Ask a project manager what machine translation post editing involves and you will often get an answer that sounds like proofreading. Read the output, fix the obvious mistakes, send it back. That description survives right up until somebody tries to price it, at which point it becomes clear that nobody agrees what the work actually is. Post editing is not proofreading, it is not translation, and treating it as either produces bad estimates and unhappy linguists.
Two Levels, Two Different Jobs
The industry distinguishes between light and full post editing, and the distinction is the single most useful thing a buyer can learn.
Light post editing aims at accurate and understandable, nothing more. Grammar errors that obscure meaning get fixed. Awkward phrasing that a reader can still parse is left alone. It suits internal documents, support tickets, large product catalogues and anything where the alternative is no translation at all.
Full post editing aims at output indistinguishable from a human translation. Style, register, terminology and flow all get attention, and the editor is expected to reject and rewrite whole segments rather than patch them. The formal requirements for this sit in ISO 18587, which defines the process and the competences a post editor needs.
The failure mode is asking for light post editing at light post editing rates and then reviewing the result against full post editing expectations. It happens constantly, and it is usually a briefing problem rather than a quality problem.
Why Good Output Can Be Harder to Edit
There is a counterintuitive effect that anybody who has done this work recognises. Weak machine output is easy to handle because the errors announce themselves. Strong output is harder, because the remaining errors are fluent.
Modern neural machine translation produces text that reads well at the sentence level and occasionally means the opposite of the source. A dropped negation, a swapped subject, a confident hallucinated number. The prose glides past, and a tired editor working at speed glides past with it.
This is why post editing rates based purely on a percentage of the translation rate tend to be wrong. The real variable is not how much text there is. It is how much attention each segment demands, and that depends on the engine, the language pair and the subject matter.
Where Machine Translation Post Editing Genuinely Pays
Repetitive, high volume, low risk content is the sweet spot. Product descriptions, knowledge base articles, internal reporting, user reviews, historical archives that nobody would otherwise fund. In those categories the economics are not close, and refusing to use the technology is simply choosing to translate less.
It pays badly, or not at all, in three places. Marketing copy, where the point is the phrasing and rewriting from scratch is faster. Very small volumes, where the setup overhead swamps the saving. And anything with legal or safety exposure, where the review burden required to be confident removes the efficiency gain entirely.
Language pair matters more than vendors like to admit. English into Spanish or German behaves very differently from English into Finnish, Hungarian or Japanese, where morphology and word order give the engine more chances to go wrong and the editor more to unpick.
Post-editing exposes a limit that raw engine output can never cross on its own: the point where a document has to be legally defensible rather than merely readable. That is exactly where certified translation services take over from the machine, because a sworn professional signs their name to the accuracy of the result. The editing pass improves fluency, but only human certification carries weight with a registry, a court or a consulate.
The Setup That Makes It Work
Three things separate programmes that deliver from programmes that quietly get abandoned.
First, the engine should be trained or at least tuned on your material. A generic engine and a domain adapted one produce visibly different amounts of editing work, and the difference compounds over a year.
Second, the existing language assets should be in play before the engine is. A well maintained translation memory and an enforced glossary catch the segments that were already solved, so the editor only sees genuinely new text. Running machine translation over content you have already translated is a common and expensive mistake.
Third, measure something. Edit distance, time per thousand words, error categories, anything that tells you whether the engine is improving. Translation quality assurance in a post editing workflow is not a final check, it is the feedback loop that tells you where to retrain.
The Human Side That Gets Ignored
Post editing is genuinely less pleasant work than translating. It is reactive rather than creative, it rewards vigilance over craft, and linguists who do it all day report fatigue faster. Programmes that treat post editors as a cheap tier of translator lose their best people within a year and then wonder why quality slid.
The organisations that hold onto good editors tend to do three obvious things: pay by time rather than by word where the content is unpredictable, mix post editing with ordinary translation work, and actually act on the error reports rather than filing them. None of this is complicated, and almost all of it is skipped.
PoliLingua's post editing strategy guide covers the process side in more detail and is a reasonable document to bring to an internal planning conversation.
A Short Answer for Anyone Deciding
Machine translation post editing is a tool for a category of content, not a replacement for translation. Sort your material by risk and by how much the phrasing matters. Push the bottom two thirds through an adapted engine with real editors and a measured workflow. Keep the top third human from the first word. Then leave the split alone for at least six months, because the temptation to move the line every quarter is what turns a working programme into a quality incident.