The machine drafts, the person signs

The question usually arrives in the form of a verdict already half-made: is machine translation good enough yet? It is the wrong question, and it has been the wrong question for some years now. Automatic translation is very good at certain things and structurally incapable of others, and the boundary between the two does not fall where most people assume. It does not fall along the line of difficulty. It falls along the line of consequence.

What follows describes that boundary honestly, without the two familiar postures — the one that says the machines have already won, and the one that says they cannot really translate at all. Both are wrong, and both are expensive to believe.

What automatic translation is genuinely good at

The current generation of systems does not translate the way the statistical engines of a decade ago did. Those worked phrase by phrase and produced text that announced its own machine origin in every clause. The systems in use today read a whole document, hold context across it, and can be instructed on register, terminology and audience. The output is fluent. Often it is more fluent than a tired human first draft.

Their real advantages are four.

• Volume. A hundred thousand words of internal documentation can be rendered overnight. No human arrangement produces that, and no client should pay for one when the purpose of the text is to be understood rather than to be relied upon.
• Consistency. Given a termbase, a machine will apply the agreed rendering of a term on page four hundred exactly as it did on page one. Humans drift. Teams of humans drift faster.
• Cost at the bottom of the market. Text that exists to be skimmed — a supplier's catalogue, an inbox of routine correspondence, a competitor's press release read for gist — has never justified a professional rate, and now it does not have to. That is a genuine gain, not a loss.
• Reach into languages nobody was serving. Whatever else is true, a great many people can now read a great many things that were previously closed to them entirely. That matters more than the professional anxiety it causes.

What automatic translation is bad at, and why the badness is hard to see

The failures are not distributed the way the successes are, and this is the part most buyers get wrong.

The old machine errors were visibly wrong. They read as broken language, and a reader with no knowledge of the source could tell that something had gone amiss. The current errors read beautifully. A modern system asked to translate a clause it does not understand will produce a confident, idiomatic, grammatically immaculate sentence that says something the original did not say. Sometimes it quietly drops a clause. Sometimes it invents a plausible one. Fluency and accuracy have come apart, and fluency is the thing a non-speaker can check.

This has a practical consequence. The cost of not reviewing machine output has risen, precisely because the output no longer looks like it needs reviewing.

Beyond that, four categories resist automation for reasons that are not about model size.

• Terms of art with no counterpart. Legal and administrative systems are not translations of one another. A Moroccan family-law instrument, a French acte de notoriété, an Anglo-American trust — each is an artefact of its own system, and rendering one into another language is a decision about equivalence, not a lookup. Someone has to choose whether to naturalise the term, borrow it, or gloss it, and that choice has to be defensible to the institution receiving the document.
• Ambiguity that requires asking. Real source texts are underspecified. A pronoun with two possible antecedents, a heading that could be a title or an instruction, a number that could be a date or a reference. A machine resolves these silently, by guessing the most probable reading. A translator resolves them by querying the client. The difference is not intelligence; it is the willingness to admit uncertainty and go and find out.
• Voice. In literary work the propositional content is the least of it. Rhythm, register, the specific temperature of a word, what an author chose not to say — these constitute the work. A machine optimises toward the expected phrasing. Literature is largely made of the unexpected one.
• Accountability. This is the decisive one, and it is discussed least.

The Arabic case, specifically

The general argument sharpens in Arabic, and anyone commissioning Arabic work should know why.

Arabic script omits short vowels. A great many written words are therefore ambiguous on the page and disambiguated only by context — which means the machine's guess is doing more work in Arabic than in French, and its errors are correspondingly quieter.

Then there is diglossia. Modern Standard Arabic is nobody's mother tongue; Moroccan Darija is spoken by tens of millions and written by comparatively few. Training data reflects that imbalance, and output into or out of the dialects is markedly weaker than the headline benchmarks suggest.

Three further hazards are mundane and cause more rejected documents than any of the above:

• Bidirectional text. Latin strings, reference numbers and dates embedded in a right-to-left paragraph are routinely reordered or mangled in ways that survive a casual proofread and fail an official one.
• Proper names. There is no single correct romanisation of an Arabic name. A machine, given the same name in three documents, will often produce three spellings. In an immigration or academic file, that inconsistency is itself grounds for rejection, independent of translation quality.
• Calendars. Hijri and Gregorian dates require conversion, not transcription, and the conversion is easy to get wrong by a day.

None of these are exotic. They are the ordinary texture of official Moroccan paperwork.

Certified work, and the thing no model can hold

Here the argument stops being about quality at all.

In Morocco, an official translation is not made official by being good. It is made official by a sworn translator — a named individual, appointed by a court, entered on a register — issuing it on secured paper supplied by the authority, under that translator's own registered stamp. The legal weight sits on a person who can be identified, questioned, and held responsible.

That is not a quality standard a system could eventually meet. It is a structure of liability. A model has no name on a register, no appointment to lose, and nobody to answer for it if the rendering of a clause costs someone a visa or a contested inheritance. Receiving institutions are not asking whether the text is accurate; they are asking whose signature stands behind the claim that it is.

Which is why the sensible way to describe the arrangement is not human versus machine but draft versus signature. Tools may sit anywhere in the production of the text. The signature is a human act, and it is the only part of the process that is legally load-bearing.

How the two actually coexist right now

The working reality in serious practices is unglamorous and has been for a while.

Machine output feeds a human reviser, who works against a translation memory and a client-specific termbase. Segments are triaged: some are accepted, some lightly repaired, some discarded and retranslated from the source. The industry calls this post-editing and distinguishes light from full — light for text that must be comprehensible, full for text that must be right. On some material the machine saves a great deal of time. On some material — dense literary prose, a badly scanned certificate, a clause of tortured legalese — post-editing takes longer than translating from scratch, because untangling a confident wrong sentence is harder than writing a right one.

Two consequences follow. First, the market is separating rather than shrinking: gist work is commoditising fast, while work whose failure is legal, financial or reputational is becoming more valuable, not less. Second, the profession's centre of gravity is moving from production to judgement — terminology, revision, cultural advice, deciding what may be automated and what may not. That is a narrower and more demanding job than the one it replaces.

Where this goes

Some predictions are safe. Document-level coherence will improve. Speech and live interpretation will get considerably better. Pipelines will consult termbases and prior work automatically rather than being fed them. The floor of acceptable quality will keep rising, and the volume of text that is never seen by a human will keep growing.

Two things are less often noticed.

The first is a data problem. As machine output floods the web, models are increasingly trained on translations rather than on originals. For well-resourced languages this is diluted by an ocean of human text. For dialects and smaller languages — where the corpus was thin to begin with — the feedback loop is a real risk, and the quality of human-made data becomes more valuable rather than less.

The second is that none of the improvement touches the accountability layer. Better models produce better drafts. They do not produce someone who can be sworn before a court. As long as institutions require a responsible party, the last step of consequential translation stays human, whatever proportion of the keystrokes preceding it does not.

The short version. Use the machine where being wrong is survivable, and pay for a person where it is not. For anything an institution will act upon — a court, a ministry, a university admissions office, a publisher — the relevant question was never whether the software is clever. It is whose name is at the bottom of the page.

This post is general and does not constitute legal advice on document requirements. Requirements vary by receiving institution and by jurisdiction, and should be confirmed with the body concerned.