Skip to main content

AI Translation - Output and Quality

How the context you set up for wxrks's AI (Translation Memory, glossary, MT and Sous-Chef profile) shapes translation quality, why the wording can still vary between runs, and how to keep the output consistent and reviewed.

When you use AI translation in wxrks, for example Context Sensitive Translate, the AI works with the context your team has set up and turned on for the project: your Translation Memory (TM), glossary, machine translation (MT) and Sous-Chef profile. How much of that context is available, and how good it is, drives the quality of the translation. This article explains how Context Sensitive uses it, why the wording can still vary from one run to the next, and what to do when you need consistent, reliable output.
​

💡 Who is this for? This guide is for Account Admins and Project Managers who need to understand what drives AI translation quality and variation, and plan their projects around it, and for linguists (Vendors) who want to understand why a suggestion changed and what to check before confirming it within the wxrks platform.


How Context Sensitive works

Context Sensitive doesn't use AI to translate from scratch. It uses AI to choose between linguistic possibilities. Working as an editing step on top of machine translation (known as automatic post-editing), it takes the MT draft, your TM matches, glossary terms and Sous-Chef profile, weighs them against each other, and turns them into the one translation that fits the segment best.

Your context guides the AI, it doesn't dictate every word

wxrks tells the AI how to rank your context: glossary terms first, then Translation Memory, then your Sous-Chef instructions, then machine translation. Each piece pins down part of the translation and leaves the rest to the AI:

Context

What it settles

What the AI still decides

The approved translation of your key terms. The AI is instructed to always use them.

Every other word in the sentence, and how the term fits into it (for example, plural or gender). On rare occasions a term can still be missed, which Smells flags.

How similar sentences were translated before.

How to adapt that wording to the new sentence. A TM match is a reference, not a copy, unless you prioritize TM matches (see below).

The domain, audience, formality and tone, plus your free-text Instructions.

Which exact words express them. Because Instructions are free text, the AI has to interpret them: "Keep it close to the source" leaves little room, while "Write in an ornate, literary style" opens up many valid wordings, so the output can vary much more from run to run.

A first draft.

How much of the draft to keep and what to rewrite so it reads naturally.

Even with all four in place, there is usually more than one good way to write the sentence, and applying the ranking to a real sentence (for example, how far to adapt a TM match to the tone in your Sous-Chef profile) is a judgment the AI makes each time it runs.
​

How the AI picks up your corrections

Your team's corrections reach the AI through your context, right away. When a linguist confirms a segment, it's saved to your Translation Memory, and a term added to your glossary or a change to your Sous-Chef profile is available immediately. The very next segment the AI translates can already use them: a similar sentence gets the approved translation as a reference, and the new term is applied.
​

This works without training the AI on your content. Each request to the AI carries the context it needs, and the AI keeps nothing once it answers (this is what's called a "stateless" model). Your data stays private, and the AI's underlying behavior doesn't drift over time. Keep in mind that a confirmed segment only helps sentences similar enough to match it in your TM, and segments that were already translated don't change by themselves: click Regenerate to apply the new context to them, but note that the wording may shift when you do.


Translation quality depends on your context

Context Sensitive can only be as good as the context it receives. The AI writes fluent text in any case, but whether that text uses your terminology, follows your approved wording and sounds like your brand depends on what your TM, glossary and Sous-Chef profile give it to work with.
​

Most suggestions need no changes at all. In wxrks's 2026 Edit Distance Report, which looked at real projects across wxrks accounts, 84 out of every 100 Context Sensitive suggestions were confirmed without a single edit, compared with 67 for machine translation alone. That's also why the rest deserve a careful read: when most suggestions are right, it's easy to stop looking closely.

Context available

What to expect

A mature TM, an up-to-date glossary and clear Sous-Chef Instructions

Translations close to your approved wording and terminology, with the least to edit and the least variation.

A glossary, but little or no TM

Key terms come out right, while the rest of the sentence follows the MT draft and the Sous-Chef profile.

No TM or glossary yet (for example, a new language or a new client)

A fluent but more generic translation, based on the MT draft and the Sous-Chef profile alone. Expect more edits to match your voice at first.

Outdated or conflicting context

The AI follows what it's given. An old glossary term is still used, contradictory TM entries pull in different directions, and vague or conflicting Instructions are interpreted as best the AI can.

The good news is that quality grows with your work. Every segment your linguists confirm becomes part of your Translation Memory and is available to the AI right away, so projects tend to need fewer edits as your TM and glossary mature. To clean up a TM that has collected outdated or inconsistent entries, see wxrks TMill: AI-Powered Translation Memory Clean-Up.
​

💡 Tip: When a translation is correct but doesn't sound like you, open Translation reasoning on the segment. It shows which TM matches, glossary terms and Sous-Chef settings shaped the result, which usually points to the context that's missing or needs fixing.


Why the wording can vary

That final sentence is still written by the AI, and at many points more than one wording fits almost equally well: which TM phrase to keep, how much of the MT draft to change, which word best matches the requested tone. Your context narrows those choices a lot, but it doesn't remove them. So the same sentence, with the same settings, can come back in two versions, often both correct:

Source (English)

First run (Portuguese)

Second run (Portuguese)

Your order is on its way.

Seu pedido está a caminho.

Sua encomenda está a caminho.

