Machine Translation and Post-Editing: The Review Became the Job

For years the easiest way to argue that humans beat machines at translation was to paste a sentence into a free engine and laugh at what came back. Those screenshots made the rounds for a decade.

The joke is out of date. Machine translation in 2026 rarely produces the howlers that built its reputation, and the interesting question moved. It is no longer whether a machine can translate. It is who is accountable for what the machine produces, and what that person actually does all day.

The short answer to the second half: they review. The machine made the first draft close to free, and the review that catches what the draft gets confidently wrong is now the thing you are buying. This post covers the terms, where the tooling actually sits, and the four error types that post-editing exists to catch.

[Average read time: 8 minutes]

The terms, plainly

A handful of words get used loosely once machines enter a workflow, so define them before using them.

• Machine translation (MT): software that converts text between languages using a dedicated neural engine.
• Neural machine translation (NMT): the model architecture behind the classic engines, now often paired with LLM features.
• AI translation: translation produced by a general-purpose large language model rather than a dedicated MT engine. The industry increasingly treats these as separate categories even as commercial platforms blend them.
• Post-editing machine translation (MTPE): a human translator revising machine output, either lightly or fully.
• Transcreation: creative adaptation that recreates the intent and effect of a message rather than its literal words. We cover it separately, because it is a different job rather than a harder version of this one.
• Localization (l10n): adapting content for a specific market, including language, formatting, currency, and cultural references, not just the words.

What large language models actually added

The technology split into two approaches and then partly re-merged, which is why the labels confuse people.

Dedicated NMT engines came first. General-purpose large language models came second, used for translation among many other tasks. The line between them is now blurry: DeepL markets a next-generation LLM translation model alongside its classic engine (Source: DeepL, 2026), and Google Cloud Translation offers a standard NMT mode, an LLM translation mode, and an adaptive mode at different price points (Source: Google Cloud, 2026). Both are fairly described as hybrid platforms rather than pure NMT, and the fact that the modes are separately priced tells you the vendors themselves treat them as different products.

What the language models added is mostly context and control. They can hold a whole document in view rather than translating sentence by sentence, follow instructions about tone and register, and apply a glossary or style guide when prompted. In a preliminary evaluation based on automatic metrics at the 2024 Conference on Machine Translation, LLM-based systems outscored traditional online and NMT systems across eleven language pairs (Source: Slator, July 2024). Keep that in perspective: it was an automatic-metric ranking, human evaluation supersedes automatic metrics, and the metrics themselves carry bias. The direction is still clear.

The trade-offs came with it. These models hallucinate, meaning they invent content that was never in the source. They can give different output on two runs of the same text. And their mistakes read fluently.

That last point is the whole reason this post exists. A clumsy human draft signals its own problems, because awkward phrasing tells a reviewer where to look. A smooth machine sentence gives no such tell. Fluency hides error, which makes post-editing a judgment skill rather than a cleanup task.

Post-editing is the default workflow now

Almost no serious translation job is press-the-button-and-ship. The dominant professional workflow is post-editing, where a human translator revises machine output instead of translating from a blank page. That workflow is named in ISO 18587:2017, the international standard for post-editing of machine translation output (Source: RWS, 2017). When you have 200 videos and a two-week turnaround, nobody translates from a blank page.

The standard names two depths, and the distinction matters when you are reading a quote. Light post-editing fixes errors that affect meaning and leaves the rest. Full post-editing brings the text up to a standard comparable to professional human translation. Only full post-editing meets the standard's quality bar; the lighter level sits in an annex. A vendor quoting "ISO 18587" without saying which depth has not told you much.

How much faster is it? It depends heavily on the language pair, the subject matter, and how clean the draft is. Post-editing is usually faster than starting from blank, and the saving is narrower than vendors like to imply. Published productivity studies exist, but most measure a single domain and do not transfer cleanly to multimedia localization, where subtitle timing and dubbing adaptation slow the editing down. Treat any single percentage as an illustration of direction rather than a guarantee for your content.

The part that matters for a buyer is what the human's job becomes. Less time generating words, more time judging them: catching the fluent-but-wrong sentence, fixing terminology, deciding what tone the audience needs, and signing off. That is not uniform across formats. For text-heavy work the machine drafts and a person owns the final version. In multimedia, where subtitle segmentation and dubbing adaptation are harder to draft cleanly, the person still carries more of the work up front.

Post-editing machine output is also not the same job as proofreading a human translation, and staffing it as if it were is a common mistake. The error distributions differ. A human translator errs on idiom, on nuance, on the slips that come with fatigue near the end of a long day, and you learn to anticipate those. A model fabricates with confidence and pattern-matches in ways that look right and are not. A reviewer trained on one is not automatically ready for the other.

