Reinventing Translation: Ethics, AI, and the Future of Meaning

The Future of Translation Studies, Agencies, and Meaning

“To translate is not to repeat — it is to dream in another rhythm.” — From my studies on translation of Lermontov

From Reflection to Reinvention

Two decades ago, I stood in a quiet classroom, comparing translations of Lermontov's I Walk Alone on the Road. My aim was purely poetic: to trace how rhythm, silence, and emotion survive migration from one language to another. Today, that same question — how meaning travels — has expanded beyond literature into technology.

In 2026, translation is no longer merely a bridge; it has become a crossroads. As AI models learn to mirror not only language but also tone and nuance, the translator's task is itself being "translated": from craft to curation, from words to architectures of meaning.

Translation of culture in its most visible form — color, ritual, and rhythm.

Where We Stand: The 2025 Language Industry

According to Nimdzi Insights, the global language services market reached approximately USD 72.6 billion in 2025 and is projected to grow to around USD 73.4 billion in 2026, with Nimdzi forecasting only linear growth of less than 1% annually in the near future. Other research houses are more optimistic — Fortune Business Insights values the market at USD 76.23 billion in 2025 and projects a 7.6% compound annual growth rate through 2034, while Mordor Intelligence sets the 2025 figure at USD 71.82 billion with an expected 5.4% CAGR through 2030. Whichever forecast one adopts, the central message is the same: growth is no longer driven by volume, but by complexity.

In its 2026 report, Nimdzi highlights two connected dynamics.

• First, many language service providers have significantly downsized in-house linguistic and project management teams — sometimes by 20 to 25 percent — to offset inflation and absorb the roughly threefold productivity gains enabled by AI.

• Second, 2026 is expected to be marked by a surge in "agentic" solutions that move beyond static workflows toward autonomous decision-making and multimodal processing, forcing traditional translation management systems (TMS) to evolve into unified data platforms.

Machine translation post-editing (MTPE) has become the industry baseline, while a new premium tier is forming at the top of the value chain — strategic linguistic consulting, language governance, and quality assurance for generative content. Quality evaluation metrics such as COMET, MetricX, and XCOMET now deliver near-instant feedback on formal accuracy, yet they remain "blind" to cultural subtext. This is precisely where the human role moves up the value chain: from typing words to managing intentionality.

Switzerland as a Case Study: Multilingualism as Infrastructure

If you want to understand where the future of translation is being quietly shaped, Switzerland is a telling example — a country where language is not merely a medium of expression, but a framework of coexistence.

According to the Federal Statistical Office (FSO) and its 2019 Language, Religion, and Culture Survey, more than two-thirds of the population aged 15 and over regularly use more than one language.

The most recent figures indicate that roughly 63% of Swiss residents are multilingual — 37% speak two languages, and 26% speak three or more. Multilingualism declines with age: among young people (15–24), 81% are multilingual, while among those over 65 the share drops to 38%.

FSO data also show that approximately 86% of residents consider knowledge of several national languages important for national cohesion. At the federal level, German, French, and Italian are official languages, while Romansh is additionally used for official purposes in Romansh-speaking areas of the canton of Graubünden. Four cantons (Bern, Fribourg, Valais, and Graubünden) are officially multilingual, and the city of Biel/Bienne is officially bilingual. All official federal documents — legislation, reports, websites, brochures — must be provided in German, French, and Italian, which places translation at the very core of how the state operates.

Alongside the national languages, English plays an increasingly prominent role as a professional language, especially in science, finance, and international communication. In the workplace, FSO data indicate that roughly 38% of employees regularly use more than one language at work (or around 26% if Swiss German and standard German are counted as one language).

The Swiss model demonstrates that in a multilingual democracy, translation is the infrastructure of inclusion and the architecture of comprehension — and, increasingly, the interface between human precision and machine efficiency.

Engineering as a universal language — beauty in structure, rhythm in motion.

The Standards that Define Value

• ISO 17100 defines translation processes, translator qualifications, and mandatory "four eyes" revision.

• ISO 18587:2017 regulates the post-editing of machine translation output and requires that the final result be comparable in quality to human translation. It is important to note that the currently valid version is still the 2017 edition; a revision is underway within ISO TC 37/SC 5, and project leader Eva-Maria Tillmann indicated in 2025 that the revised standard will, among other things, more clearly distinguish between machine translation, generative AI output, and post-editing tasks, and will align more closely with ISO 17100.

• ISO 30042:2019 is the standard for managing terminology resources (TermBase eXchange, TBX) and describes the metamodel, data categories, and XML styles. A revision is already in preparation (ISO/AWI 30042).

• ISO 24495-1:2023 is the first international standard for plain language and sets out four governing principles: relevance, findability, understanding, and usability.

  
  

  In a world flooded with instant text, traceability is becoming a form of prestige.

Technology in 2026: The New Toolbox

Generative large language models (LLMs) have substantially reshaped the workflow over 2024 and 2025. Long-context models (e.g., Gemini 2.5 and its successors) make it possible to maintain consistency across documents spanning thousands of pages. Real-time voice translation (DeepL Voice, Gemini Live) has moved beyond demonstration and opens up a new role: the linguistic pilot — a human expert who supervises live translated content and intervenes in cases of hallucination or cultural misfire.

