Rethinking productivity, quality and the human factor in AI-assisted translation
Translation technology has made remarkable progress. Translation memories, machine translation, AI-assisted tools and increasingly sophisticated workflows promise faster production, greater consistency and lower costs.
And in many situations, these technologies genuinely help.
Yet a question deserves more attention: Does technological progress always translate into greater efficiency in the real working lives of professional translators?
For many language professionals, the answer is more complicated than productivity figures might suggest.
The challenge is not simply whether a tool can produce a suggestion faster. It is whether the entire workflow enables translators to deliver the right result, at the required level of quality, within a realistic amount of time.
The productivity paradox
In translation, productivity is often associated with the number of words processed within a given period. Translation memories offer matches, machine translation generates drafts, and AI tools suggest terminology, phrasing and revisions.
These capabilities can reduce the time needed for certain tasks. Research into machine translation post-editing and translation-memory use has documented productivity benefits in specific contexts, while also highlighting the importance of output quality and the effort required to correct it.
For example, Guerberof Arenas examined productivity and quality in a professional localisation project. The findings suggested that, under the conditions studied, translators could achieve higher productivity when working with machine-translated output than when translating without it. The results also showed that the relationship between productivity and quality depends on the type of assistance provided.
More recently, a real-world case study by Martikainen et al. (2025) examined the implementation of GPT-4 in an in-house translation team in the Finnish social and healthcare sector. Post-editing was, on average, 14% faster than translating from scratch. However, individual productivity changes ranged from a 2% decrease to an improvement of 102%.
The study also identified challenges involving inconsistent terminology, insufficient document-level context and limited system customisation.
These findings illustrate an important distinction: measurable productivity gains are possible, but they are not necessarily consistent across translators, texts or working environments.
The difficulty arises when the speed of generating or processing text is treated as a reliable measure of the efficiency of the entire translation process.
A suggestion may appear instantly. Evaluating whether it is accurate, appropriate and consistent with the client’s requirements can take considerably longer.
The time saved at one stage of the process may create additional work at another.
This does not mean that translation technology is ineffective. It means that productivity needs to be assessed across the complete workflow, rather than through a single metric.
The hidden work behind a match
Translation memories are a familiar example.
A high match percentage can be useful, but it does not automatically establish that a segment is suitable in its current context. A sentence may be identical to one stored previously while the surrounding text, intended meaning, terminology or communicative purpose has changed.
Even a 100% match may therefore require verification.
Research by Screen examined the effects of Translation Memory fuzzy matches on the translation process. The study found that correcting and adapting these matches could increase productivity, but also involved longer initial reading times and more self-revision than manual translation.
This is an important distinction: a match can save time in one respect while changing the nature and distribution of the work involved.
The same principle applies to machine-translated text and AI-generated suggestions. Fluent language can create an impression of reliability, but fluency alone does not establish factual accuracy, contextual appropriateness or compliance with client-specific terminology.
The translator’s task is not simply to accept or correct a suggestion. It is to determine whether the suggestion should be used at all.
This judgement is an essential part of professional translation. Yet it can be difficult to capture in productivity measurements that focus primarily on word counts, match rates or processing speed.
Research into post-editing also demonstrates that the effort involved depends on the quality of the output and the nature of the task. In her review of the research, Koponen highlights the different dimensions of post-editing effort and the conditions that influence whether machine translation actually reduces the work required.
More recent research by Sarti et al. (2025) examined how word-level quality estimation tools influence professional post-editing. In a study involving 42 professional post-editors across two translation directions, the researchers found that the usefulness of automatically highlighted potential errors depended on factors such as domain, language and the speed of individual editors.
Their findings point to a relevant gap between the technical accuracy of an automated tool and its practical usefulness in a professional workflow.
When reference material becomes a bottleneck
Another challenge lies in the information surrounding a translation project.
Clients may provide extensive reference documents, terminology lists, previous translations, style guides and project-specific instructions. These resources are valuable: they help establish consistency and clarify expectations.
However, their usefulness depends partly on how accessible they are within the actual workflow.
If relevant information is scattered across files, platforms, email threads or lengthy reference packages, translators may have to interrupt their work to locate the right passage, confirm a term or determine which instruction takes precedence.
The problem is not necessarily a lack of information. It may be the difficulty of retrieving the right information at the right moment.
These interruptions can turn an apparently streamlined process into a sequence of small but time-consuming research and verification tasks.
Better integration of reference material, searchable documentation, clearer terminology management and accessible client instructions could therefore improve efficiency as much as faster text generation.
The human factor is not an optional extra
The more automated a workflow becomes, the more important it is to define what human review is expected to accomplish.
Post-editing is not a single, uniform activity. The required level of intervention depends on the purpose of the translation, the quality of the machine output, the subject matter and the quality standard agreed with the client.
A text intended for internal understanding may require a different level of revision from a text intended for publication, legal use or communication with customers.
Research has shown that post-editing can improve productivity in particular circumstances, but that poor output may require substantial effort.
A 2025 study by Sun, Wang and Jia (2025) compared post-editing and human translation in Chinese-English and English-Chinese translation. The researchers found that post-editing generally offered advantages in speed and output quality in their experimental setting. However, these advantages differed depending on translation direction. The study highlights why productivity assumptions should take language direction and working conditions into account.
The way translators interact with technology also matters. Briva-Iglesias, O’Brien and Cowan (2023) compared traditional and interactive post-editing in a study involving professional translators. Their findings suggested that translators preferred interactive post-editing, partly because it gave them greater control over their interaction with machine translation. The study also found productivity benefits for the interactive approach under the conditions examined.
