When machine translation is the right tool

Start with the part nobody in this trade says out loud

Machine translation is very good, it is getting better, and this office sells it. It is on the rate card at 11 cents a word for the commonest languages, roughly half of 24 cents a word for the same text translated from scratch, and the same person's name goes on the result as on any other job here.

A company arguing that the machine cannot do work it is quietly using the machine for is not making an argument. So this page is the honest half of the case, and the one absolute limit is at the end where it belongs rather than at the front doing the selling.

Four places the cheaper option plainly wins

Understanding a document that has arrived. A letter comes from an office abroad, or a landlord, or a hospital, and what you need this afternoon is to know what it says. Paste it into a machine. It will be right about the substance far more often than not, and it costs nothing. If what it says turns out to matter, buy the translation then — and you will buy the right one, because you will know what the document is.

Triaging something long before you spend anything. The use that saves the most money and almost nobody thinks of it. A hospital file runs to a hundred pages and eleven of them are relevant. Run the whole thing through a machine, work out which parts matter, and have those parts done properly: this office would rather quote for eleven pages than sell you a hundred.

Internal drafts. Anything read at speed by your own colleagues, for the gist of it. A memo, an internal e-mail, a set of notes, a first look at a supplier's terms before anybody decides whether they are worth negotiating.

Repetitive technical text. Technical documents repeat themselves, which is why this is the cheapest of the professional rates and the work where machine translation with a human editor genuinely earns its keep. The part that is not cheap is the safety text: a warning is written to be obeyed by somebody holding a tool, and a warning that reads as a suggestion has failed even if every word of it is correct.

What the middle option actually buys

Between the free tool and full human translation there is the service on the rate card, which is machine output read end to end against the source by a translator and corrected. We offer this openly, we price it at about half, and it is on the rate card next to everything else. The reason is simple: a company arguing that machine translation cannot do certain work, while quietly using it for the work it can do, is not making an argument. What you get is machine output read end to end by a translator against the source, corrected, and delivered with the same person's name on it as any other job. What you do not get is a discount on the reading — the reading is the work, and a review that skims is the failure mode this service has.

What that includes is worth being specific about, because the phrase is used loosely.

• Every line compared against the source, not spot-checked
• Terminology forced to a glossary, which is the thing raw machine output is worst at across a long file
• The same translator answerable for the result as on any other job here
• A note of anything the machine got confidently wrong, because that is what you are paying to have caught
• A candid answer about whether your file is suitable, before you buy it

Four places it costs more than starting clean

Anything that needs a certificate. Anything that needs a certificate, because the certificate asserts a person did the translating

Marketing copy. Marketing copy, where the machine draft anchors the writing to the wrong sentence and costs more than starting clean A headline that works in Spanish is frequently not a rendering of the English one at all, and reading a literal draft first anchors the writer to a sentence they then spend an hour escaping.

Handwriting and poor scans. A machine reads what it can see; what it cannot see, it does not mark as unreadable but replaces with something plausible, and plausible is the failure you cannot catch by reading the output alone.

A document whose layout carries meaning. A table where the column decides what the number means, a form where a tick in the third box is the whole content, a record where a marginal note qualifies the entry beside it. Flattened into a stream of sentences, all three lose what they were saying.

The one place it is not about quality at all

Everything above is a judgement about which tool is cheaper for the job. The limit below is not a judgement and it does not move as the models improve.

That is not the case and this office will not argue it. Machine translation is often very good and we sell it, with human review, at about half the rate. The real limit is narrower and it is absolute: a certificate of accuracy is a statement by a person who can be named, questioned and held to it, and a model is not a person who can be any of those.

Look at what the certificate asserts. Somebody is competent in both languages. Somebody has compared the translation against the original and says it is complete and accurate. Somebody's name is on it, and if that turns out to be false there is a person to answer for it — in a deposition, in a professional complaint, in a claim against an insurer. Every one of those is a fact about a legal person, and a model has no legal identity to put in the box. That is a statement about accountability rather than about quality, and two more follow from the same place: a state appointing sworn translators appoints a named individual, and an interpreter in a proceeding takes an oath.

And one thing to think about before pasting

A document you paste into a consumer service is a document you have disclosed. For a news article that does not matter at all. For a sealed record, a client's file or somebody else's medical history, it is a decision that deserves to be taken deliberately rather than at eleven at night with a deadline in the morning.

Everybody who touches a file is under a written undertaking, and that includes the freelance translators. Documents are not put through any service we have not contracted with, which is a sentence worth reading twice in a trade where pasting a page into a free website is a keystroke away. Ask us what a file will touch and we will tell you.

So: use the machine to understand, to triage and to draft. Buy the middle option where somebody has to stand behind a long repetitive file. And where a person has to answer for the result, buy a person, because that is the one thing on sale here a machine cannot be.