come out formal than a clipped one.
6. Idioms translated word for word
Well-known idioms are usually handled, because they appear often in training data. Less common ones, regional expressions, slang and jokes often come through literally. "It's raining cats and dogs" is fine in most major languages; a new internet meme or a local proverb may turn into nonsense or, worse, into something that sounds meaningful but is not.
How to spot it: a phrase that is grammatically fine but strangely vivid or out of place ("he pulled my leg" in a medical context) is a strong hint that an idiom was taken literally.
7. Silent omissions
The most dangerous error is text that is simply missing. Neural models sometimes skip a clause, a list item or a qualifier like "only" or "at least", and the resulting sentence is perfectly fluent. It is also common for untranslatable fragments (names, codes, emoji-only lines) to be dropped.
How to spot it: compare lengths. If a paragraph with five sentences becomes four, or a list of six items becomes five, something went missing. Different languages have different lengths, but the number of sentences and list items should match.
Quick checks that work without knowing the language
- Back-translation. Translate the result back into the original language (or into English) with a different engine. If the meaning survives the round trip, it was probably right.
- Two engines. Translate the same text with two services. Where they agree you can relax; where they differ is exactly where the ambiguity lies.
- Count things. Sentences, list items, numbers, names. They should all match the original.
- Read the surrounding text. A sentence that contradicts its neighbours is more likely to be mistranslated than to be a surprise.
- Translate more, not less. Give the translator the whole paragraph instead of one line. More context almost always means fewer mistakes.
When to stop relying on the machine
For reading social media, news, comments and everyday chat, machine translation is more than good enough, and the checks above catch most of the important slips. For contracts, medical instructions, legal notices, anything you will publish under your name, or anything where a single wrong word costs money or safety, use a professional translator or at least a fluent speaker to review the text.
The bottom line
Modern translation errors are quiet: no broken grammar, no obvious warnings. Once you know the seven usual suspects (negation, numbers, ambiguous words, pronouns, politeness, idioms and omissions), you will start to notice them, and a few quick checks will make you a far more confident reader of text in languages you do not speak.