Machine translation plus human review.

Two camps argue past each other about machine translation. One says the machines are good enough now, just ship the output. The other says translation is a human craft and machines produce fluent nonsense. Both camps are wrong in the same way: they answer a per-document question with a policy for all documents. The practitioners who actually move institutional volume settled this argument years ago, and their answer is boring and correct: machine translation plus human review, with the depth of review matched to the stakes of the document. This post lays out that position properly.

Why both extremes fail

"MT is good enough for everything" fails on the errors that matter. Modern engines are remarkably fluent, and fluency is exactly what makes their failures dangerous: a dropped negation, a wrong number, a mistranslated legal term all read perfectly. In casual contexts these errors cost nothing. In a contract, a regulation, or a published annual report, one of them is a headline. No serious practitioner ships raw machine output under an institution's name.

"Only humans can translate" fails on arithmetic. Institutions produce far more bilingual content than human translation can absorb at human speed and human cost, and the practical result of an all-human policy is not better translation. It is less translation: documents that never get translated at all, or get pasted into a free web tool by someone under deadline, which is the worst of both worlds. Meanwhile, a human reviewing good machine output works several times faster than a human translating from a blank page, because judging a sentence is quicker than composing one.

A triage framework: four levels of stakes

The workable policy assigns review depth by consequence. Four levels cover most institutional documents.

• Gist. You need to know what a document says, not to reuse its words: an inbound foreign-language letter, a competitor's press release, triage on a large pile. Raw machine output is fine, provided it is labeled as machine output and nobody forwards it as finished work.
• Internal. Working documents your own colleagues read: memos, meeting notes, draft summaries. Light review by one person, fixing errors of meaning (wrong numbers, dropped negations, mistranslations) and leaving awkward-but-accurate phrasing alone. Correct, not beautiful.
• Published. Anything that leaves the building under your name: reports, web content, press releases. Full post-editing to publication standard, meaning terminology, style, and tone all reviewed, every segment carrying a human sign-off before release.
• Legally binding. Contracts, regulations, court submissions. Full review by someone with legal competence in both languages, and in many jurisdictions a certified or sworn translation on top, which machine translation does not replace and is not trying to.

What makes review fast instead of painful

The triage framework only works if review is genuinely fast, and that is a tooling question. Reviewing a translated document by reading two files side by side in two windows is miserable; nobody sustains it. What makes review quick is structure:

• Segments and statuses. The document arrives as rows, source beside machine translation, and every segment carries a state: machine, edited, accepted, rejected, locked. Progress is a count, not a feeling, and "done" has a definition: every segment human-approved. Bulk accept clears the clean stretches in seconds.
• Automated QA. Mechanical errors deserve mechanical detection: number mismatches between source and target, untranslated segments, doubled spaces, punctuation problems, and length ratios that suggest dropped content, with Arabic-specific checks on top. QA will never tell you a sentence sounds stiff; it will catch the wrong number a tired human eye slides past.
• Translation memory. Segments a reviewer approved once come back pre-filled and locked the next time they appear, so the second annual report starts from the decisions made in the first. Review effort compounds instead of resetting.
• Glossaries. The official rendering of every ministry, title, and product name is pinned once and enforced everywhere, so reviewers stop re-litigating terminology sentence by sentence.

This is the loop TranslateX ships: a side-by-side review editor with segment statuses, track-changes diffs against the original MT, per-segment glossary and memory lookups, automated QA checks, and a final document regenerated with the original layout intact once review closes. The mechanics are covered step by step in our academy lesson on the post-editing workflow.

Where humans stay mandatory

Honesty about the boundary keeps the whole framework credible. Certified and sworn translation for courts, immigration, and official filings is a legal institution, not a quality tier: the translator's accreditation and liability are the product, and no machine output substitutes for it. High-stakes legal drafting needs bilingual legal judgment, because the reviewer is checking legal effect, not just language. Literary and high-visibility persuasive writing remains human craft. And any document where your organization would want a named person accountable for the words should have one.

None of that weakens the argument; it completes it. Machine translation plus human review is not a compromise between two extremes. It is the discipline of spending scarce human judgment where consequences live, and letting machines do the typing everywhere else. The teams that get this right translate more, publish faster, and make fewer embarrassing errors than either camp, which is presumably the point.