A reading aid, not a source
A translation engine produces the sentence that best fits the context, in the target language, from the source text. For reading a grocery list, a product page, or an unfamiliar article, this is close to magic: the meaning arrives, and you move on.
What it does not produce is a faithful record of what the author wrote. It produces a reading — a good one, usually, but a single reading, made without knowing what you need the sentence for. When that reading is saved as if it were the text, the difference between "translated" and "said" quietly disappears.
Rule of thumb: translation can be how you understand a source. It should never be the source itself.
What it handles well
Translation quality is not uniform, and knowing where it is strong saves you effort in both directions.
- Gist and navigation. Deciding whether an article is worth reading is exactly the kind of task translation is good at. You do not need precision to know it is the wrong paper.
- Straightforward expository prose. News reporting, product documentation, technical descriptions with standard vocabulary — short, declarative sentences come through reliably.
- Familiar domains. The more common a subject, the more training material exists, and the more natural the output.
Where it bends meaning
Negation and scope
A misplaced negation inverts a sentence while keeping its grammar intact. The translated sentence reads perfectly and says the opposite. This is the single most damaging failure, because nothing in the output signals it.
Terms that have no single equivalent
Legal, medical, and administrative vocabulary is full of words with an established translation in the target language that does not mean quite the same thing. The translation sounds professional and carries a different obligation.
Names, units, and numbers
Proper nouns get localised, sometimes into a different person. Units and date formats convert or fail to convert inconsistently. Numbers written with different separators — decimal commas, thousand dots — are a frequent source of silent error.
Idiom and register
Register is the hardest thing to preserve: whether the author was formal, ironic, dismissive, or quoting someone else. A flat rendering can make a criticism sound neutral, or a joke sound like a claim.
Quote in the original, translate beside it
If a sentence will be quoted anywhere — an essay, a report, a slide, a message to a colleague — the quoted text must be the original. Translation goes alongside, clearly marked as yours.
A workable format that takes ten extra seconds:
- the original sentence, in the original script,
- your translation on the next line,
- the source and the date you accessed it.
This survives every later problem. If someone asks what the author actually wrote, the answer is on the same line. If a second translation disagrees with yours, the disagreement is visible rather than buried.
If you keep only the translation, keep it without quotation marks, and label it a translation. A paraphrase that has lost its label becomes a misquote within a month.
Keeping terminology stable
If you read more than a handful of sources in one field, you will meet the same term repeatedly, and translation engines will render it differently each time depending on surrounding words. That drift makes notes hard to search and easy to misread later.
The fix is small: keep a glossary note with three columns — the original term, the translation you have chosen, and one line on why. When a source uses a term that is already in the glossary, use your own translation, not the engine's.
After a dozen entries, the glossary starts doing something useful on its own: it shows which concepts the field distinguishes and your language collapses together.
Where the translation happens
There is a second question that has nothing to do with accuracy: where the text is processed.
Translation run on your own device never sends the text anywhere. Translation sent to a service does — which is fine for a public article and less fine for a contract, a medical letter, or anything covered by confidentiality.
For most reading this distinction never comes up. It becomes relevant exactly once: the day the text is not yours to forward.
When not to translate at all
Some things should be left exactly as written. Titles of works keep their original form when you are citing them, even if you also give the translated title in brackets. Legal citations, standards numbers, product names, and code identifiers are not translated — a translated identifier stops pointing at the thing it names.
Personal names are the same problem: a localised name is a different name, and searching for it later in the original language will not find it.
A practical habit: anything you might have to search for again, or hand to another person, stays in its original form.
A short workflow
- Translate to decide whether to read. Do not save the translated output.
- Read the original. Keep a translation open beside it, not instead of it.
- When something matters, capture the original text.
- Add your translation underneath, labelled as yours.
- Record the source and the date — translated pages are revised more often than people expect, and the wording you relied on may not be there next year.
How Papscroll handles it
Translation works differently across Papscroll's clients, on purpose. The Android app translates on the device with an on-device model, so the text does not leave the phone and works without a connection once the language pair is downloaded. The web app and the browser extension send text to a translation service and return the result.
Either way, the note stores what you captured — and if you keep the original beside your translation, both stay in the same note, with the source attached.