AI translation can produce a useful first draft, but a fluent-looking result is not automatically accurate, complete, or safe to publish. This multilingual quality assurance checklist gives editors a repeatable way to review meaning, terminology, tone, formatting, cultural references, and sensitive claims before releasing AI-generated translations.
Overview
Reviewing an AI translation is more than checking whether every sentence sounds natural. A translation can be grammatically polished while changing a qualification, omitting a warning, mistranslating a product term, or using a tone that does not fit the target audience. The goal of quality assurance is to compare the translation with the source and confirm that it communicates the intended message in an appropriate way.
Use this process for website translation, marketing copy, support content, internal documents, and other multilingual publishing workflows. The level of review should match the risk of the content. A casual social post may need a focused language check, while a legal, medical, financial, safety-related, or public-facing document requires careful human review by someone competent in both languages and familiar with the subject.
Before reviewing, preserve the source text, the AI output, the model or tool used, the date, and any instructions supplied in the prompt. This record makes recurring errors easier to identify when you change tools, add a language, or update a content batch. For broader terminology and consistency work, use a translation quality assurance checklist alongside this AI-specific review.
What to track
Track the same quality variables for every batch rather than relying on a general impression. A simple spreadsheet can include the source segment, translated segment, issue category, severity, correction, reviewer, and approval status.
1. Meaning and completeness
Read the source and translation side by side. Confirm that the translation preserves the main claim, conditions, permissions, prohibitions, and intended audience. Look specifically for omitted sentences, repeated phrases, invented details, reversed meanings, and changes in certainty. Words such as “may,” “must,” “usually,” and “not” deserve close attention because a small change can alter the practical instruction.
Check headings, captions, buttons, footnotes, alt text, metadata, and text inside graphics as well as the main paragraphs. AI workflows sometimes process these elements separately, which increases the risk of inconsistent or missing content.
2. Names, numbers, and technical details
Verify names of people, places, products, organizations, file formats, units, dates, times, percentages, prices, URLs, email addresses, and reference codes against the source. Do not assume that a plausible-looking number is correct. Check decimal and thousands separators, date order, currency notation, measurement conversions, and plural forms where relevant.
Names should be handled according to the project’s agreed convention. Some names remain unchanged; others have an established local form. Record the decision in a glossary or style guide so later batches do not introduce variation. The translation glossary and style guide template can help organize these decisions.
3. Terminology and consistency
Compare key terms across the entire batch, not just within one sentence. Product features, interface labels, legal concepts, industry terms, and recurring calls to action should use approved equivalents. Watch for false friends, overly literal translations, unexplained abbreviations, and terms that shift between formal and informal forms.
Build a short term list before review begins. Include the source term, approved translation, terms to avoid, grammatical information, and a context note. When the AI repeatedly chooses the wrong equivalent, correct the term in the glossary and update the prompt or translation workflow rather than fixing each occurrence without documenting it.
4. Grammar, naturalness, and readability
After checking meaning, read the translation as a target-language reader would. Look for agreement errors, incorrect word order, unnatural collocations, excessive repetition, awkward pronouns, and sentence structures copied too closely from the source. A multilingual readability check can help you assess clarity and audience fit, but it should support—not replace—human judgment.
Check whether the level of formality matches the channel. A customer support article, product interface, executive announcement, and social caption may require different degrees of directness. Natural language is not necessarily informal language; the appropriate choice depends on the audience and purpose.
5. Tone, culture, and sensitive claims
Review humor, idioms, metaphors, references, honorifics, gendered language, and culturally specific examples. A literal rendering may be understandable but inappropriate, confusing, or unintentionally offensive. Decide whether to translate, adapt, explain, or remove a reference based on the target audience and the content brief.
Flag claims involving health, safety, law, finance, discrimination, identity, or public instructions for additional human review. Check that the translation has not strengthened a claim, removed a limitation, or introduced wording that could be interpreted as a guarantee. AI output should not be treated as an independent authority for sensitive content.
