84 of 100 AI Translation Suggestions Confirmed Unchanged

Everyone wants to know what AI is really doing to translation. Your team already works with it. Every suggestion passes a person who changes what needs changing and confirms the rest. How much they change is the number your leadership asks for, and it usually gets answered with an opinion.

We had a way to measure it. Over two years, people on wxrks confirmed 27 million translated sentences. Each one arrived as a suggestion, from a memory match, from our context-sensitive AI, from plain machine translation, or from a colleague, and we recorded how much the person changed before confirming it. Nobody read the text. The study saw only the measurements, in groups, with no client identifiable.

The result is The Edit Distance Report, our second study on this question and six times the size of the first. Here is what it found, and what it means for the person who runs the work.

What edit distance measures

A suggested translation goes in front of a person. They read it, fix what needs fixing, and confirm. Edit distance is how much they changed. Zero means they kept it as written. One hundred means they rewrote every word. We count it as words changed in every 100, with the translation edit rate the research field has used since 2006, so the 2024 study and this one compare.

It's a blunt instrument, and that is its value. It doesn't ask anyone what they think of the AI. It records what they did with it, one sentence at a time. We've written before about edit distance and translation effort. This is the first time we've had enough of it to say something at scale.

15 words in 100 for machine translation, 8 for context, 6 for a person

Since January, every segment on the platform records where its starting text came from. That let us put the sources on one scale.

Plain neural machine translation, a general engine with none of your context, gets 15 words in 100 changed. A context-sensitive suggestion, built from your team's memories, glossaries, and preferences, gets 8. A translation written by a colleague gets 6.

So the suggestion sits 2 points from a person and 47% below plain machine translation. Our first study, published in 2024, found 39% with the same metric. Measured the same way on both editions, the paper's note puts the move at 22% to 47%. Either way the direction is the same. Machine translation engines got better in two years, and context got better faster.

84 of every 100 suggestions are confirmed as they came

The number that got us is the share with zero edits. 84 of every 100 context-sensitive suggestions are confirmed without a single change. Plain machine translation manages 67. A colleague's translation, 87.

Across the whole corpus, half of all segments ship untouched, and one in seven has more than half its words rewritten. The middle, the patch-it-up editing that post-editing was named for, is the thinnest part of the picture. We argued that machine translation post-editing is over. The data now shows what replaced it, acceptance on one side and a rewrite on the other.

Editing falls as the knowledge base fills

Every confirmation and correction on wxrks teaches the account's knowledge base, never a model. If that's true, a team should edit less the longer it works. It does.

A new account changes 41.7% of its text in its first month, because its knowledge base is empty. By month ten it changes about 16%, and it stays there through the end of the data. Each account counts once per month in that curve, so no large customer moves the line.

That curve is the case for keeping the learning where you can see it. A trained model would hold it inside a vendor's weights. Here it lives in your translation memory, your glossaries, and your preferences, in your own account, and it compounds.

Where people still edit

The spread across languages is wide. English into Spanish for the US gets 9 words in 100 changed. English into Italian gets 26. Japanese, Chinese, French, and German sit between 19 and 24.

That map shows where judgment spends its time, and where a glossary or a style preference pays back first. Our guess, and the paper calls it a guess, is that the 16% left at the floor is no longer mistakes. It's choices about tone, audience, and preference that no memory holds yet, which is the work Sous-Chef, our agent for tone and style, is built to carry.

Even at zero edits, someone decides

Here is the part that matters if you run a team. Imagine the technology gets perfect. Zero edits. The person is still reading. Per word, a context-sensitive suggestion takes 1.35 seconds to review and plain machine translation 1.45, a gap of about 7%. The time was never in the typing. It's in reading, and in deciding what gets read.

The second pass makes the point. After the first translation pass, review changes 0.5% of what it reads, and in-country review changes 0.2%. Almost all of a reviewer's day goes to text they will confirm as written.

So the gain from AI doesn't show up in the editor. It shows up in the workflow, and someone designs the workflow. From the data, that person decides three things.

What gets reviewed. Keep the full read where the risk is, in legal and in campaigns. Everywhere else, Smells, our check that flags meaning problems like an omission or a distortion, marks the segments where meaning moved, and a reviewer can work the flags. The paper lists review by exception as a hypothesis to test, so read "can," not "does."

What gets reused. A correction that stays in the memory and the glossary is what takes a team from 41.7% changed to 16%. A correction that lives in one file and dies there is paid for again next month.

How the company's context is organized. Which memories apply to which content, what the glossary says, who the audience is, and what tone it gets. That setup decides what the suggestion looks like before anyone reads it.

That someone is the localization manager. Or the translation agency. Or whoever designs the workflow. That is where the result gets decided. The 8 words in 100 that people still change are the judgment, and they have never mattered more.

What the data leaves open

The data is observational. Accounts choose which content goes to a person and which goes to the AI, so the sources may not carry the same kind of text. Acceptance is not quality. 84 of 100 confirmed unchanged says the suggestions were good enough to keep, and only a blind quality review can say more. The paper names eleven open questions, five from the first study and six from this one, and says how each could be tested.

The full paper is open, with the numbers for each language pair, the method, and every figure. Read The Edit Distance Report, bring it to your leadership, and tell us where else the person orchestrating the workflow adds value.

Every number here is from The Edit Distance Report (wxrks Research, WP-2026-01, revised October 2026), built from anonymized group aggregates. No client is identifiable and no text was read.