The evaluation landscape is evolving just as quickly.

MetricX is currently at version MetricX-25, presented at WMT 2025 and built on the Gemma 3 model. XCOMET combines sentence-level scoring with the detection of errors at the span level; analyses from WMT24 show that COMET-22 achieves a system-level correlation with human MQM ratings of around 0.69, while XCOMET reaches approximately 0.72. As one of the co-authors of COMET (R. Rei) has acknowledged, however, no single metric is sufficient for complex scenarios; practice is shifting toward combinations of COMET, MetricX, CHRF, and LLM-as-a-judge approaches.

Terminology has become "data gold". Instead of simple glossaries, organizations now build knowledge graphs that feed corporate retrieval-augmented generation (RAG) systems and ensure a consistent voice even when the respondent is an agent.

Technology does not replace translators — it redefines them. The translator of the future is a meaning engineer who combines linguistic intuition with metric literacy.

The language of images — where every translation begins with seeing.

How Translation Schools Can Reinvent Themselves

From Translator to Language Operations Specialist

When I translated my first book from Russian at nineteen, I understood translation as pure art — a dialogue between two souls across languages. Since then, I have translated across disciplines: from literature to law, from cultural essays to corporate reports. At twenty-five, I led translator training at Slovenia's second-largest translation company, which opened the door to industry standards, the first CAT tools, and the earliest serious debates about machine translation. That journey has taught me that, at its best, translation is a system of meaning management — not a solitary act of word transfer.

And yet many translation studies programs still teach as if the world had stopped in 2005. While some universities have embraced technology, others continue to treat translation as a static craft rather than a dynamic ecosystem. As AI systems, localization workflows, and multilingual data pipelines reshape communication, the question for the academy is no longer how to preserve translation, but how to evolve it.

I propose eight content modules for a one- or two-year graduate program:

• Algorithmic hermeneutics and AI interpretation. Students learn to analyze why a model selected a specific cultural framing and how to steer it through advanced prompt design for nuanced linguistic output.

• Post-editing and quality evaluation. In-depth training under ISO 18587 and mastery of modern metrics (COMET, MetricX, XCOMET, CHRF), complemented by LLM-as-a-judge approaches.

• Terminology and ontology design. Building knowledge graphs and multilingual ontologies (aligned with ISO 30042) that connect to RAG systems.

• Digital sovereignty and cultural revitalization. A specialized module for mid-sized and smaller languages (such as Slovenian) — how to prevent the language flattening caused by dominant AI models.

• Regulated contexts and accessibility. Plain language (ISO 24495-1:2023), compliance with the European Accessibility Act (EAA, applicable since 28 June 2025), and the relevant provisions of the EU AI Act (Regulation (EU) 2024/1689), whose obligations for general-purpose AI models entered into force on 2 August 2025, with the bulk of rules for high-risk systems applying from 2 August 2026.

• Audiovisual AI and real-time orchestration. Managing live voice-to-voice translation (DeepL Voice, Gemini Live), hallucination monitoring, and terminological consistency.

• Transcreation and brand voice curation. Cultural adaptation for the digital economy, UX microcopy, and tone-of-voice manuals that guide both human teams and AI agents.

• Professional standards, ethics, and data stewardship. Applying ISO 17100 and 18587, GDPR, and intellectual property in the age of generative AI; which data may legitimately be used to train models and how to safeguard client confidentiality.

The program should culminate in a twelve-week Language Operations Lab, in which students work with real clients (NGOs, municipalities, financial institutions) and manage the full multilingual workflow — from setting up the AI pipeline to the final human-led quality certification.

How Agencies Can Evolve

From Language Service Providers to Meaning Architects

The decade leading up to 2026 has proven that the industry does not belong to those who merely translate, but to those who transform. In a landscape where AI produces fluent text, audio, and video in seconds, competitive advantage shifts to the stewardship of meaning — the ability to combine technology, empathy, and ethics into measurable value.

The agency of 2026 no longer sells "words per page"; it sells clarity per context. The path leads away from linear production toward circular collaboration.

• New service lines include advanced human-in-the-loop post-editing under ISO 18587, supported by benchmarking through MetricX, XCOMET, and LLM judges; multimodal localization and AI dubbing that orchestrate voice-to-voice translation and video localization with attention to emotional resonance and lip synchronization; terminology as a knowledge asset, developed as living ontologies embedded in RAG systems; linguistic governance and AI auditing that provide ethical and technical QA for LLM outputs, including bias detection, hallucination checks, and compliance with the EU AI Act; and global accessibility and plain language services aligned with ISO 24495-1 and the EAA.

• New revenue models are emerging as well. Strategic language retainers replace ad hoc tasks with continuous maintenance of linguistic assets and LLM fine-tuning. Value-based or outcome pricing substitutes metrics of engagement, clarity, and impact for per-word rates. Compliance and risk packages bundle translation with GDPR, ESG, and data-integrity audits.

