Insights
MTPE vs Human Translation: When Post-Editing Is Enough
By Asiatis · Published 24 August 2026 · Last updated 24 August 2026
Machine translation post-editing (MTPE) is machine output corrected by a qualified human linguist, defined by ISO 18587. It typically costs 40–70% of full human translation and suits high-volume, low-risk content such as internal documentation and support articles. It is not appropriate for contracts, regulatory submissions, marketing copy or anything that carries legal, financial or safety consequence.
Three things that are often confused
| Raw machine translation | MTPE | Human translation (TEP) | |
|---|---|---|---|
| Human involvement | None | Linguist corrects engine output | Linguist translates, second linguist revises |
| Governing standard | — | ISO 18587 | ISO 17100 |
| Relative cost | ~5% | 40–70% | 100% |
| Accountability | None | Post-editor and provider | Translator, reviser and provider |
| Suitable for publication | No | Depends on content type | Yes |
| Terminology control | None | Partial to full | Full |
The category error that costs money is treating these as points on a single quality slider. They are three different products with three different risk profiles.
What MTPE actually involves
Light post-editing corrects errors that change meaning: mistranslations, omissions, additions, wrong numbers. It does not aim for stylistic polish. The result is accurate and readable, and reads like what it is.
Full post-editing additionally aligns terminology with the client glossary, fixes register and style, and produces output intended to be indistinguishable from human translation. Full post-editing on difficult content often takes as long as translating from scratch — which is why the discount narrows, and sometimes disappears entirely, as content difficulty rises.
Why fluent output is the dangerous part
Neural and LLM-based engines produce grammatically clean, confident text. This is precisely what makes their errors expensive. Modern output produces a perfectly formed sentence that says something the source did not. Typical failure modes:
- Polarity flips. A negation dropped or added. “The supplier shall not be liable” becomes “shall be liable” and reads perfectly.
- Terminology drift. The same defined term rendered three different ways across a contract, each individually plausible.
- False confidence on ambiguity. A source ambiguity resolved silently in one direction, with no flag.
- Numbers and units. Decimal separators, thousands separators, dates and measurement units converted or preserved inconsistently.
- Cultural and legal non-equivalence. A common-law term mapped to a civil-law term that is not its equivalent.
- Invented plausibility. LLM-based engines can smooth over a gap in the source rather than reproduce it.
A reader who does not have the source cannot detect any of these. That is why post-editing has to be done against the source by a qualified linguist, not by proofreading the output on its own.
Can DeepL, Google Translate or ChatGPT do this?
| Google Translate | DeepL | General LLMs (ChatGPT, Claude, Gemini) | |
|---|---|---|---|
| Strength | Coverage, speed, free | Fluency in major European languages | Follows instructions, handles context and tone |
| Typical weakness | Literal renderings, weak on register | Narrower coverage, silent omissions on long input | Can smooth over or invent where the source is unclear |
| Terminology control | None in the free tool | Glossary in paid tiers | Only if supplied in the prompt, not guaranteed |
| Consistency across a document set | None | Partial | Varies between runs |
| Confidentiality | Free tiers not appropriate for confidential content | Enterprise tiers offer controls | Depends on the plan and its data terms |
| Audit trail | None | Limited | None by default |
Three properties disqualify all three from unsupervised use on consequential content: no self-assessment (none tells you which sentence it was unsure about), no stability (the same input can produce different output), and no accountability (if a translated warning is wrong, the manufacturer is liable, and under the EU Machinery Regulation there is no “original version” to fall back on).
What about using an LLM to check a translation?
It catches obvious omissions and some terminology inconsistencies, and it is genuinely useful as a first-pass screen. What it does not do is verify meaning against the source with authority — it will confidently approve a fluent mistranslation, and just as confidently flag a correct idiomatic rendering as wrong. Use it to triage, never to sign off.
