Skrivanek’s Arturs Peha on AI Translation, Human Review and Language Quality

Instant translation makes communication possible across language barriers, but access is not the same as understanding. Arturs Peha, chief strategy officer at language services company Skrivanek, sees neural machine translation (NMT) as a way to speed up work and open new kinds of communication. He also points to a harder question: what happens when people accept quick, imperfect output as good enough?

Peha’s view comes from working on technology and service design at Skrivanek, including its NMT engine and Language Knowledge Assessment Tool. His account shows how providers are combining machine output with human review, adapting their services to new tools, and weighing the practical benefits against effects on workers, data and language diversity.

Translation work is changing shape

AI and NMT are altering what language specialists spend their time doing. Automated systems can take on repetitive translation tasks, while people check, correct and assess the resulting text. That can create room for work requiring subject knowledge, cultural awareness and judgment, but Peha cautions that a change in tasks doesn’t guarantee higher earnings for translators.

“AI tools boost efficiency and take over routine tasks, yet the insight of humans is vital for addressing the subtleties, cultural nuances, and specific areas of knowledge.”

Post-editing machine-generated text has become a familiar part of the discussion across fields such as marketing, IT and legal services. For translators, the shift can also feel like a threat to their livelihoods. Peha acknowledges that the industry won’t retain every existing role. Some people who relied on freelance assignments are looking for permanent positions to gain more stability.

That transition makes staffing and skills important parts of adopting translation systems. A provider needs people who can judge the output, but finding qualified post-editors can be difficult for some language pairs. The question for a project manager is not simply whether a tool can translate a file. It’s whether the workflow fits the content, the client’s requirements and the available expertise.

Fast access comes with a quality tradeoff

For Peha, one promising use is interpreting mixed-language speech. Someone might switch between Polish and English within a conversation, and a machine interpreting system could recognize the combination and produce a transcript or instant translation in another language. That kind of support could make everyday communication easier, especially when people need to understand the gist of a conversation without knowing every language involved.

But the spread of instant tools may also lower expectations for language quality. Peha worries that people are becoming more tolerant of rough translations, with potential consequences for smaller languages. Many systems use a widely represented language as an intermediary when translating between less common ones. That process can flatten distinctive expressions and cultural features, making languages more alike in the output.

The reach of professional language services is limited compared with the volume of communication happening around the world, Peha argues. Self-service translation can help fill that gap, but choosing the wrong tool for a task can create problems in personal and professional settings. Translation systems may convey a basic message while missing the cultural references, tone or context that shape what the message means.

Peha describes better translation technology as a broad benefit for people who want to communicate in languages they don’t speak and are unlikely to learn. Still, machine-mediated conversation won’t necessarily provide the depth of communication between two native speakers. Cultural nuance is especially easy to lose. Businesses also approach providers for help when machine output doesn’t meet expectations, including for marketing materials where wording needs to fit a brand and resonate with its audience.

Instant translation can change how people shop, learn and travel. Machine translation already helps make e-commerce content available to customers in other languages. Translation tools can also open access to educational material that might otherwise be inaccessible, and help travelers engage with local experiences. In personal communication, they allow people to attempt conversations that would have been difficult without a dictionary or a fluent speaker nearby.

The stakes vary by use. A rough translation may be sufficient to get a product page online or understand the basic meaning of a message. A mistranslated marketing email, however, could have a financial impact. The practical distinction is not whether machine translation is useful in general, but whether its likely errors are acceptable for a particular audience and purpose.

How Skrivanek applies machine translation

Skrivanek began exploring machine translation when statistical systems were common, according to Peha. Its research and development team now works on internal tools and ways to adapt newer technologies to language services. One result is the company’s NMT engine, developed for clients in fields including medicine, law and technology. The company has also built a Language Knowledge Assessment Tool, which Peha says is used by IT companies and human resources teams.

Skrivanek is also exploring generative AI for routine work and possible client services. Its teams offer editing and rapid review of AI-generated text, video, audio and graphics, whether the company helped produce the material or a client created it independently.

Peha describes NMT as a way to reduce the time and effort spent on repetitive work, potentially lowering costs and letting specialists work more independently. It can also free translators to concentrate on idioms, cultural context and stylistic choices. That division of labor can be useful when companies adapt marketing materials for audiences with different expectations.

Two examples illustrate how the service can be tailored to a deadline or budget. An e-commerce company needed to launch a product quickly. Skrivanek first used unedited machine translation to put the content online, allowing search engines to begin indexing it. The provider then post-edited the text to improve its quality and match the company’s brand.