ℹ️ Note: The translation examples in this article are made up for illustration. They don't come from real projects or customer content.
​

Why the same request can give a different translation

Even with exactly the same context, on the very same segment, asking for a new translation (for example with Regenerate) can return different wording. This comes from how large language models (LLMs), the kind of AI behind Context Sensitive, work.
​

Before writing anything, the AI reads everything it was given at once: the source sentence, the TM matches, the MT draft, the glossary terms and your Sous-Chef settings. It then writes the translation out in order, and every word it picks takes all of that into account, along with the words it has already written. That's also true when the result ends up close to a TM match or the MT draft: the AI writes it out again rather than pasting it in.
​

At each step, it scores the possible next words and, on the most consistent setting, which is what wxrks uses, it picks the highest-scoring one. Often two options score almost exactly the same, like "Seu" and "Sua" at the start of the example above. The gap is so small that tiny differences in how the AI provider's servers do the math (they process many requests at once, never in quite the same way twice) can be enough to put the other option on top.
​

Once one word changes, the rest of the sentence is written to follow it: after "Sua", the AI continues with "encomenda" instead of "pedido". That's how a near-tie on a single word turns into a differently worded sentence. This is true of every tool built on LLMs, not something specific to wxrks, and it's why wxrks gives you ways to lock wording down where it matters (see How to get consistent, reviewed output).
​

Two other things can change the wording:

  • Each segment is translated on its own. The same sentence in two places in a file can come out worded differently. To keep repetitions identical, copy the confirmed version to the other segment, or turn on Prioritize TM matches over Context Sensitive translation and regenerate it, so it reuses the confirmed translation as it is.

  • The AI models are updated. AI providers improve their models over time, and a newer model can phrase the same sentence differently. This also applies if your account uses its own AI provider.


What this means for your projects

Most variation is harmless: two correct ways of saying the same thing. Occasionally, though, AI output can include small artifacts that don't belong in the translation. They're uncommon, but easy to miss on a quick read, for example:

  • Strange characters: Welcome to our store! → Bem-vindo à nossa loja! Ֆ

  • Punctuation, or an invisible character: Welcome to our store! → Bem-vindo à nossa loja!। (an invisible character looks perfectly normal on screen, but can break character limits or how the text displays; the wxrks Editor shows it, so reviewers can spot and delete it)

wxrks runs automated checks and clean-up on AI output and keeps improving them, but no automated check catches everything, which is why a review step still matters even when these issues are rare.


How to get consistent, reviewed output

The AI in wxrks is built to speed up and support your linguists, not to replace them. Your team stays the final judge of every translation.

  • Reuse approved wording exactly. Turn on Prioritize TM matches over Context Sensitive translation. Segments with a strong enough TM match then get the approved translation as it is, without the AI. Keep the threshold at 100% or 101%: below that, a fuzzy match is reused as it is, without being adapted to the new sentence. See Let strong TM matches win over Context Sensitive.

  • Keep your context clean. An up-to-date glossary, a Translation Memory without outdated entries and clear, specific Sous-Chef Instructions leave the AI fewer wrong options to choose from. If consistency matters more than style, keep the Instructions focused on what must not change.

  • Add a review step to the workflow. A linguist reading every segment against the source is the most reliable way to catch anything that doesn't belong. See How to review in wxrks.

  • Use the quality checks as a second pair of eyes. Translation Smells and the QA Check Tool point reviewers to likely problems. A flagged segment is worth a look, but a segment with no warnings isn't guaranteed to be correct.

  • Check the final file. For subtitles, UI text or web pages, look at the exported file or published page too: some problems, such as a hidden character breaking a length limit, only show up there.

💡 Tip: Content where a mistake is costly (legal, medical, financial or safety text, prices and product claims) deserves a review step every time.


Reviewing AI output before you use it

AI translations are machine-generated, so they may contain errors. Think of them as a strong first draft: they're meant to be reviewed before you rely on them, and your team decides when a translation is ready and right for its purpose. Our Terms of Use cover this in more detail.
​

If something looks like a platform problem rather than a one-off AI mistake (for example, the same unusual characters showing up across many segments or projects), let us know at [email protected] with the project and a few example segments, and we'll look into it.


Troubleshooting

The translation is correct, but it doesn't sound like us. The AI is missing context. Check that the project uses the right TM, glossary and Sous-Chef profile, and open Translation reasoning to see what was used. In a new language or for a new client, quality improves as your linguists confirm segments.
​

I clicked Regenerate and got a different translation. This is expected, even when nothing in the context changed (see Why the same request can give a different translation). Choose the version that fits best, or edit it yourself.
​

The same sentence was translated differently in two segments. Each segment is translated on its own. Copy the confirmed version to the other segment, or turn on Prioritize TM matches over Context Sensitive translation and regenerate it, so it reuses the confirmed translation as it is.
​

We corrected a translation or updated the glossary, but another segment didn't change. Check that the corrected segment was confirmed, so it's saved to your Translation Memory, and remember it only helps sentences similar enough to match it. For a key term, add it to your glossary. Segments translated before the change need a Regenerate to pick it up.
​

A translation has odd characters at the end. Delete them in the Editor or click Regenerate.
​

Still stuck? Search the wxrks Community or contact [email protected].
​

Related articles

Did this answer your question?