The four error types, and the method for each

1. Hallucinations: check segment against segment, never the target alone

A hallucination is content the model invented that was never in the source. An added clause, a fabricated specific, a plausible detail that is simply not there. The engine produces it because its job is to generate fluent text, and a smooth invented sentence scores well by that measure even when it is false.

Take a corporate training script where the English source says "complete the form." The output reads "complete the form within five business days." Fluent, plausible, absent from the source.

The method: review at segment level. Source and target side by side, line for line, confirming that every claim in the target traces back to the source. Never review the target in isolation, because in isolation there is nothing wrong with it. We run this segment-level check on every production.

2. False friends: a native speaker of the target language, not a fluent one

A false friend is a word that resembles a word in another language and means something else. Models map by surface similarity, so they are prone to them. The English "actually" is not the French "actuellement," which means "currently" (Source: FluentU). The Spanish "embarazada" does not mean "embarrassed"; it means "pregnant" (Source: The Spanish Experiment).

So an English line reading "we were embarrassed by the delay" can come back as a sentence about a pregnancy.

The method: keep a false-friend checklist per language pair, and staff a native speaker of the target language as reviewer. A native speaker hears the wrong word the way you hear a wrong note. Fluency in the language is not the same qualification.

3. Gender and agreement: track referents at document level

In gendered languages the model has to guess grammatical gender and the gender of the people in the script, and it guesses from local context that is often too thin. It mis-genders a speaker, or breaks an agreement chain when the referent sits several sentences from the word that has to agree with it.

Picture a testimonial where a woman describes her experience and the engine, working sentence by sentence, renders her later self-references in the masculine because the immediate sentence gave it nothing to anchor on.

This is an old failure that survived the upgrade to language models. Translating from a language with no gendered third-person pronoun, such as Turkish or Uzbek, the model guesses from training data, so "the doctor" drifts male and "the nurse" drifts female.

The method: track every referent across the whole script rather than segment by segment. Note who each "they," "it," or implied subject points to, then flag any place where gender or agreement does not hold across the full text.

4. Register: fix it in the brief, before anyone edits

Register is the formality level, and many languages encode it grammatically in a way English does not. Spanish has "tú" and "usted," German "du" and "Sie," French "tu" and "vous" (Source: Wikipedia, T–V distinction). An engine picks one with no view of your audience, and it will switch inside the same script without noticing.

A banking app addressing customers wants "Sie" throughout in German. The engine drops into "du" for a few lines and the brand voice cracks.

One caveat worth keeping: the formal-informal split is not always a clean binary. Spanish also has plural forms like "vosotros" and "ustedes," so the choice can run wider than two options.

The method: decide the register before post-editing starts, write it into the project brief, and check the whole script against that one decision.

What stays hard, even now

Language models closed some gaps that used to be reliable human territory, so it is worth being specific about what genuinely remains difficult.

Intensity within a register. "Happy" and "joy" point at the same feeling, but you would not swap them freely. A model optimizing for a literal match can pick the technically correct word that lands at the wrong emotional level for a wedding invitation versus a legal notice.

Homonyms. The English word "register" can mean a level of formality, the machine at a checkout, or the act of signing up to vote. The right choice depends on context the model may not have.

Idiom, unevenly. Common ones like "raining cats and dogs" now tend to come through. Reliability still varies by language pair, by how rare the idiom is, and by how much context the sentence gives.

And there is a category beyond all of these where the job is not to translate the words but to recreate the effect: a slogan, a joke, a brand voice that has to land the same way in a new market. A model is genuinely useful there as a generator of alternatives a copywriter can react to. What it does not do is know which one lands. That is a different craft and we cover it in our transcreation post.

Two threads worth keeping apart

Two legal threads touch this work, and they are separate.

The EU AI Act's Article 50 transparency obligations for AI-generated media take effect on 2 August 2026, with the disclosure duty falling on the party that publishes. Separately, performer consent for voice cloning and digital replicas is governed in the US by SAG-AFTRA agreements and by state statute, not by EU law. One is a disclosure duty owed to an audience; the other is a consent regime owed to a person. Our disclosure and provenance post covers the first, our voice-cloning post the second. Have counsel confirm what applies to your specific production.

Where this leaves your workflow

AI moved the cost of a first draft close to zero, and it did so honestly: speed, coverage, consistency on the repetitive parts. That was never the hard part of localization.

The hard part is the segment-level check, the native ear, the referent tracked across forty lines, the register fixed before anyone hit record. If you are localizing at volume, the question is not whether to use AI for the draft. It is whether your post-editing catches what the draft gets confidently wrong, and whether the person doing it can name the four things they were looking for.

Our white paper, Localizing in the Age of Generative AI, sets out where that review belongs in the pipeline.