← Blog · 2026-10-13

Machine translation post-editing: MTPE and ISO 18587

Author: Yuri Vodostoy — Technical translator and founder of Insight Translation

Machine translation post-editing (MTPE) is a human service: a linguist edits and corrects machine translation output. ISO 18587:2017 defines two levels. Full post-editing aims at a product comparable to human translation. Light post-editing aims at a text that is merely comprehensible, and the standard sets requirements for the full level only. HardTrans does not sell post-editing. We produce the draft a post-editor starts from.

What post-editing means in ISO 18587:2017

To post-edit is to "edit and correct machine translation output." That is the definition in ISO 18587:2017, Translation services — Post-editing of machine translation output — Requirements, published in April 2017.

The scope takes three sentences. The standard "provides requirements for the process of full, human post-editing of machine translation output and post-editors' competences." It is meant for translation service providers (TSPs), their clients, and post-editors. It is "only applicable to content processed by MT systems."

The introduction of the 2017 edition gives the reason the service exists: "there is no MT system with an output which can be qualified as equal to the output of human translation." That sentence is nine years old, and the engines have changed since. The draft of the next edition, covered below, widens the subject from machine translation to "non-human translation" output.

The full text is paid. Everything we quote from the standard comes from the public abstract, the free preview (introduction, scope, terms, table of contents), and a summary by the ISO committee of the Association of Translation Companies (ATC).

Light vs full post-editing

Light post-editing makes a text understandable, full post-editing makes it publishable, and the amount of editing follows from that target.

The last two rows come from "A Comparative Study of Post-editing Guidelines" by Ke Hu and Patrick Cadwell of Dublin City University (Baltic Journal of Modern Computing, 2016), which compares five sets of guidelines and quotes TAUS and Common Sense Advisory. The same paper cites a 2014 blog post by Lee Densmer of Moravia on both levels. Light post-editing is not an easy job for linguists, who have to "turn a blind eye" to minor errors. And the effort to bring machine output to human level "may exceed the effort to have it translated by a linguist in the first place."

What ISO 18587 requires, and what it leaves out

ISO 18587:2017 is a process standard. It regulates how a post-editing job is run and who does it.

The table of contents in the free preview shows the shape. Clause 4 covers the post-editing process: pre-production, production, post-production. Clause 5 covers post-editors: competences, qualifications, professionalism. The annexes deal with post-editor training, light post-editing, pre-editing, client–TSP agreements, and automatic post-editing.

The detailed lists are in the paid text. ATC's summary states the principle: the standard "defines the qualifications of professionals involved in the process," and its revision reinforces "that only qualified post-editors should be involved in MTPE."

The standard does not evaluate MT systems; ATC's explanation is that technology changes too fast for a standard to keep up with it. Nor does it set requirements for light post-editing, which it only defines and discusses in an annex.

Because it is a requirements standard, a provider can be certified against it. ATC names language service companies, freelance translators, and internal language departments.

How ISO 18587 relates to ISO 17100

ISO 17100:2015 is the requirements standard for translation services, and its abstract excludes post-editing explicitly: "The use of raw output from machine translation plus post-editing is outside the scope of ISO 17100:2015."

So an ISO 17100 certificate does not cover a provider's MTPE work. If a vendor offers you post-editing "under ISO," ask which standard. For post-editing, the relevant one is ISO 18587.

ISO 18587 is being revised

As of October 6, 2026, the 2017 edition is still the published one, and the revision is at the draft stage. The revision started in October 2023, and ATC's ISO committee expected the new edition in October 2026. According to national standards catalogs, the ballot on ISO/DIS 18587 opened on September 4, 2026, for 12 weeks. The German standards body DIN published the same draft as DIN ISO 18587:2026-10 and takes comments on it until November 11, 2026.

The draft's title is the first visible change: Translation services — Post-editing of non-human translation output — Requirements.

ATC's summary of the work in progress, written after a committee webinar in February 2025, lists what else was planned:

• New terms: "non-human translation," "artificial intelligence translation," and "automated post-editing." The committee wanted "non-human translation" to cover generative AI and LLMs, not just MT systems.
• Light and full post-editing both stay, but the level applied is "guided by the customer specifications."
• More weight on customer requirements, needs analysis, and risk assessment.
• A dedicated clause and annex on feasibility, to help project managers decide whether MTPE suits a given text.
• A structure aligned with ISO 17100, on the grounds that the process and the quality reached by human translation or MTPE "should be similar."

That summary is older than the draft now on ballot. It says the committee had decided to keep the title, and the draft changes it, so other points may have moved too. Check the published edition before you cite any of this in a contract.

How MTPE is priced

The usual model is a rate per source word, discounted from the translation rate. The research firm Nimdzi cites a TAUS survey that found it the most common, and a 2024 survey in Belgium found the same for agency work. Public data on actual rates is thin, so read the figures below as data points, not market rates.

