Machine translation has become good enough that its remaining failures are more interesting than its successes.
It fails where the correct depends on something not present in the text.
A in a language without grammatical gender must be assigned one in a language that has it.
Nothing in the sentence determines the answer, and the system supplies the answer its training data made likely.
The result is a sentence that something the original did not, with no signal that this has occurred.
Human translators face the same problem and resolve it by asking or by leaving a note.
Neither option is available to a system that must return a complete sentence.
This is a design rather than a limitation of the technology, and it is rarely presented as one.
A translation interface that could return 'this sentence is in the following respect' would be more useful and less impressive.
is the second trap and is more serious because it is invisible.
Early systems produced output that was obviously wrong and therefore checked.
Current systems produce output that is fluent and occasionally wrong, which is checked much less.
A reader's confidence tracks fluency and fluency no longer tracks accuracy.
Errors that survive are therefore the ones that read most naturally, which is the worst possible selection.
Professional practice has adapted by changing what a translator is paid to do.
a machine draft is now a larger part of the work than translating from nothing.
It requires a different and less pleasant skill: reading fluent text with active .
Translators report that it is harder to maintain for a full day than translation ever was.