Google Translate vs HardTrans: real examples

Author: Yuri Vodostoy — Technical translator and founder of Insight Translation · Updated 2026-09-09

Google Translate on the left, HardTrans on the right: four passages, three language pairs into English, in engineering, legal, medical and scientific text. They are short passages, not whole documents, and under every one is a note on what a reviewer looks at first. Three panels ran with no glossary. The fourth ran with a glossary loaded.

How to read these panels

Left panel, Google Translate. Right panel, HardTrans, with the header saying how it ran: no glossary, or with your glossary loaded. Both panels translate the same source passage, so words facing each other came from the same source. Green marks the form we would keep; red marks the form a reviewer would challenge. Sometimes that is a point of English, sometimes only the form this document has settled on; each section says which. The calls are ours, made by hand, not scored by a tool; who makes them is on the about page. Three sentences show a term choice, a register choice and how a modal verb lands. They do not show whether the term still holds on page 40.

One German noun, Leistung, covers both the rated figure of a machine and the way that machine performs. A specification uses one sense at a time, and every instance has to come back in English as the same term. Three sentences about a pump, its acceptance report and its seals:

Engineering — German → English

Google renders «Leistung» two ways, output then performance; HardTrans keeps it consistent, turns the literal «must» into normative «shall», and picks the right verb, «inspected». With your glossary it also locks house terms, for example casing, or the steel grade as X6CrNiTi18-10 (AISI 321).

What to look at

• One term, two targets. Output and performance on the left are two English terms for one source noun inside three sentences. Power output stands twice on the right.
• must against shall. Both are grammatical English. In requirement drafting shall carries the obligation, and ISO/IEC Directives, Part 2 keeps must for constraints that come from outside the document rather than from the document itself. A specification built on must reads like a different genre of text.
• checked against inspected. The clause ends in an acceptance report. Inspection is the act a report of that kind records, and the verb sets the reader's expectation about whether the result is written down.

None of that needed a glossary. The house terms named in the caption, the casing and the steel designation, are the part that does.

French damages law splits loss into what can be counted in money and what cannot. English judgments call the second head non-pecuniary damage. Moral damage is the French shape carried across word for word, and it appears in translated texts often enough to look settled, which is what makes it the easier of the two to keep by accident.

Legal — French → English

Even with no glossary, HardTrans reaches for the term of art, «non-pecuniary damage» rather than the literal «moral damage», plus the binding «shall be deemed». With your glossary it standardises pecuniary / non-pecuniary loss throughout. For legal documents, professional review is recommended.

What to look at

• Term of art against calque. Non-pecuniary damage is the form an English court uses. Moral damage is the French wording in English words, and nothing in the sentence flags it.
• shall be deemed against is deemed. Present simple states a fact about the world. Shall be deemed states the consequence the contract imposes. In a clause about what happens to a contrary provision, that difference is the sentence.
• claim against request compensation. Claim is the verb that goes with a right of action; request reads as correspondence.

No glossary was loaded here. Where a house style prefers pecuniary and non-pecuniary loss over damage, that is one entry, applied to every file in the job.

Medical paperwork exists in registers. The leaflet in the box and the summary of product characteristics for the prescriber describe the same product in fixed, and different, words. The instruction to take a tablet has a set form in each, and the two panels below split on it.

Medical — German → English

With no glossary supplied, HardTrans still uses the leaflet register, «is to be administered». With your glossary it also locks the house terms, such as medicinal product and symptoms. For medical documents, professional review is recommended.

Why this wording is fixed by the pharma template, and where adverse event and adverse reaction stop being interchangeable — in the medical document translation guide.

What to look at

• should against is to be. Should reads as advice. A dosing instruction in product information is not advice, and is to be administered is the form that carries it.
• One source noun, two defensible English words. Discomfort names a sensation and narrows the source; the right-hand panel takes the broader clinical noun. Neither is wrong in the abstract, and the red mark means only that a reviewer would query discomfort against the register the document keeps, not against English. That is why the house term the caption names, symptoms, belongs in a glossary rather than in an argument between reviewers.