6. Formatting and user experience
Inspect lists, tables, headings, links, placeholders, variables, markup, line breaks, capitalization, and punctuation. Confirm that translated text fits buttons, navigation labels, forms, and mobile layouts. Languages expand and contract differently, so a sentence that is correct in a document may still be unusable in an interface.
For websites, review the translated page in its actual layout and test links, forms, search fields, structured content, and language selectors. The website translation and localization checklist provides a useful companion for launch and maintenance checks.
Severity levels for corrections
Use a consistent severity scale. A critical error changes a safety instruction, legal meaning, essential number, or required action and must be fixed before publication. A major error affects meaning, terminology, audience suitability, or important formatting. A minor error affects polish, consistency, or style without changing the message. This classification helps editors prioritize when a batch contains more issues than they can correct immediately.
Cadence and checkpoints
Run quality checks at several points instead of waiting until the end. First, perform a sample review of the AI output before translating the full batch. Use representative content: short and long sentences, headings, lists, terminology-heavy passages, and any sensitive sections. This early check can reveal that the prompt, glossary, source formatting, or selected tool needs adjustment.
Next, review the first completed portion in detail. Confirm that the workflow preserves placeholders, tags, links, and formatting. If the first sample contains repeated errors, pause the batch, revise the instructions or glossary, and reprocess where appropriate. Continuing without correcting a systematic issue usually increases the cost of review.
At completion, carry out a full targeted pass. Recheck critical content, all numbers and names, approved terminology, and any segments previously flagged by the reviewer. Then conduct a separate read-through for naturalness and layout. Keeping accuracy and style as distinct passes makes it less likely that fluent wording will hide a meaning error.
For recurring work, maintain a monthly or quarterly quality log. Track the language pair, content type, tool or model version if available, number of segments reviewed, critical and major issues, recurring error categories, and unresolved questions. The purpose is not to create a misleading single score; it is to see patterns and decide where process changes are needed.
How to interpret changes
A lower issue count does not always mean that translation quality has improved. A batch with simpler content may naturally produce fewer errors, while a new product area or language pair may expose issues that were not previously visible. Interpret results alongside content difficulty, reviewer time, sample size, and the types of errors found.
Separate systematic problems from isolated slips. If the same term is wrong throughout a batch, improve the glossary, source wording, prompt, or terminology constraints. If errors occur mainly in long sentences, revise the source for clarity or split complex segments before translation. If the output is accurate but consistently too formal, add audience and tone examples to the workflow.
Compare human corrections over time, but do not use an automated similarity score as the sole publication decision. Surface-level closeness can reward literal wording, while a more substantial edit may be necessary for natural and culturally suitable communication. Human review remains especially important when a translation affects rights, safety, money, reputation, or a user’s required action.
When changing tools, compare a controlled sample using the same source text, glossary, instructions, and review criteria. Record what improved and what regressed. A tool that performs well for product descriptions may not be equally suitable for legal notices, creative copy, or highly specialized documents.
When to revisit
Revisit this checklist whenever you add a target language, change an AI translation tool, revise a glossary, launch a new content type, or receive complaints about a published translation. Schedule a monthly or quarterly review for active multilingual programs, even if no tool has changed. Recurring review helps reveal terminology drift, new formatting problems, and error patterns that are easy to miss in individual batches.
Use the following practical routine:
- Keep a current glossary, style guide, and record of approved corrections.
- Sample new AI translations before approving a full batch.
- Review meaning, omissions, terminology, numbers, names, tone, cultural references, and formatting separately.
- Assign a severity level to each issue and block publication for critical errors.
- Record recurring causes, not only individual corrections.
- Recheck a representative sample after changing the tool, prompt, model, source format, or target market.
- Escalate sensitive or high-impact content to a qualified human reviewer.
For ongoing terminology management, see the translation glossary template. If the work is intended for search, pair linguistic review with the multilingual keyword research workflow so accurate language also reflects how the target audience searches. A repeatable review record turns AI translation from an unchecked shortcut into a controlled editorial process.