Skip to content

The Hidden Risk of Post-editing

AI-assisted translations have become an increasingly important part of the language services industry. As post-editing becomes more common, many translation projects that would once have been assigned to translators as human translation are now delivered as AI draft plus expert post-editing (MTPE).

By Toprak Deniz Odabaşı, Translation Project Coordinator at Alafranga Language Solutions

I am not writing this from the outside. Post-editing has been a named workflow in our own project management system since 2017, running across DE>EN, EN>DE, EN>TR, TR>EN, IT>TR and DE>TR. We buy it and we sell it. The observations below come from watching the same discount applied to jobs that turned out to be nothing alike.

The 50% assumption

The underlying assumption is understandable. If a AI translation system produces a useful first draft, the translator should have less work to do. In some cases, this is exactly what happens. A good AI translation can save a translator considerable time, particularly when the source text is straightforward and the system handles the terminology and style well.

The problem is that this is far from consistent. The amount of work required to post-edit a AI-generated translation can vary dramatically from one project to another, and sometimes even from one segment to the next. A translator may occasionally be able to review and lightly correct an output in a matter of seconds. In other cases, the output may require substantial restructuring, terminology research, contextual interpretation and rewriting before it reaches the quality expected from a professional translation.
 
Despite this variation, MTPE rates are often set at around half the rate of human translation, or even lower. The amount of work required, however, does not necessarily decrease by the same proportion. A project priced at 50% of a human translation rate does not automatically require 50% of the effort.
 
This creates a difficult economic equation for translators. When the rate offered for a project is significantly lower, spending the same amount of time on each segment as they would during a human translation project may no longer be commercially viable. The translator is therefore more likely to focus on correcting the AI output efficiently rather than approaching every sentence with the same level of attention they would give to a translation produced from scratch.
 
This does not mean that translators are deliberately delivering lower-quality work. It is a predictable consequence of the relationship between compensation, time and workload. If a task is priced on the assumption that it requires substantially less effort, but the actual effort required remains high, there is a clear incentive to reduce the time spent on it.
 
The result can be a quality gap that is difficult to explain simply by looking at the AI translation output itself. A AI-generated draft may appear promising, yet the final quality of the MTPE process depends heavily on how much time and attention the post-editor can reasonably devote to it.
Post-editing is still professional translation work 
One of the challenges facing the industry is the tendency to treat post-editing as inherently easier than human translation. In reality, post-editing requires many of the same linguistic skills as translation. The translator still needs to understand the source text accurately, identify ambiguities, make appropriate terminology choices, preserve meaning and register, and produce natural target-language text.

In some cases, the translator also has to determine whether the AI translation has misunderstood the source in a way that is not immediately obvious. A grammatically correct sentence can still be inaccurate, misleading or inappropriate in context. Detecting such problems requires the same kind of linguistic judgment that is essential in human translation.
 
The technology itself can also introduce additional complications. AI translation systems do not always sit cleanly inside established translation workflows. There can be friction between AI-generated content, CAT tools and translation memories, while segmentation, formatting, terminology and context can create additional work during post-editing. A system that produces a strong translation in isolation may therefore still require considerable effort to use efficiently within a professional translation environment.
 
For this reason, measuring post-editing effort solely by the quality of the initial AI output can be misleading. The full workflow needs to be considered.
When MTPE creates more work
There is another issue that deserves attention: some types of translation are still poorly suited to an MTPE model. Subtitling is a good example. AI translation may produce a technically understandable rendering of the dialogue, yet professional subtitling involves much more than transferring meaning from one language to another. Timing, character limits, segmentation, reading speed, dialogue flow and the conventions of the target language all need to be considered.

The clearest case in our own records is a 2024 EN>TR subtitle post-editing project. The machine output carried the dialogue accurately enough. It carried none of the things that make a subtitle work: reading speed, line breaks that fall where the sense falls, character limits, timing against the shot. Every one of those had to be built from zero, on a rate set on the assumption that the translation was already half done.
 
When the initial machine output is inadequate, the post-editor may therefore have to spend time identifying what is wrong with the suggestion before producing a suitable translation. In some cases, the fastest route is to disregard the machine output and translate the segment from scratch. The translator is still being paid at an MTPE rate, even though the task has required the kind of linguistic work associated with human translation.
 
