What is AI translation QA, and what does it catch that rule-based QA does not?

AI translation QA is a check where a language model reads every translated segment next to its source and asks one question: does the target say the same thing? It catches what rule-based QA cannot see, such as a reversed meaning, a missing sentence or the wrong register, in a segment whose tags, numbers and terms are all correct.

That is the short answer. The rest of this page explains how the two layers work together, shows two real examples and says where the AI stops.

Why rules are not enough any more#

Rule-based QA tools compare form. Is the tag still there? Does the number match? Is the glossary term used? Those questions have a yes or no answer, and a rule answers them the same way every time. That is exactly what you want for tags, numbers, placeholders, URLs and terminology.

What a rule cannot do is read. For years that was fine, because the typical error was mechanical: a tired translator dropped a tag or typed 205 for 250. Much of what we review today was drafted by a machine, and machine output fails differently. The tags are right, the numbers are right, the sentence reads well. And it says something else.

How What's Wrong does it#

1. The rules run first. Tags, numbers, placeholders, URLs and email addresses, punctuation and spacing, untranslated segments, glossary and do-not-translate terms, locale formats, your own rules and any checklist you imported. None of this needs a model. Spelling runs beside the rules in every language: a dictionary proposes the candidates and the AI confirms each one before it is reported, and for a language no dictionary covers the AI proofreads the segments itself. Spelling findings are listed, never scored.
2. A model reads the rest. It reads every segment the rules passed, plus the ones they only warned about, one segment at a time with its neighbours as context. It looks for mistranslations, omissions, additions, terminology and register.
3. A second opinion on the rule findings that need a reading. A model also reads each critical or major rule finding that calls for a reading, such as a term, a number, a length or a checklist hit, and says whether it agrees, disputes it or cannot tell. The deterministic checks never go to it: a missing tag or placeholder stands as found, same file, same finding. A disputed finding is never hidden. It stays in the report with the reason and counts at half weight in the score until the reviewer decides; the rules the second opinion most often disputes rightly, such as a segment that looks untranslated or a checklist hit, are shown as minor while the dispute stands. The report's Confirmed view holds what the AI confirmed and what an exact check found. That is how the false positives that fill a rule-based report get cleared first.
4. A person decides. Every finding comes with its reason and a suggested fix. The reviewer accepts it, ignores it or fixes the segment in their own CAT tool, and signs the report off. A reviewer can also teach: one line written for the AI on a finding (this is the client's house phrasing, this term stays in English) becomes, once a PM approves it, guidance the AI review and the second opinion read on the next run. That is Learn, on the Business plan.

A worked example#

This one is from our public sample report: an example file, real output.

A customer reading the Turkish page goes looking for a discount code that does not exist. No rule would flag this segment, because there is nothing wrong with its form.

A second example: the false alarm everyone knows#

The second opinion works the other way too. A settings menu has a line that says Wi-Fi, and the Turkish target says Wi-Fi. The rule sees a target identical to its source and raises it as untranslated, at Major.

Anyone who has run Xbench on a software file knows this one. It is not rare either: in the reports run on What's Wrong up to mid-October 2026, the second opinion disputed 695 of the 891 findings of this rule, 78%. The rule was right to look, because an untranslated segment is a real error when it happens. The model reads the context and says why this one is fine. The reviewer decides, and on the Business plan one click on the finding sends Wi-Fi to the profile as a Fine as is suggestion, so once a PM approves it the next report does not ask again.

What it does not do#

• It does not replace a reviewer. The model proposes. A person confirms or overrides, and no result is final without one.
• It is not always right. That is why every finding carries its reason and why a disputed rule finding is shown rather than hidden.
• It does not search a whole project for you. Looking up how a term was translated across forty files is still a job for a desktop tool such as Xbench.
• It does not work offline. The AI review runs on servers. If your contract forbids any online tool, read what leaves your workspace and what never does first, or switch the AI features off and keep the rule checks.

Where to go next#

• The full sample report and how the check works.
• Security and data handling: where the file goes, how long it stays and which providers see the text.
• Check a translation from an AI assistant through the MCP server.

There is a free plan, no card needed. See the plans.