In another project, the client needed four million words translated within a few days and had a limited budget. The proposed workflow used NMT for the full project. Peha says the client was satisfied because the result met its business needs, while making clear that he didn’t consider the machine output flawless.

“Did we personally think that the quality of this machine translation was flawless? Of course not. But it clearly corresponded to client expectations.”

That example captures the central tradeoff: quality is judged against what the work needs to accomplish, not against an abstract ideal. Modern NMT and project-management systems can also include quality-assurance tools to help teams check output. Even so, some assignments aren’t appropriate for machine translation. Clients may rule it out through their own agreements or internal requirements, so providers need alternative workflows.

Tool choice, review and limits

Machine translation still has trouble with specialist terminology, idioms and cultural context. Skrivanek addresses those weaknesses by using machine output alongside computer-assisted translation software and human review. Translators with knowledge of the subject area can correct errors and decide whether a phrase makes sense in context.

Industry-specific terminology databases and more relevant training material can improve how systems handle complex terms. Regular quality checks and feedback loops can help teams identify recurring errors and adjust their processes. These measures support better output, but they don’t remove the need to assess each project on its own terms.

Skrivanek combines its own engine with services such as Google Translate, DeepL and Bing, according to Peha. Human language specialists may also use generative tools such as ChatGPT to clarify source passages, summarize material, adapt documents to formal requirements, or adjust tone and style. People remain responsible for checking whether the result is accurate and suitable for its audience.

Clients can impose additional constraints. Some don’t permit post-editing, sometimes because of clear project requirements and sometimes because of earlier poor experiences. Meanwhile, providers face the broader task of introducing generative AI into daily work. That involves deciding how the tools fit existing workflows and what employees need to know to use them well.

Data and language ethics

Translation projects can contain medical records, company secrets, banking details or ordinary marketing material. Each category raises different questions about how information is handled. Peha says providers need clear rules for when and how AI tools may process client content, as well as an understanding of what each client considers confidential.

Skrivanek says it trains staff and external partners to use AI tools responsibly, with training tied to its work toward ISO 27001 and ISO 9001 standards. For projects, the company chooses closed commercial or proprietary NMT systems and agrees not to process sensitive information through AI or NMT systems. It also works with clients to define confidential data, sign relevant agreements and follow the workflows both sides have established.

Bias is another concern. Translation systems learn from human-created material, which can contain prejudice. That bias may shape the translations a system produces. Smaller languages face a separate risk: when an NMT system uses a more common language as an intermediate step, distinctive meanings and cultural features may be lost along the way.

Peha also identifies worker displacement, misuse of AI and disagreements over intellectual property as issues that need attention. Together, these concerns make tool selection and human oversight part of responsible delivery, not merely steps to polish a translation after the fact.

What providers and clients may need to adapt

Peha expects AI use for specific tasks to become commonplace over the next five years. He says large companies will tend to push their teams toward adoption, while small and medium-sized businesses can sometimes move faster. Pay-as-you-go access may make tools easier to adopt without a large upfront commitment.

He also expects more people to build specialized services by adapting large language models to narrow problems. The same period may bring new jobs and make it easier for people to produce digital work such as games and videos. Peha also warns of increased misinformation, radicalization, and challenges in protecting privacy and financial information. The effects, he says, will depend in part on each country’s economy and openness to change.

The language-services market is already seeing companies expand beyond their original roles. Some IT businesses and call centers are adding translation as a secondary service. Language providers are taking on IT work such as system testing and setup, alongside social media management. Public relations firms are broadening their offerings as well. Peha expects more providers to combine services and greater use of neural machine interpreting, with speed and convenience encouraging acceptance of average-quality results.

He cautions that companies that don’t adapt may struggle, and predicts that market consolidation could follow, with some smaller providers losing value as larger companies grow. His broader point is that AI adoption affects how firms organize and sell their services as well as how they produce translations.

A practical approach to adoption

Peha’s advice to businesses starts with defining the task. Not all translation work has the same requirements, and AI may suit some assignments better than others. For content where precision, specialist knowledge or a personal connection matters, review by experienced people remains important.

He also advises teams to keep learning and checking how their translations perform. Tools change, and a workflow that suits a project today may no longer be the best option later. Customer feedback can surface phrases that are confusing or feel wrong for the audience, providing a useful signal for improving future work.

The operational question is how to combine machine speed with human judgment. Used selectively, NMT can help businesses publish sooner, manage large volumes and reach people across language barriers. Its limits become more visible when an error carries a high cost, a phrase depends on cultural context, or the language involved has little representation in the system’s data. For clients and providers alike, choosing the right workflow—and keeping people accountable for the final result—will shape how much value instant translation delivers.