Smart Revision — How to Use AI to Review Your Translation, Not Replace You

There’s a trap many translators fall into without realizing it until the damage is done. They finish a translation, ask an AI model to “review and improve it” — and get back a completely different text. Eloquent, smooth, error-free, and nothing like their voice or the source author’s style. Their translation wasn’t reviewed. It was rewritten.

The problem isn’t the model. It’s the instruction. “Review and improve” gives the model open permission to rewrite. What we need is a fundamentally different instruction: we want a critic who identifies problems, names them precisely, and suggests alternatives — not a ghostwriter who replaces the translator. This fourth article in the Translation Prompt Library series builds that precise critic.

The Difference Between Review and Rewrite — and Where They Blur

Professional revision operates on one principle: preserve what works, fix only what’s broken. But when an AI reads a text with “improvement” as its objective, its underlying training pushes it toward standardization, smoothing, and completion — exactly the operations that erase style and produce homogeneous output.

The difference between what we want and what the model does by default is like the difference between a doctor who diagnoses and a pharmacist who dispenses. You want the diagnosis: what’s the problem, where is it, and why. The wrong prompt gives you the prescription immediately, without telling you what your condition was — and the prescription is often for someone else’s illness.

This doesn’t mean AI is always right in its diagnosis either — it isn’t. But with the right framework, the diagnosis becomes something you can discuss, challenge, and reject, instead of a finished text presented as a fait accompli.

If you want to understand the common errors that damage translation prompts before you even reach the review stage, we’ve covered them in detail here:
(See our article: 10 Mistakes That Kill Your Translation Prompt Results | With a Ready Fix for Each)

Good AI review doesn’t give you a better text — it gives you a precise list of problems that you can resolve using your own professional judgment.

What Types of Review Do You Actually Need?

Not all revision is the same type of task. Before writing a review prompt, identify exactly what you need:

Accuracy review: Does the translation fully reflect the source meaning? Is there any unintentional omission, addition, or distortion?

Style review: Is the Arabic linguistically clean and contextually appropriate? Are there awkward sentences or imprecise terminology?

Consistency review: Are terms unified throughout the text? Is the register stable from start to finish?

Fluency review: Does the text read naturally in Arabic, or does it feel translated? Have foreign syntactic structures crept in?

Each type calls for a different prompt — or at minimum, different instructions within the same prompt. The common mistake is asking for “everything” at once, because a model asked to review accuracy, style, and consistency simultaneously tends toward rewriting rather than diagnosis.

The Ready-to-Copy Prompt: Scoped Critical Review

The core principle of this prompt is explicitly limiting the model’s authority: it diagnoses, points, and suggests — but it does not rewrite unless you ask it to.

You are a professional translation reviewer. Your role is diagnosis, not rewriting.

I will give you the source text and the Arabic translation together.
Review the translation only on these dimensions that I specify:
[Choose what you need: accuracy / style / consistency / fluency]

Review rules — follow without exception:
- Do not rewrite the translation in full under any circumstances
- For each problem you find: identify its location precisely (the sentence or phrase),
  name it (accuracy error / awkward phrasing / imprecise term / calque / register break...),
  then suggest one or two alternatives only, with a brief justification
- If a passage is good, say so explicitly — do not suggest improvements
  to what doesn't need improving
- Order your notes from most critical (affects meaning) to least critical
  (stylistic preference)
- At the end: give me a one-line assessment — what is the strongest aspect
  of this translation, and what needs the most attention?

Source text:
[Insert source text here]

Arabic translation:
[Insert your translation here]

The final one-line assessment isn’t a formality. It forces the model to prioritize and synthesize rather than producing an endless list of minor observations that overwhelm more than they help.

How to Choose the Right Review Dimension

Choosing your review focus depends on the text type and your stage of work:

First draft, straight-through translation: Start with accuracy — confirm the meaning is complete before worrying about style. Style-polishing an incomplete translation is a waste of effort.

Accuracy confirmed, raising the bar: Use the fluency dimension — it surfaces the foreign syntactic structures we become blind to when we’re deep inside a text.

Long or complex text (contract, report, article): Use the consistency dimension — terminology unification in long texts is the most commonly overlooked issue and the one most likely to embarrass you with a professional client.

Literary or marketing text: Add the style dimension after accuracy — and always include the author’s voice description we built in the second article of this series.
(See our article: Voice and Style — How to Preserve the Author’s Tone in Translation)

Full Working Example: Reviewing a Legal Translation

Here’s the prompt in action. A clause from a consulting agreement (simplified for illustration):

Source text:
“The consultant shall not disclose, directly or indirectly, any confidential information obtained during the term of this agreement to any third party without prior written consent of the client.”