This can make the process particularly inefficient. Instead of starting with a blank segment and translating it according to the source and the requirements of the target language, the translator first has to assess an unreliable suggestion, determine which parts can be used, identify what needs to be replaced and then reconstruct the translation. When the machine output is fundamentally unsuitable, this additional evaluation can take more time than simply translating the content independently.
 
The same principle can apply to other content types where meaning depends heavily on context, style, creativity or highly specific conventions. In such cases, the presence of an AI-generated suggestion should not automatically be treated as evidence that the translator's workload has been reduced.
 
This is an important distinction when setting MTPE rates. A machine-generated draft only creates an efficiency gain when it actually reduces the amount of work required from the translator. If the post-editor has to spend significant time assessing, correcting or replacing the output, the workflow may provide little practical benefit while still placing the project in a lower MTPE rate category.
A more realistic approach to MTPE rates
The industry needs to move towards a more realistic understanding of what post-editing involves. MTPE can certainly reduce translation costs and turnaround times when the technology performs well. However, the savings should not automatically be assumed to come entirely from the translator's compensation.

Scale does not resolve this, it multiplies it. In 2025 we post-edited one German technical documentation set into eleven target languages at once: Bulgarian, Croatian, Czech, Turkish, Latvian, Polish, Slovak, Slovenian, Estonian, Romanian and Hungarian. The engine performed well in some of those directions and poorly in others. The rate category was identical across all eleven. That is not a pricing model, it is an average applied to work that has no average.
 
If post-editors are expected to deliver the same level of linguistic quality as professional human translators, their compensation needs to reflect the actual work involved. This does not necessarily mean that every MTPE project should be paid at the same rate as human translation. The appropriate rate will depend on the quality of the machine output, the subject matter, the required quality level, the workflow and the amount of post-editing required.
 
What matters is that the rate should correspond reasonably to the work that the translator is expected to perform.
 
There is a standard for this. ISO 18587:2017 covers the full post-editing of machine translation output and sets out what a post-editor is expected to be competent in. It is worth noting because ISO 17100:2015, the one most buyers ask about, covers human translation with independent revision and explicitly does not cover machine-drafted content. When a supplier quotes ISO 17100 against an MTPE line item, that is worth a question.
 
There is also work where the question does not arise at all. Content that contractually cannot reach a third-party AI provider is translated by people from the start, and priced on that basis rather than discounted against a draft that was never produced.
 
A sustainable MTPE model therefore requires a better balance between technology, pricing and quality. When the technology genuinely reduces the translator's workload, the efficiency gains can benefit everyone involved. When extensive human intervention is still required, the economics of human expertise need to be recognised accordingly. I have argued elsewhere that every updated manual is a retranslation project. The pricing problem described here is the same one, seen from the supplier's side.
 
As AI-assisted translation continues to develop, the industry's understanding of post-editing should develop with it. The question is no longer simply how much a machine can translate. It is also how much professional work remains after the machine has finished, and whether that work is being valued accordingly.
 
Frequently asked questions
 
Why are MTPE rates usually set at about half the human translation rate?
The 50% figure is a market convention rather than a measurement. It assumes a usable machine draft removes roughly half the translator's work. That holds when the source is straightforward and the engine handles terminology and register well. It does not hold when the output has to be restructured, researched or rewritten, and the rate is usually fixed before anyone has seen how the engine performs on the actual content.
 
Does post-editing always take less time than human translation?
No. Effort varies between projects and between segments within the same project. A segment may need a few seconds of correction, or terminology research, contextual interpretation and full rewriting. A grammatically correct machine sentence can still be inaccurate or wrong in register, and detecting that requires the same linguistic judgment as translating from scratch.
 
Is machine translation post-editing suitable for subtitling?
Often not. Subtitling depends on timing, character limits, segmentation, reading speed and dialogue flow, none of which a raw machine output addresses. When the suggestion is unusable, the post-editor has to assess it, decide what can be kept and reconstruct the subtitle, which can take longer than writing it directly.
 
Does ISO 17100 cover machine translation post-editing?
No. ISO 17100:2015 covers human translation with independent revision and excludes raw machine translation output. The standard that addresses post-editing is ISO 18587:2017. The two should not be presented as interchangeable in supplier documentation.
 
What should a fair MTPE rate depend on?
The quality of the machine output on that specific content, the subject matter, the required quality level, the workflow and the amount of editing actually needed. The practical test is whether the discount applied to the rate matches the reduction in work, and that can only be established by looking at the content rather than the workflow label.