Machine Translation and Post-Editing: Where It Works, Where It Fails

Machine translation stopped being a joke some years ago. Neural systems produce fluent, often accurate text for general content in well-supported language pairs, and used properly they save real money. They also fail in a specific and dangerous way: the output reads confidently whether or not it is right.

That combination — fluent and unverified — is what makes the buying decision matter.

The service levels, named honestly

Raw machine translation. No human involvement. Appropriate for understanding an incoming document, triaging a mailbox, or searching a large archive. Never appropriate as a deliverable to a client, an authority or the public.

Light post-editing. A translator fixes what makes the text wrong or unusable: mistranslations, omissions, broken terminology, anything misleading. Style is left alone. Result: accurate, serviceable, visibly utilitarian text — fine for internal documents, support content and high-volume material nobody will quote.

Full post-editing. The translator brings the text to the standard of human translation: terminology, register, flow, consistency. Result: publishable. Cost: closer to translation than to editing, especially in specialised fields.

Human translation with revision. Translation by a specialist and review by a second linguist, as described in translation quality assurance. The standard for anything legally binding, published or persuasive.

The mistake is not using machine translation. The mistake is buying one level and expecting another.

Read also: Website and App Localization: A Practical Guide to Launching in Arabic

Where it works well

• Internal and operational content — status reports, internal e-mails, meeting notes, knowledge-base articles for staff.
• High-volume, low-stakes material — product listings, user reviews, support tickets, catalogue data.
• Repetitive technical content with a mature glossary and translation memory — release notes, repeated manual sections. See technical translation.
• Triage — deciding which of 500 documents need a professional translation.
• Draft stage — a first pass for a human to work on, when the pairing is mature and the source is clean.

Read also: Expanding Into Arabic Markets: A Language and Document Checklist

Where it fails, predictably

• Legal and contractual text. Obligations hinge on modal verbs and defined terms; a confident paraphrase can change liability. See legal and contract translation.
• Medical and pharmaceutical content. Dosages, units and abbreviations need a specialist who queries rather than guesses. See medical translation.
• Certified documents. The deliverable is a certified artefact, and the certifier takes responsibility for accuracy — see running a certified translation office.
• Marketing and brand voice. Persuasion does not survive literal rendering; this is copywriting, not translation.
• Anything with missing context. Interface strings, table headers, isolated labels — a machine cannot know what a two-word button does.
• Safety instructions. The cost of a plausible error is unacceptable.

The Arabic-specific risks

Machine translation into and out of Arabic has improved substantially, and still carries characteristic failure modes:

• Dialects. Systems are trained predominantly on Modern Standard Arabic; Egyptian, Gulf, Levantine and Maghrebi content is handled unevenly.
• Absent diacritics. Ambiguity that a human resolves from context is resolved by the machine by guessing.
• Names. Personal and place names get translated, transliterated inconsistently, or mistaken for common nouns — and in official documents names must match the passport exactly. See Arabic to English translation.
• Register. Output tends toward a neutral tone that is wrong for legal, religious and formal administrative texts.
• Mixed-direction content. Numbers, codes and Latin-script terms inside Arabic text are frequently mangled.
• Terminology drift. The same term rendered three ways across a long document — the exact failure a glossary exists to prevent.

Confidentiality: read the terms before you paste

Putting a client's contract or a patient's report into a consumer tool can mean transferring personal or privileged data to a third party, outside the control of the person who owes a duty of confidence. Before any machine translation touches sensitive material, establish:

• whether the content is used to train models or retained;
• where it is processed and for how long it is stored;
• who inside the provider can see it;
• what the provider has committed to in writing.

For legal, medical, HR and financial documents, the safe answer is a provider with explicit data-handling terms — not whichever tool is open in a browser tab.

How to buy it

Say what you want, in these words:

• the service level — raw, light post-editing, full post-editing, or human translation with revision;
• the purpose — understanding, internal use, publication, submission to an authority;
• the engine and glossary, if you have preferences or an existing memory;
• the confidentiality terms;
• and who is accountable for accuracy on delivery.

Then price it on that basis. Post-editing discounts are legitimate where the machine genuinely reduces effort, and dishonest where it does not: asking a specialist to post-edit a technical or legal document at half rate usually produces either a loss for the translator or a worse document for you. Our guide on per page versus per word covers how to compare rates when the unit changes, and how much translation costs covers the market picture.

A simple decision rule

Ask one question: what happens if a sentence in this document is confidently wrong?

• Nothing much — raw or lightly post-edited machine translation is fine.
• Someone is misinformed but can check — light to full post-editing.
• Money, health, rights, compliance or your reputation are at stake — human translation with independent revision, by a specialist.

Most organisations need all three, for different documents. The discipline is deciding deliberately, per document type, and writing it down — rather than deciding by accident, in a hurry, on the document that mattered most.