How to Review Machine Translation and Turn It Into Professional Work
A field guide to the most common MT errors into Arabic and how to catch them before your client does.
Review isn’t correcting errors — it’s making decisions. The difference between a strong reviewer and a weak one doesn’t show in the mistakes they catch, but in the judgments they make when there’s no obvious mistake at all.
Why Reviewing Is a Skill in Its Own Right
In our previous article on AI in translation, we saw that the new working model makes the human translator a smart reviewer of machine output rather than a competitor to it. But this immediately opens a practical question: how exactly do you review?
Surface-level review — reading the text and feeling it “seems fine” — isn’t professional review. A translator who takes a machine output and sends it after minor touch-ups is delivering a service below what clients expect, and exposing their reputation to real risk.
Professional review of AI output is an independent skill combining three things: a clear methodology, mature linguistic judgment, and conscious cultural decision-making. This article teaches you how to build all three.
First: Before You Start Reviewing — Setting Your Mind Right
The biggest mistake people make when reviewing machine translation is reading the translation first. This locks the mind into the choices the machine already made and turns the reviewer into a copyeditor rather than a translator who judges.
The correct order is:
Step one: read the source text in full before opening the translation. Understand its subject, purpose, audience, and tone. Hold these questions in mind: what is this text trying to achieve? Which terms are sensitive? What level of formality does it require?
Step two: read the machine translation in full, once, without editing. The goal of this reading is general diagnosis: what is the overall quality level? Where did it succeed and where did it fail? Are the problems systematic (a repeated wrong term) or scattered (random slips)?
Step three: begin the systematic review with a mind that knows what to expect.
Practical rule: if your first read of the translation shows that more than 30–40% of the sentences need substantial revision, it’s faster to translate from the source directly. Fully reviewing a weak translation takes longer than translating from scratch.
Second: The Layers of Review — From Easier to Deeper
Professional review works in successive layers, each building on the one before.
Layer One: Informational Accuracy
The question: was the original meaning transferred with precision?
This is the minimum threshold of any review. It includes: comparing each sentence to its source to verify nothing was added or omitted. Watching for negation markers — the machine sometimes drops them or inverts them. Checking all numbers, dates, proper names, and specific terms — these transfer correctly or they don’t. Verifying pronouns and references — in long texts the machine sometimes loses track of a pronoun’s antecedent.
Layer Two: Terminological Consistency
The question: was the same term translated the same way every time it appeared?
The machine often translates the same term differently in two different passages — especially in longer texts. This confuses the reader and strips the text of professionalism. Build a working terminology list during your first read and check for consistency throughout.
Layer Three: Linguistic Naturalness
The question: does the text read as though it was written in the target language, or does it feel translated?
This is where the difference between an average and a professional reviewer becomes visible. A sentence may be grammatically correct, faithful in meaning, and still feel artificial. The most common causes in machine translation output:
Source-language sentence structure carried over literally. Nominalized constructions where verbs would read better. Long embedded clauses that work in English but weigh heavily in Arabic or vice versa. Prepositions that are grammatically acceptable but idiomatically wrong.
Layer Four: Cultural and Contextual Appropriateness
The question: is the text appropriate for its target audience — in their culture, register, and expectations?
This is the deepest and hardest layer to teach — it requires real field experience with both the subject and the cultures of both texts. Some examples: the level of formality appropriate for a Gulf corporate audience differs from what suits a North African one. Humor and wordplay in the source — how should they be handled in the target context? Religious expressions in marketing copy — the machine sometimes misses their sensitivity entirely. Literary and cultural allusions that resonate in the source and land flat in the target.
Third: The Most Common MT Errors Into Arabic — A Field Guide
Based on real work with Claude, ChatGPT, and specialized translation tools, these patterns recur consistently and deserve particular attention in every review:
Gender agreement: the machine makes errors matching verbs and adjectives to grammatically feminine nouns — especially non-biologically feminine ones. “The company decided” becoming masculine in Arabic is a common slip.
Broken plural forms: choosing the correct plural from among several valid Arabic options — the language has complex plural patterns — requires linguistic judgment, not statistical selection.
Proper names and international terms: the machine hesitates between Arabization and transliteration, sometimes mixing both within the same text. Decide your standard at the start and apply it consistently.
Prepositions: Arabic prepositions don’t map onto English ones directly, and subtle errors here produce text that is technically acceptable but idiomatically wrong.
The term that looks right but isn’t: this is the most dangerous category. A word that is semantically close but contextually wrong, or a term that is standard in one market and non-standard in another. These quiet errors pass unnoticed in surface-level review and emerge when the client notices them.
Fourth: Tools That Speed Up Review Without Reducing Quality
These tools don’t replace the human reviewer — they free them for the tasks that actually need human judgment.
Specialized terminology databases: resources like the EU’s IATE (which includes Arabic) or domain-specific glossaries from professional bodies. Faster and more authoritative than personal judgment each time.
CAT tools (Computer-Assisted Translation): programs like OmegaT (free) or SDL Trados store your terminology decisions and suggest them automatically in future projects — solving the consistency problem in long-term work.
Claude as a secondary checker: after your human review, you can ask Claude for a limited grammar pass: “Review the following text for grammatical and spelling errors only — do not change the style or content.” This catches what you missed. But your human decision remains the final word. The prompt template for this is in our article How to Write a Prompt That Gets You What You Want.
Fifth: For Professionals — Building a Repeatable Work System
A Field-Specific Checklist
Each specialization has a different checklist. In legal translation: is every modal verb (shall/must/may) translated with precise terminological accuracy? Are the parties’ names consistent throughout the contract? In medical translation: are all dosages and units correct? Are drug names verified against authoritative sources?
Build your own checklist gradually from errors you’ve encountered in past projects. This list is your intellectual capital — it grows with experience and is yours alone.
Your Personal Glossary
A simple file in Notion, Excel, or even a word processor document that stores your fixed terminology decisions: “We always translate ‘Consideration’ in English law contracts as العوض not المقابل.” This saves time, ensures consistency, and accelerates your work on similar projects.
An Acceptance and Rejection Criterion
Before reviewing any project, decide in advance: what would make me translate from scratch rather than post-edit? A clear criterion protects you from wasting hours on a translation that cannot be salvaged.
Sixth: How to Price Machine Translation Review
There is no universal fixed formula, but these principles help:
Light PE (Light Post-Editing): grammar and spelling review with minor adjustments. Typically priced at 40–60% of the full translation rate.
Full PE (Full Post-Editing): substantial rewriting while keeping the machine output as a structural base. Typically priced at 70–85% of the full rate — and often takes nearly the same time anyway.
When to charge for full translation: if you find the machine output requires substantial rewriting of more than 50% of the content, inform the client and price accordingly. Honesty here builds a sustainable professional relationship.
We expand on pricing strategies in the freelance context more broadly in our article AI for Freelancers — 10 Tasks in Half the Time.
The Takeaway: Review as an Investment in Reputation
Ultimately, professional review of AI output isn’t only a technical task — it’s an investment in your professional reputation.
The client who receives an unreviewed machine translation and discovers its errors won’t return. The client who receives a professionally reviewed translation won’t ask how it was produced — they’ll ask when they can send the next project.
AI produces a draft quickly. Good human review turns that draft into a product that carries your name. The difference between the two is exactly what a professional rate is worth.