• New roles are taking shape: the editor-curator as architect of the "human signature" in machine-generated content; the linguistic data inquisitor who detects and mitigates bias in training datasets; the Language Operations (LangOps) specialist who bridges localization and IT; the real-time interpretation pilot who supervises live streams; and the knowledge ontologist who designs the multilingual frameworks that prevent AI from flattening a brand's voice.

Every pattern tells a story — every fold a translation of heritage into movement.

Measuring Meaning — Not Just Output

In the next era of translation, success will not be measured by how much we produce, but by how deeply we connect. A meaningful combination of metrics blends automated scoring (COMET, MetricX, XCOMET for baseline quality), terminology consistency (the share of approved terms across brand touchpoints), LQA error rate per 1,000 words, turnaround time reduction, and engagement metrics (clarity, empathy, and trust in communication). No single indicator is enough; we need multiple signals at once, as discussions within the WMT community confirm.

Translation quality must now include ethical fidelity — alignment with values as much as with vocabulary.

Ethics and Responsibility

As translation merges with AI, three duties remain distinctly human.

• The first is to protect privacy. Sensitive data must never be used to train public models without consent. Confidentiality is not a choice; it is the foundation of trust.

• The second is to credit authorship. Transparency about the balance between human and machine contribution is essential. Clients and readers have the right to know how meaning was made.

• The third is to preserve cultural integrity. Language is never neutral. Bias, erasure, or cultural flattening must be consciously named, measured, and corrected.

The ethics of translation is not about avoiding error — it is about choosing awareness.

10. The Future of Meaning

Translation will always be more than transfer — it is transformation. In an age of automation, its purpose expands: to ensure that what is understood remains human. The agencies, educators, and creators who embrace this calling will not compete with machines — they will teach machines what meaning means.

11. What You Can Take Away

If you are a translator, master the new literacy (MTPE under ISO 18587, metrics such as COMET/MetricX/XCOMET, human-in-the-loop workflows); specialize where machines cannot reach (terminology, transcreation, plain language); build a portfolio that demonstrates transformation (before-and-after examples, stylistic refinements, reach, clarity gains); and treat AI as an instrument, not a rival.

If you run an agency, move from production to curation; use ISO 17100 and 18587 as arguments for traceability and premium positioning; diversify your offer with retainers for terminology, accessibility, and quality-evaluation maintenance; and lead with responsibility. The future of translation is not repetition — it is stewardship.

12. Closing Reflection

    
    

Translation will always be an act of listening — across languages, systems, and generations. Its purpose is not to copy the world, but to keep it coherent. In an age of automation, those who translate with integrity will remain its custodians — guardians of the dialogue that keeps meaning alive.

13. References & Further Reading

    
    

Industry reports and market analyses (2025–2026)

• Nimdzi Insights, The 2026 Nimdzi 100 (https://www.nimdzi.com/nimdzi-100-2026/)

• CSA Research and ALC, global language services market report (cited via Statista, data of December 2024)

• Fortune Business Insights, Language Services Market Size, Share, Industry Report, 2034

• Mordor Intelligence, Language Services Market — Industry Analysis, Size & Trends Report, 2030

• Nimdzi Insights, Language Technology Market Size (2024–2025)

Standards and regulatory frameworks

• ISO 17100:2015, Translation services — Requirements for translation services

• ISO 18587:2017, Translation services — Post-editing of machine translation output — Requirements (revision under way at ISO TC 37/SC 5)

• ISO 24495-1:2023, Plain language — Part 1: Governing principles and guidelines

• ISO 30042:2019, Management of terminology resources — TermBase eXchange (TBX) (revision under way, ISO/AWI 30042)

• Regulation (EU) 2024/1689 (EU AI Act); timeline: prohibited practices since 2 February 2025, general-purpose AI obligations since 2 August 2025, most rules on high-risk systems from 2 August 2026

• Directive (EU) 2019/882 (European Accessibility Act), applicable since 28 June 2025

Technology and evaluation metrics

• Guerreiro et al. (2024), xCOMET: Transparent Machine Translation Evaluation through Fine-grained Error Detection, Transactions of the Association for Computational Linguistics, vol. 12, pp. 979–995

• Juraska et al. (2024), MetricX-24: The Google Submission to the WMT 2024 Metrics Shared Task

• Google Translate (2025), MetricX-25 and GemSpanEval: Google Translate Submissions to the WMT25 Evaluation Shared Task (arXiv:2510.24707)

• Proceedings of the WMT 2024 and 2025 conferences (ACL Anthology)

• Documentation of the open-source COMET framework (Unbabel)

Institutions and public sources

• Swiss Federal Statistical Office (FSO), Language, Religion and Culture Survey 2019; press release of 25 January 2021 and subsequent updates

• Federal Office of Culture (BAK/OFC), Switzerland: documentation on language policy

• European AI Office, Guidelines for Providers of General-Purpose AI Models (July 2025)

• European Disability Forum, analyses of the implementation of the European Accessibility Act

Author's note: All statistical and regulatory figures have been verified against sources available as of April 2026. The status of ISO standards, EU AI Act timelines, and market forecasts is continuously updated; before relying on them in legal or contractual contexts, verification with the original issuing body is recommended.