Section headings, frequency categories and the reporting formula that closes section 4.8 are looked up rather than translated, and no engine guesses them reliably.

Purity test is ordinary English, and with no glossary HardTrans writes test here too. The red mark on this panel is narrower than on the others: it is not a verdict on the English, but on the form this document has fixed. The house term list says assay, and the right-hand panel is HardTrans with that list loaded.

Science — Spanish → English

Here generic MT and HardTrans agree out of the box — both write «test». This is exactly where a glossary earns its place: tell HardTrans your house term is «assay» and it holds it everywhere, as an invariant.

What to look at

• The rest of the passage is identical. One word is the whole difference between the panels, and neither engine had anything in the sentence to decide it on.
• The red mark is a glossary decision, not a language one. The caption says it: out of the box both engines write test. What the panel shows is one run, HardTrans with the term list loaded, and the single word that list changes.
• A glossary records decisions your team already made. One entry holds everywhere in the file, including the places where the surrounding sentence gives no clue, and including the sentence you would never have thought to check.

If a page tells you a tool is better everywhere, this is the panel it left out.

What these panels cannot show

A term, a register and a modal verb fit inside three sentences. These do not:

Consistency across a document. We test that on whole files. On a German regulation, August 25, 2026, Google Translate needed 13 different English forms for six recurring terms, DeepL 8; HardTrans used exactly one form per term, six for six. That run is on the DeepL alternative page; the drift behind it is in why machine translation breaks on technical documents.

Tags. Every panel here is plain prose. A real file carries inline tags — bold, a footnote, a placeholder your CAT tool inserted — and HardTrans keeps them through the round trip: the same tags, the same file structure, out as in. See translating XLIFF online.

DeepL. There is no free anonymous DeepL API, so any DeepL output we publish is a dated manual capture, and those runs live on the DeepL page.

A human reader. There is no human in our loop. Anything carrying liability needs a specialist before it goes out, and we refuse a scanned PDF with no text layer.

Method

June 2026. Four passages, three language pairs into English: a pump specification from German, a damages clause from French, medical product information from German, a laboratory method note from Spanish. Each went through Google Translate (public endpoint), pasted without edits, and through HardTrans. The panels carry what each engine returned, including the sentences where they agree.

Three passages ran with nothing supplied but the text itself. The science panel is the run with a glossary loaded: the house term arrives as an entry and replaces the general word wherever it occurs.

We do not publish those passages as source files, and we do not rewrite a panel later. A newer run gets a new panel and its own date.

Common questions

Does Google Translate keep terminology consistent across a long document?

In the engineering panel it does not. One German noun, Leistung, comes back as output and then performance inside three sentences, and nothing in the text marks the two as one term. Across a whole file the drift is counted rather than eyeballed, and the dated whole-document run sits on the DeepL alternative page.

Can I upload my own glossary?

Yes. CSV, TSV, TXT or a Smartcat XLSX, several files at once, applied to every file in the job rather than to one paste. Before you pay, the estimate shows the term count and any conflict where one source term has two targets.

Do I need a glossary to get results like these?

Three of the four panels ran without one. A glossary settles the cases where more than one rendering is correct and your house has picked one, which is the science panel exactly.

Can a three-sentence passage prove terminology consistency?

No, and it is not meant to. Consistency is a property of a document and it is measured on one. Read the panels here for term and register choices, and the dated whole-document runs for consistency.

What happens to my files?

Files are deleted from our servers automatically after 7 days. Your text is never used to train models — neither by us nor by Anthropic, the model provider. Anthropic retains API data for a limited period under its policy and does not train on it.

If you are still choosing between a general engine, a chat window and a document tool, that comparison is in what makes an AI translator good for documents.

Upload the file you are actually stuck on: a specification, a clause, a section of product information. $1.50 per 250-word page, first 10 pages free, no signup.

Real output: Google Translate vs HardTrans, June 2026. The first three examples used no glossary; the science example shows a glossary term. For legal and medical documents, professional review is recommended.