AI for translators

At three in the morning on September 21, 2017, a Kansas state trooper typed “Can I search your car?” into Google Translate and turned his laptop toward the driver. The screen read “¿Puedo buscar el auto?”, which a Spanish speaker understands as “Can I find the car?”. The driver said “yeah, yeah go.” The following June a federal judge threw out everything the search found, because nobody can consent to a question they were never asked.

That is the whole problem with machine translation in one sentence. The output was instant, free, fluent, and wrong in the one place that mattered. Translators have lived with this for a decade. What changed with large language models is that the drafts got good enough to use and the failures got harder to see.

Let the model write the draft, on a machine you control, and spend your hours where it fails: idiom, tone, legal precision, and languages with little training data. The job did not shrink. It moved to the part a client cannot check. This post shows where the line sits with real numbers from the profession’s own surveys, a test we ran on three open models, and a setup that keeps client files on your own disk.

The fluent draft is the new problem

The best machines now sit next to professionals on ordinary text. At the 2025 edition of WMT, the yearly shared task where translation systems are scored by human annotators, Gemini 2.5 Pro landed in the top cluster for 14 of 15 language pairs, and the human reference translations made the winning cluster in only six. The organizers had to invent a harder sampling method because the old test sets were no longer telling systems apart. In the same report, the specialized test suites found that state-of-the-art systems still struggle with non-standard input, linguistic complexity, domain terminology, and gender agreement in particular language pairs.

Translators see both halves of that result every day, and their opinion of the output is falling as they use it more. The 2026 European Language Industry Survey found only 23% of independent professionals rating machine or AI translation quality as high or very high, down from 40% a year earlier. The report’s own reading: “the effect of the increased fluency that GenAI introduced is wearing off.” The American Translators Association put the danger more plainly in its May 2025 statement: AI output can appear accurate to the general observer, which is exactly what makes its errors hard for a non-expert to catch.

Most translators already use it, at about a 20% discount

Whether to use machine output is no longer the question. ELIS 2026, which gathered answers from 564 independent language professionals in 45 countries, found that automated translation is now involved in 63% of independent translators’ work, split evenly between post-editing text a client had already run through a machine and translators using a model on their own initiative. The share depends on who is asking: under 15% of work from direct business clients and public bodies arrives pre-translated, against 53% of the work outsourced by large language companies.

The money follows the machine. Translator forums talk about 70% and 90% discounts for post-editing; the survey data show a narrow band centered on 20%, from 13% reported by independents to 29% reported by language companies for their external post-editors. The Belgian translators’ association measured the same gap in euros: in its 2025 rates survey of 176 freelancers, the median agency rate for Dutch to French translation was €0.09 per source word, and the median for post-editing the same pair was €0.06, a third less. The academic version of that finding, published in Perspectives in December 2025, is that dissatisfaction comes from low rates rather than from the pricing method, and that professional satisfaction stays high anyway.

The mood is worse than the rates. Only 22% of ELIS respondents gave a confident yes to whether freelancing is a sustainable long-term activity, 17% are considering leaving, and the report describes free text comments mentioning income drops of more than half. An ISO standard, 18587, has defined full human post-editing since 2017. What no standard defines is which half of the work you should be charging full rate for.

Hand the model the volume, keep the judgment

The split is the same one we drew for lawyers: the machine does volume, the professional does judgment, and every job has both halves.

The first draft is the obvious win, and the prompt is where most of the quality is decided. Tell the model who will read the text, which register to use (German has two forms of “you”, and all three models in our test had to guess), which terms are fixed, and which names, codes and figures must come through untouched. A glossary pasted above the text does more for consistency than any amount of editing after the fact. Then read every sentence against the source, not against the draft, because a fluent draft argues for itself.

The second win is the one translators underrate: checking your own work. Ask a model to back-translate your German into English and read the result cold. Ask it to list every number, date and proper noun in both versions side by side. Ask it which sentence in the source has no counterpart in the target. These are mechanical questions with checkable answers, which is what models are good at. What they are bad at is grading quality, so never ask the model that wrote the draft whether the draft is good.

It fails on idiom, tone, legal terms and small languages

Grammar is not the weak point any more; meaning carried by culture is. Appen’s 2026 cultural nuance benchmark had five native speakers per language rate translations from seven multilingual models into 15 languages. Mean full-text quality was 1.68 on a 0 to 3 scale. By category, holiday references scored 2.20 and culturally embedded concepts 2.19, while idioms scored 1.65 and puns 1.45, and idioms were the segments most often left untranslated altogether. The strongest model, GPT-5, reached 2.10 overall.

Smaller languages get a worse version of every problem. A study at EAMT 2026 ran 15 models over 22 language pairs and found that pairs not involving English scored consistently lower, with quality tracking how much of a language’s vocabulary the model actually activates when it writes. At WMT25, the organizers could not even score English to Bhojpuri or Maasai with the usual automatic metrics, because no metric has been shown to work for them, and fell back to a character-overlap measure. Coverage claims hide this. Meta’s NLLB-200 translates 200 languages, Google Translate added 110 in one June 2024 release, and a language being on the list says nothing about whether a contract in it comes out right.

