What machine output changes
Post-editing is a different cognitive task from translating — and its characteristic error is text that reads well and means something slightly wrong.
Post-editing is a different job from translating, with a different failure mode: the errors are fluent.
The shape of the failure
A translator working from scratch makes visible errors: a missed clause, an unconvincing word choice, a sentence that halts because the structure hasn't resolved. The errors signal themselves. Post-editing machine translation inverts this. The output arrives fluent — plausible rhythm, correct-looking grammar, the confident prose of a system trained on billions of words — and the errors hide inside that fluency. A number is off by an order of magnitude. A negation has been dropped. A conditional has quietly become a statement of fact. None of this breaks the surface.
The standard specifies qualifications, a second qualified person revising the whole text, and records — because the failure it exists to prevent is physical. What ISO 17100 asks for
This is not a new observation. Researchers studying post-editor behaviour have consistently found that editors are less likely to catch errors when the machine output is of high surface quality, because fluency lowers vigilance. The text reads like text that has been checked. The eye moves through it as if it has been verified. It has not.
What the job actually requires
Translation, at its core, is a transfer of meaning: the translator holds the source text in mind and constructs a target text that carries that meaning into another language. Post-editing is something else. The post-editor must hold two texts simultaneously — source and machine output — and resist the pull of the output. That resistance is the skill. The brain's default mode is to read the target text as a completed object and to find meaning in it, whether or not that meaning corresponds to the source. Catching a fluent error requires the editor to override this, to return constantly to the source, to treat every plausible-looking sentence as a hypothesis rather than a result.
This is measurably more cognitively demanding than it appears from the outside, and different in kind from conventional revision. In revision by a second translator, the reviser compares a human text against a source; the errors tend to cluster around difficulty — ambiguous terms, long-range dependencies, culture-specific content. In post-editing, difficulty and error are decoupled. The machine handles difficult passages with apparent ease, and introduces errors in sentences a competent junior translator would never get wrong.
Revision by a second translator is not proofreading; it is a comparison against the source, line by line. Somebody else reads it first
The practical implication is that post-editing cannot be treated as faster translating with less to do. A linguist who edits at reading speed, relying on the machine's fluency as a proxy for accuracy, is not post-editing — they are proofreading in the wrong direction.
What it changes for the text
There is a secondary effect on the target text itself. Machine translation optimises for the statistically probable rendering, which tends to be the most common, most central, most unspecific version of a meaning. Terminology — one object, one name, used consistently across a document set — is where this shows up first. A machine may render the same technical term three ways across a long document, each one defensible in isolation, none of them the agreed term in the client's term base. The post-editor who is moving quickly will normalise these into something reasonable. Reasonable is not the same as correct.
The post-editor must hold two texts simultaneously — source and machine output — and resist the pull of the output
The same logic applies to legal text, where the Court of Justice of the European Union requires precise equivalence across its working languages, and to regulated technical content, where ISO standards specify that the target text be verified against the source by a qualified second person — a requirement that machine-assisted workflows have not dissolved, regardless of how they restructure the steps.
The skill that doesn't transfer automatically
None of this means post-editing is harder than translating, or easier. It means it is different, and that translators moving into post-editing workflows need to be trained specifically for the failure mode they are now managing, not for the one they already know. Fluency is not accuracy. A text that reads well has only passed the lowest test. Post-editing is the discipline of knowing that, and not forgetting it mid-page.