This matters because efficiency cannot be understood solely as speed. It also involves the effort required, the reliability of the result and the conditions under which professionals can exercise their judgement.
The human contribution should not be reduced to a final check of text that has already been assumed to be correct. Professional review is part of how the quality of the result is established.
When deadlines and expectations fail to adapt
Technology can make it possible to process larger volumes of text. But when increased technical capacity automatically leads to shorter deadlines, higher output expectations or reduced budgets, the benefits may not be distributed evenly.
The time needed to verify terminology, investigate ambiguity, consult references and review potentially misleading suggestions does not disappear simply because a tool generates text more quickly.
If the time allocated to these tasks is insufficient, the risk is that important checks are missed or performed under unnecessary pressure.
This is not an argument against efficiency or innovation. It is a reason to distinguish between a tool’s technical capacity and the realistic capacity of a complete quality-controlled workflow.
A process that produces more words but leaves too little time for appropriate verification may look productive while failing to achieve the intended quality consistently.
Rethinking what we measure
If the industry wants to understand whether translation technology is improving productivity, it needs to look beyond the number of words processed.
A more complete assessment could consider:
- Total effort: How much time is spent on translation, post-editing, research, terminology checks and final review?
- Quality of the result: Does the output meet the agreed requirements and serve its intended purpose?
- Information accessibility: Can translators find the right reference material and instructions without unnecessary interruptions?
- Workflow design: Are tools, deadlines and quality requirements aligned with the actual work involved?
- Professional experience: Do translators have sufficient time, control and information to make sound decisions?
These questions do not diminish the value of technology. They help identify where it creates genuine benefits and where the surrounding process needs to improve.
They can also help clients, language service providers, project managers and technology developers make better-informed decisions about how translation work is organised.
Technology should improve the process, not just accelerate the output
The future of professional translation will undoubtedly involve increasingly capable AI systems and more integrated technological workflows.
The important question is how these developments are implemented.
Technology can support translators by reducing repetitive work, improving access to terminology, helping identify inconsistencies and making relevant information easier to retrieve. But these benefits depend on thoughtful implementation, suitable quality criteria and realistic expectations.
The European Commission’s Directorate-General for Translation provides an interesting example of how technological development and professional language services can coexist. Its work combines high-quality translation, editing services and AI-based multilingual services such as eTranslation. The Commission’s description of the DGT’s responsibilities illustrates that technology is one part of a broader language-service infrastructure.
The objective should not be to make translators work faster at any cost. It should be to create workflows in which technology genuinely reduces unnecessary effort while preserving the time and judgement needed to produce reliable translations.
That requires more than better tools. It requires a better understanding of the work those tools are intended to support.
A question worth discussing
The debate about AI and translation often focuses on what technology can now do. We should also examine what happens to the work around it.
Are we measuring productivity in a way that reflects the complete translation process? Are reference materials and instructions accessible when they are needed? Do deadlines account for the verification that quality requires?
And perhaps most importantly:
When translation technology becomes faster, are our workflows becoming more efficient, or are we simply moving the work somewhere else?
These questions deserve a constructive, evidence-based discussion involving translators, project managers, language service providers, clients and technology developers.
Because technological progress is most valuable when it improves not only the speed of production, but also the quality, reliability and sustainability of the work.
Selected references
The following sources are listed in order of their relevance to the article’s central themes: real-world AI productivity, post-editing effort, Translation Memory, human interaction and professional translation quality.
- Martikainen, M., Smolander, K., Sanmark, J., & Sanmark, E. (2025). “Evaluation of Generative Artificial Intelligence Implementation Impacts in Social and Health Care Language Translation: Mixed Methods Case Study.” JMIR Formative Research, 9, e73658.
https://doi.org/10.2196/73658 - Sarti, G., Zouhar, V., Chrupała, G., Guerberof-Arenas, A., Nissim, M., & Bisazza, A. (2025). “QE4PE: Word-level Quality Estimation for Human Post-Editing.” Transactions of the Association for Computational Linguistics, 13, 1410–1435.
https://doi.org/10.1162/tacl.a.46 - Sun, S., Wang, H., & Jia, Y. (2025). “Direction matters: Comparing post-editing and human translation effort and quality.” PLOS ONE, 20(7), e0328511.
https://doi.org/10.1371/journal.pone.0328511 - Guerberof Arenas, A. (2014). “Correlations between productivity and quality when post-editing in a professional context.” Machine Translation, 28(3–4), 165–186.
https://doi.org/10.1007/s10590-014-9155-y - Koponen, M. (2016). “Is machine translation post-editing worth the effort? A survey of research into post-editing and effort.” The Journal of Specialised Translation, 25, 131–148.
https://doi.org/10.26034/cm.jostrans.2016.303 - Screen, B. A. (2016). “What does Translation Memory do to translation? The effect of Translation Memory output on specific aspects of the translation process.” Translation & Interpreting, 8(1), 1–18.
https://doi.org/10.12807/ti.108201.2016.a01 - Briva-Iglesias, V., O’Brien, S., & Cowan, B. R. (2023). “The impact of traditional and interactive post-editing on Machine Translation User Experience, quality, and productivity.” Translation, Cognition & Behavior, 6(1), 60–86.
https://doi.org/10.1075/tcb.00077.bri - European Commission, Directorate-General for Translation. “Translation.”
https://commission.europa.eu/about/departments-and-executive-agencies/translation_en