Translation submitted for review:
“لا يجوز للمستشار أن يُفصح عن أي معلومات سرية حصل عليها خلال فترة هذه الاتفاقية، بشكل مباشر أو غير مباشر، لأي طرف ثالث دون الحصول على إذن كتابي مسبق من الطرف الأول.”

Dimensions requested: Accuracy and terminological consistency.

What the prompt returned:

Note 1 — Accuracy (critical): The phrase “الطرف الأول” (first party) does not appear in the source text, which uses “the client.” If the contract defines parties as “First Party” and “Second Party” in its preamble, this is legally appropriate. If it doesn’t, introducing this term modifies the contract rather than translating it. Suggestion: use “العميل” (the client), or verify the contract’s definitions section first.

Note 2 — Consistency (medium): “Confidential information” is translated as “معلومات سرية.” This is correct, but verify that this exact term appears wherever “confidential information” appears throughout the full contract — not “المعلومات الخاصة” or “المعلومات المحظورة” in other clauses.

Note 3 — Style (low priority): The placement of “بشكل مباشر أو غير مباشر” after the verb and before the object is grammatically fine but slightly heavy. An alternative: “لا يجوز للمستشار أن يُفصح، مباشرةً أو بطريقة غير مباشرة، عن أي معلومات سرية…” — not an error, purely a stylistic option.

One-line assessment: The translation is accurate in substance with clean legal register. The point that deserves the most attention is verifying the “الطرف الأول” term against the full contract’s definitions.

Notice what happened: three notes, ordered by importance, with a clear distinction between what is an error and what is a stylistic preference — and a final verdict that gives you a clear priority. The translation was not rewritten. We can now respond to each note with our own professional judgment: accept, reject, or investigate.

Good revision doesn’t diminish the translator’s role — it magnifies it. It ensures that the final professional call belongs to the translator, not to the model that happened to run a pass over the text.

Advanced Tip: The Fluency-Specific Review Prompt

Fluency review — detecting text that “smells like a translation” — needs a different approach because it’s measuring a feeling rather than a rule. This specialized prompt works by shifting the model’s position entirely:

You are an Arabic-speaking reader who doesn't know this text is a translation.
Read the following text and identify:

1. Any sentence or phrase that made you pause because it didn't feel
   natural in Arabic — not grammatically wrong, just "translated-sounding."
2. For each: what specifically makes it feel that way?
   (foreign syntactic calque / word used in an unfamiliar sense /
   sentence that opens in an unusual way for Arabic / other)
3. How would a native Arabic writer express the same idea?

Do not suggest revisions for sentences that feel natural — ignore them completely.

[Insert text here]

This prompt is particularly effective because reframing the model as “an Arabic reader who doesn’t know it’s a translation” changes what it notices. It’s no longer scanning for rule violations — it’s reading for naturalness, which is the actual standard a fluency review needs to meet.

For a deeper look at how this approach applies specifically to machine translation post-editing — which has its own distinct workflow — we’ve covered it separately:
(See our article: The 8 Best Post-Editing Prompts for Machine Translation | A Practical Guide for Freelance Translators)

Special Case: When You’re Reviewing Machine Output, Not Your Own Work

Everything above applies to reviewing a translation you produced yourself. When the task is reviewing machine translation output (from DeepL, Google Translate, or a similar system) before delivery to a client, the approach shifts slightly.

Here the model is not reviewing its own work — which removes one layer of bias — but it still needs stricter instructions about what to preserve versus what to change. In most cases, using a different model from the one that produced the original translation reduces the tendency to validate existing choices rather than scrutinize them. And the standard shifts: post-editing machine output has different benchmarks than reviewing human translation, because the starting point is different.

For a full toolkit including the software translators use alongside AI review, this article covers the professional landscape:
(See our article: The Professional Translator’s Toolkit: Dictionaries and Software That Make the Difference)

And if you’re thinking about where AI fits in translation more broadly — as a tool, a competitor, or a partner — this article addresses the question directly:
(See our article: AI in Translation — Partner or Competitor)

Practical Takeaway Before the Next Article

AI is an excellent critic when you limit its authority clearly. Its problem isn’t its inability to spot issues — it’s its instinct toward full repair when given open-ended space. A good review prompt doesn’t ask “is this good?” — it asks “what exactly is the problem, where, and why?”

Three things to apply starting now:

1. On your next revision job, define the review dimension before writing the prompt — never ask for “everything” in one pass.
2. Use the “Arabic reader who doesn’t know it’s a translation” framing specifically for fluency issues — it catches what technical scanning misses.
3. Make the one-line final assessment a non-negotiable requirement in every review prompt — it keeps the model’s output focused and gives you a clear work priority.

In the fifth article, we tackle the most technically demanding challenge in a translator’s career: specialized language — how to handle legal, medical, and technical terminology efficiently without needing to be an expert in every field you translate.
(See our article: Technical Language — Handling Legal, Medical, and Technical Terminology with AI)