We wanted to see the failure up close in a language pair that is as well-resourced as any, so we wrote a five-sentence business email with two idioms, a legal term of art, a term of trade and an unstated register, and sent it to three open-weight models through OpenRouter. One prompt, first result each, no rerolls.

On the easy parts the models agree. All three chose the formal “Sie”, all three found a German idiom for “out of the woods” instead of translating the trees, and all three rendered “the ball is in your court” with the same German phrase, which is in real use. In two seconds and for well under a tenth of a cent, that is a serviceable draft. Then look at the clause. Qwen produced the standard German formula. Gemma produced a shorter phrase German lawyers understand, though the longer formula is the usual one in Germany. Mistral wrote “ohne Vorbehalt”, which means “without reservation”: a letter that was supposed to reserve every right now gives them all up, in perfectly good German, in the one sentence the client cared about.

The term of trade splits them three ways too: a literal compound, the English word kept, and an invented one, with no way to know from the output which the client’s own contracts use. That is what the translator is for. The three drafts took under eight seconds in total. Reading them against the source took far longer than that, and the reading is the job.

Client documents should never leave your machine

In September 2017 the Norwegian broadcaster NRK found that contracts, workforce reduction plans and dismissal letters from the oil company Statoil could be read by anyone with a search engine. Employees had pasted them into a free translation website that stored submissions and let search engines index them. When the trade publication Slator searched the same site it found a physician’s correspondence with a pharmaceutical company, a bank’s staff performance report and termination letters, names and phone numbers included. The Oslo Stock Exchange responded by blocking its staff from the site and from Google Translate. The company explained that until late 2015 it had used volunteers, so submitted text had been public by design.

The free tools of 2026 are more careful and still not yours. DeepL’s privacy policy says content uploaded to the free translator is processed “for a limited period of time to train and improve our neural networks”, and that you may not use the free version for “texts containing personal data of any kind”; only the Pro tiers promise deletion after the job and no training. The International Federation of Translators’ June 2025 position paper names confidentiality as a critical concern with free online services and says AI cannot yet ensure it in legal proceedings or medical consultations. The ATA statement warns that tools which store or reuse text can put a translator in breach of a confidentiality agreement.

That sorts translation tools into three tiers, the same three we drew for lawyers. A free web tool is for text that is already public. A paid API or Pro plan with a written no-training, no-retention clause is for text the client has cleared for it. And for everything under a non-disclosure agreement, which for most translators is most of the work, the only setting that needs no trust at all is a model running on your own disk. We covered what happens to text sent to a cloud service in what happens to your AI chats; for client files the short version is that the question should never come up.

A local setup that fits a translator’s laptop

Two open families cover most working language pairs well. Google’s Gemma 4 ships under Apache 2.0 with out-of-the-box support for more than 35 languages and pre-training on more than 140, in sizes from a phone model to a 31B desktop model. Alibaba’s Qwen3 line was trained on 119 languages and dialects and its current 9B and 27B builds run on the same hardware. Unbabel’s Tower+ is tuned for translation specifically across 26 language variants, but its license is non-commercial, so it is a tool for study rather than paid work. Our Gemma 4 guide and Qwen guide cover the families in depth.

The hardware bar is memory, not speed. A 16 GB laptop runs the 12B and 9B models with room for the document; 32 GB opens the two larger models that produced the cleanest drafts in our test. Our RAM guide explains why memory, not the processor, decides what you can load. A local draft arrives slower than a cloud one (our every-size-of-machine post has measured speeds per class of computer), but it is still faster than typing and it leaves no copy anywhere.

In CSuite the models above run through the bundled local runtime, and the same translation actions also work through a cloud provider when the text is public. Select a passage and the writing menu’s Translate to offers twelve languages, with the translation streaming in over the selection; for a whole document, duplicate it first, then give the instruction in the composer with your glossary and register pasted above the text. The text guide walks through the page by page workflow for long files.

The model drafts, the translator signs

Three rules cover it. Nothing under a non-disclosure agreement leaves your machine unless the client has said in writing that it may. Every idiom, number, name and legal phrase in the draft is read against the source, not against the draft, before your name goes on it. And the rate you quote covers that reading: the survey band is a 20% discount for post-editing, not 70%, and the clause Mistral reversed is the reason.

The Kansas court drew the line better than any vendor has. It said an officer may reasonably use a translation tool for basic information, but not to obtain consent. Basic information is the model’s half. Consent, meaning anything a person will act on, is yours. If you translate for family rather than for clients, the same rule holds at a smaller scale, and our post on learning a language with AI covers the other side of the same tool. For working translators, the machine has taken the typing. Keep the signature.

Disclaimer: This is general information, not legal advice. Tool licenses, content-usage rights, and platform policies summarized here change frequently and reflect sources available as of October 2026. Verify the current terms of each tool and the rules of each platform or marketplace before publishing commercial work, and consult counsel where real money or rights are at stake.