The recent figures come from a survey of freelance language professionals in Belgium, run by researchers from the University of Geneva and UCLouvain with the Belgian Chamber of Translators and Interpreters and published in September 2025. It has 176 valid responses, collected between June and August 2024:

• 51% of respondents offer post-editing. For those who gave a figure, it is 27% of their workload on average (median 20%).
• For agency work, the source-word rate dominates: 63 respondents mention it, 9 mention an hourly rate. Asked which method fits post-editing best, 59 of 81 chose the hourly rate.
• Average agency word rates for post-editing are about 40% lower than the mean translation rate, which is about €0.10 per source word for agencies before CAT tool discounts.
• 21% of those who post-edit for agencies are satisfied with the income. It is the lowest figure in the survey.

These are numbers for one country. Among the concerns respondents raise about post-editing are CAT tool discounts they consider unfair and machine output of uneven quality, where poor output needs extensive rework "without advance warning or rate adjustment."

Nimdzi's article on post-editing rates (February 2020) draws a line between the levels. Discounts of up to 35–40% off the regular word rate "make sense" for light post-editing. For full post-editing, "sometimes it takes even more effort to process the MT output rather than translate from scratch." Nimdzi also calls edit distance, the count of changes made, a poor indicator of effort saved, because it "may not quite correlate with the time spent on editing."

As the side that produces drafts, we see a flat discount as a bet that the draft will be good. Before agreeing a rate, we would ask to see the raw output for a sample of the actual document. A post-editor can use the same sample to argue for an hourly rate.

When an AI draft is enough, and when it is not

In our view, an unedited AI draft is enough when the reader needs to understand a document and nobody will act on its exact wording. It is not enough when the text will be published, signed, filed, or followed as an instruction.

The 2017 edition has no checklist for this decision. According to ATC, the revision adds a feasibility clause for it. Until that is published, these are the four questions we would ask:

1. Who reads it? Internal reading for information tolerates what ISO calls a merely comprehensible text. A customer or a regulator does not.
2. What does one wrong term cost? A mistranslated torque value in a maintenance manual and a mistranslated greeting in an internal memo are different risks. Price the review accordingly.
3. Will someone be accountable for the text? If a named person or company must stand behind the translation, a qualified human has to read all of it.
4. Is the draft consistent enough to be worth editing? A post-editor fixes errors. Re-harmonizing terminology across a long document is rework on a different scale.

On the fourth question, the expensive errors are the ones spread across a document, like a term that changes from one page to the next, because every occurrence has to be found. We describe these failures in why machine translation breaks on technical documents, and our side-by-side examples against Google Translate show them on real text.

Where a HardTrans draft fits

HardTrans supplies the draft, and your linguist or vendor post-edits it in their own CAT tool.

Before translating, the system reads the whole file, or the whole job if there are several files, and then holds one term to one translation across all of it. Your glossary applies to every file in the job: upload a Smartcat XLSX export, CSV, TSV, or TXT. The system also builds its own project glossary and gives it to you as a separate CSV, which a post-editor can use as a term list to check first.

For post-editors the relevant format is XLIFF. Upload an XLIFF file from Trados, memoQ, Phrase, or Smartcat and you get the same file back with target filled in and tags in place. Open it in your CAT tool and edit there.

The limits are the ones a post-editor will hit first. We have no translation memory, no segment editor, and no QA checks, and we do not set segment status. Confirming segments happens in your tool. The system reads XLIFF 1.2 only; if your tool exports 2.0, export 1.2 or write to [email protected]. Scanned PDFs with no text layer are not supported.

Pricing is $1.50 per page of 250 words, with a 75% discount on repeated segments. The first document up to 10 pages is free, with no signup and no card. Upload a sample from the job you are about to quote and judge the draft before you agree a post-editing rate.

FAQ

Does ISO 17100 cover post-editing?

No. The abstract of ISO 17100:2015 states that raw machine translation output plus post-editing is outside its scope. Post-editing has its own standard, ISO 18587.

Is light post-editing covered by ISO 18587?

It is defined there and discussed in an annex, but the requirements of the 2017 edition cover full post-editing only. According to ATC, the revision keeps both levels and lets customer specifications decide which one applies.

Can you get certified to ISO 18587?

Yes. It is a requirements standard, so translation service providers can be certified against it: language service companies, freelance translators, and internal language departments, according to ATC. An ISO 17100 certificate does not cover post-editing.

Is MTPE cheaper than translation?

Per word, usually. In a 2024 survey of freelancers in Belgium, average agency word rates for post-editing were about 40% lower than the mean translation rate. The saving depends on the draft. Lee Densmer (2014) and Nimdzi (2020) both note that full post-editing can take more effort than translating from scratch.

Does HardTrans offer post-editing?

No. HardTrans is a self-service translator that produces the draft. Editing it, or hiring a post-editor for it, is up to you.