From MTPE to AI Post-Editing: A Methodology Upgrade

The craft of post-editing (PE) is undergoing a methodology upgrade as the engines change generations. In the neural MT era, post-editing (MTPE) was chiefly about "correcting errors": MT failure modes were relatively predictable — omissions, inconsistent terminology, stiff phrasing, mangled long sentences — and a veteran post-editor could locate problems at a glance.

Large language model output fails in an entirely different way: fluency is extremely high, with almost none of MT's "translationese" — but it "errs confidently": fabricating plausible details, quietly altering figures and proper nouns, deleting or adding information in pursuit of smoothness. Such errors are highly covert, nearly impossible to catch in a read-through; sentence-by-sentence verification against the source is required.

Accordingly, the center of gravity in AI post-editing shifts from "correcting" to "verifying": figures, dates, proper nouns, negations and qualifiers in legal clauses must be checked item by item; terminology consistency needs tool-assisted validation; fluency, conversely, is not the problem. The competency profile of post-editors shifts with it — language-polishing skill yields priority to fact-checking and a nose for risk.

Industry training is following: multiple professional associations and training bodies updated their post-editing curricula in 2023, adding "verification methods for LLM output" as a core module. As tools evolve, methodology must evolve — this is what professionalism concretely means in the AI era.