AI translation quality review for output you already have
You already translate with AI or machine translation (MT), and you need to know whether the output holds up before more of it goes live. Our AI translation quality review is a language quality assessment (LQA) of that output. Quality assurance (QA) agents check the text, a native linguist validates every finding, and you get a report you can take to your stakeholders: each issue classified by severity and category, the patterns behind them, and a recommended service level for the content.
To see the method on your own text first, start a free check on up to 300 source words.
What we review
Send us the AI or MT output you want assessed, in a language pair we offer:
- The source text and the output, aligned, or a bilingual file
- The name of the tool or engine that produced the output
- The content type, such as help-center articles, product content, or website pages
- Your glossary or style guide, if you have one
Agents run automatic checks and separate QA checks for accuracy, grammar, style, and terminology. A linguist who is a native speaker of the target language and a subject specialist then validates the findings and assesses the sample. A finding counts only when it names a concrete defect and points to a reference: the source, your glossary, a rule, or your instructions. Preferences are logged separately from errors.
How findings are classified
Each issue gets one severity: critical, major, or minor. Each issue also gets one category:
| Category | What it covers |
|---|---|
| Accuracy | The target says something the source does not: a mistranslation, an omission, an addition, or a changed number |
| Fluency | Grammar, spelling, and punctuation |
| Terminology | A term that does not match your glossary, or a term used inconsistently |
| Style | Tone, register, unidiomatic phrasing that reads like a translation, and wording that does not follow your style guide |
| Locale | Conventions of the target market: units, dates, number formats, and regional vocabulary |
| Tags and format | Tags, placeholders, variables, line breaks, punctuation pairs, and length limits |
| Client requirements | Instructions you gave us for your program, such as how to handle interface labels or product names |
What the report contains
- The scope: language pair, content type, and what was assessed
- The sample method and size
- Issue counts by severity and category
- An issue log with each segment, its source and target, the category, the severity, an explanation, and a suggested fix
- The main patterns behind the issues
- A recommended service level
- The label “Assessment by Locstars.”
If the sample is too small for a reliable rating against an acceptance level, the report says so and states the sample size. Scores appear only in your own report, never on this site.
From findings to a decision you can act on
The patterns matter more than any single error, because they tell you what to fix in your program, not only in this file. The report recommends a level based on what the issues show and on the risk of the content. For example, content with scattered minor issues may be fine at AI + Check, if the content type allows it, while repeated meaning or terminology errors point to AI + Edit. Regulated and high-risk content points to AI + Edit + Review or the no-AI route.
Fixing what we find
Correction is quoted separately, at the level the report recommends. For most content that is AI + Edit: a native post-editor edits every segment, QA agents check the result, and every delivery comes with the quality-gate result, an issue log, glossary changes, and a query log, plus sampled quality reports on the schedule agreed for your program. Regulated and high-risk content is corrected at AI + Edit + Review.
| Row | Free check | Quality review |
|---|---|---|
| What you send | Up to 300 source words of existing AI or MT output | A larger sample of existing AI or MT output, agreed with you when we scope the review |
| What you get | An issue log, up to three patterns that matter most, and a recommended level. Findings on the text you sent. Indicative, not a rating of your full content. | A report with issue counts by severity and category, the sample method and size, an issue log, the main patterns, and a recommended level you can plan a program on |
| Next step | A sample on your own file | Correction, or a pilot at the recommended level |
Questions about AI translation quality review
How do you classify issues?
We classify every issue by severity (critical, major, or minor) and by category, from accuracy and terminology to tags and format. The quality page explains the classification in full.
Can you review output from any AI or MT tool?
Yes. The review works on the text, so the tool does not change the method. Tell us which tool produced it, because that helps us read the patterns.
How is this different from the free check?
The free check covers up to 300 source words and returns an issue log, up to three patterns that matter most, and the service level we would recommend. Its findings are indicative, not a rating of your full content. The quality review assesses a larger sample that we agree with you. It adds issue counts by severity and category, states the sample method and size, and gives a recommended level you can plan a program on.
Can you fix the errors you find?
Yes. Correction is quoted separately, at the level the report recommends: usually AI + Edit, with a native post-editor editing every segment, or AI + Edit + Review for regulated and high-risk content.
Who reviews the output?
A linguist who is a native speaker of the target language and a subject specialist validates every finding. QA agents run their checks first.