Where MTPE works well, and where it should not be used
Works well: internal documentation and knowledge bases; support tickets; high-volume technical documentation with an enforced termbase; user-generated content; e-commerce catalogues at scale; gisting; repetitive documentation updated in small increments. Common characteristics: high volume, low individual consequence, repetitive language.
Should not be used: contracts and anything creating a legal obligation; regulatory submissions, clinical and patient-facing materials; financial statements and prospectuses; marketing and brand copy; safety instructions and warnings; patents; regulated product documentation (CE marking files, IFUs, safety information).
The test is simple: what does one undetected error cost? If the answer exceeds the saving from post-editing, the saving is not a saving.
The confidentiality question
Sending content to a public machine-translation service can mean transmitting it to a third party outside your control, potentially outside your jurisdiction, and depending on the terms of service, retaining it for model improvement. For content covered by PDPA obligations, an NDA, banking secrecy, legal privilege or a pre-publication embargo, the relevant questions are:
- Which engine is used, and is it a secure enterprise deployment or a public endpoint?
- Where is the content processed and stored, and for how long?
- Is the content used to train or improve the engine?
- Are the individual post-editors under NDA?
- Can machine translation be disabled entirely for designated content?
Any provider that cannot answer these in writing should not be handling confidential content, whatever its price.
Key facts
- MTPE is defined by ISO 18587; full human translation by ISO 17100.
- MTPE typically costs 40–70% of full human translation, and the discount narrows as content difficulty rises.
- Light post-editing corrects meaning only; full post-editing also aligns terminology and style.
- Neural and LLM output is fluent but not necessarily accurate, which makes errors harder to detect than in older systems.
- Post-editing must be performed against the source, not by proofreading the output alone.
- Public machine-translation engines may retain and reuse submitted content — a material issue for confidential material.
Related: the service-level comparison, how quality is measured, how to choose a translation agency.
Frequently asked questions
Is machine translation good enough for business documents now?+
For internal, low-consequence content with an experienced post-editor and an enforced glossary, often yes. For contracts, regulatory filings, financial reporting and published marketing, no. The limitation is not fluency, which is excellent, but reliability: the output is confident whether or not it is correct, and a reader without the source cannot tell the difference.
How much does MTPE actually save?+
Typically 30 to 60% against full human translation, but the saving depends heavily on content type and language pair. On repetitive technical documentation with a mature termbase, savings are at the top of that range. On dense legal or creative content, full post-editing can take longer than translating from scratch and produce no saving at all.
Can I tell whether a supplier used machine translation?+
Not reliably by reading the output, which is the point. What you can do is ask for the answer in writing, request the workflow logs or the CAT tool project package, and include a disclosure clause in the contract. Suppliers who post-edit openly and price it as such are usually the safer choice.
Does post-editing need a different skill set from translating?+
Yes, and it is not a lesser one. A post-editor must detect errors in text that reads well, resist being anchored by a plausible wrong suggestion, and decide quickly whether to fix or retranslate a segment. ISO 18587 sets out the competences required precisely because post-editing is a distinct discipline.
Is DeepL better than Google Translate for European languages?+
For the major European languages, DeepL is generally regarded as more fluent, particularly into German and French. That advantage is about readability rather than reliability: both produce confident output on ambiguous source text, neither flags what it was unsure about, and neither maintains terminology across a document set without a configured glossary. The choice of engine changes the amount of post-editing, not the need for it.
Can I just ask ChatGPT to translate my manual?+
For understanding a document, yes. For producing one you will publish, no. General-purpose models do not guarantee consistent rendering of the same term across a long document, do not enforce a termbase, produce different output on different runs, and carry no accountability for an error in a safety warning. For regulated documentation, the absence of an audit trail is disqualifying on its own.
Should I ever use a free online translator for business content?+
Only for understanding a document you have received, never for producing one you will send. Beyond accuracy, free services may retain submitted content and use it to improve their systems, which is incompatible with confidentiality obligations under an NDA, PDPA duties or legal privilege.