How does AI translation accuracy change between a contract, a manual, and an ad?
Accuracy changes with the job. A contract is judged on duties and defined terms, a manual on names, steps, and numbers, and an ad on whether the line still does its work for that occasion. The same engine can mark all three as fine and be wrong in a different place on each file.
On a Friday a product manager dropped three files into one window: a supply contract, a valve manual, and a six-word line for a summer booth. The screen put a green mark beside each. On Monday the lawyer found that “shall” had softened in the indemnity. The technician found the valve called by the spare-parts name. The booth line rhymed, and it invited people to a season the campaign had already left.
Three jobs, three ways to be wrong
A contract is a stack of obligations. The costly misses are short words and names that must stay put: shall, may, including, and a defined term such as Affiliate that has to mean the same person from the first clause to the last schedule. A sentence can sound like ordinary English after the duty has been narrowed. Someone who is only asking “does this read well” will pass it.
A manual is a set of parts, warnings, and sequences. The costly misses are a renamed fitting, a torque that slid into the next row, or a figure callout that still points at last week’s drawing. Fluency is a weak test here. The person at the bench has to finish the step with the part that is actually in the box.
An ad is a line for a moment, a place, and a reader who will not study it. A word-for-word rendering can be the failure. A pun, a rhyme, or a local joke can be carried across with every word accounted for and still be useless on the booth wall. The question is whether the line performs the original job. A match score does not ask that.
Where the contract actually broke
The indemnity said the supplier shall cover claims arising from the goods, including claims brought by subcontractors. The draft said the supplier should cover claims about the goods, and then stopped. One verb softened the duty. The shortened list left the subcontractor claims outside the sentence. Both edits looked small on a screen. Both changed what the clause would do in a dispute.
Defined terms break in the same quiet way. Affiliate was capitalised because the contract had already said who that is. The draft varied it, once as “related company,” once as “partner,” because variety reads more naturally to a model that was praised for style. In a contract, that variety is a new party.
What the booth line needed instead
The booth line had to be said aloud in a noisy hall, in a language where the original rhyme has no twin. Keeping every source word produced a sentence a bilingual person could map back, and nobody would print it on a banner. The manual’s rule, same part and same name, would have made the banner worse. The contract’s rule, same duty and same scope, does not govern a joke.
Check the three files as three different readings. On the contract, compare operative verbs and defined terms with the source, clause by clause. On the manual, read names, numbers, warnings, and cross-references against the drawings. On the ad, read the line in the place it will be used, and be willing to rewrite it when the occasion demands a new sentence. One shared score hides that switch.
Does a high score on the manual say anything about the contract?
It says the engine handled a lot of repeated technical sentences, and it says nothing about the handful of verbs that carry the contract. Read those verbs against the source even when the dashboard is green.
Should the ad be given an accuracy score at all?
A score can show whether words were carried across, and it cannot show whether the line belongs on that wall, in that week, for that reader. Have a person read it where it will actually be seen.
The related service page is machine translation.
Quick FAQ: AI Translation Accuracy
- How accurate are AI translators? Accuracy is often high for repetitive or general content, while domain-sensitive content still needs expert review.
- How to improve AI translation quality? Use glossary control, domain prompts, QA checks, and human post-editing in one workflow.
- Where does human translation still win? Legal, medical, and high-stakes brand content usually requires human nuance and accountability.