Palabras Prohibidas

By Fernando Castillo

I have vague memories of how I used to translate articles and book chapters during my university years. But I am certain that in the early 90s and mid 2000s, translating a fragment of text was a true odyssey, ranging from consulting physical bilingual dictionaries to consulting a bilingual person.

Although many free online translation tools saw the light of day in the mid-to-late 1990s—well before the birth of Google Translate, which launched on April 28, 2006—their initial iterations relied on technology based on grammatical rules and dictionaries (provided mainly by SYSTRAN), yielding results that were often overly literal, rigid, and full of comical mistakes.

However, over the years, the translation process evolved toward neural network-based translation (AI) and expanded its catalog exponentially to support nearly 250 languages. We are talking about artificial intelligence capable of processing massive volumes of information in seconds with a syntactic fluency that, compared to its beginnings, is undeniably impressive.

Nevertheless, as a linguist and an active participant in the growing use of Artificial Intelligence in almost every aspect of daily life, several questions arise: Is current translation speed comparable to deep understanding? To what extent can the apparent grammatical perfection of an algorithm replace critical judgment, empathy, and the cultural sensitivity of a human being?

We have advanced by leaps and bounds from rule-based machine translation (RBMT) to neural machine translation (NMT), and more recently toward the integration of Large Language Models (LLM) and Adaptive Generative Translation (AGT). While traditional neural engines excel in sentence-level morphosyntactic precision, environments powered by generative artificial intelligence offer highly advanced capabilities in supra-sentential contextual understanding and document-wide tone adaptation.

Today, translation management platforms integrate these models to dynamically query translation memories and specialized glossaries within prompt guidelines, tailoring content to a brand’s tone of voice. However, using these models without proper supervision still results in variations in terminological accuracy, keeping the debate centered on how reliable synthetic drafts truly are.

The Human Filter is Always Indispensable

Faced with the need to manage large volumes of digital content under tight deadlines without compromising rigor, the global industry has consolidated a hybrid framework known as Machine Translation with Post-Editing (MTPE). In this workflow, an automated system generates an initial draft, and a professional linguist reviews, corrects, and optimizes the final output.

To regulate this practice and prevent time savings from leading to a decline in quality, the international standard ISO 18587:2017 establishes quality management criteria for post-editing, complementing standards such as ISO 9001 and ISO 17100. This standard defines two clear operational levels:

• Light post-editing: Focused on correcting severe semantic errors to guarantee a clear and accurate text, typically intended for internal documentation or rapid information analysis.
• Full post-editing: Required for commercial, legal, technical, or advertising texts, where a fluent, natural, and stylistically equivalent result to a professional human translation is demanded.

To focus human intervention on the most critical segments, organizations turn to the Multidimensional Quality Metrics (MQM) framework and Quality Estimation (QE) systems such as COMET-QE, which evaluate the quality of the generated text before routing it to a specialist.

It is here that we encounter a fundamental ethical and technical dilemma: the fluency paradox. Unlike the translation tools of a decade ago, which exhibited obvious syntax errors, modern language models generate nearly flawless text that nevertheless contains subtle hallucinations or plausibility errors.

An AI-generated translation may appear natural at first glance, yet hide terminological inaccuracies or distortions of meaning that demand a higher cognitive effort to review than fixing an imperfect grammatical structure.

For this reason, reference frameworks like the ISO 18587 standard and position papers from international organizations agree that post-editing should not be imposed indiscriminately: highly creative content, emotional marketing campaigns, or sensitive legal documents preferably require direct human translation from scratch.

Ethical commitment with a human touch

The consolidation of the hybrid paradigm demonstrates that technology must be conceived as a collaborative tool rather than a substitute for human beings. Entities like the International Federation of Translators (FIT) and the Spanish Association of Translators, Proofreaders, and Interpreters (ASETRAD) emphasize that linguists not only act as guarantors of final quality, but also as instructors of the system, feeding memories and models with their own professional criteria.

Furthermore, emerging artificial intelligence regulations in the European Union and data governance guidelines promote transparency and explicit traceability of AI-translated content, reserving ethical and legal accountability solely for the professional who reviews and certifies the final text.

Ultimately, translation has never been a mere mechanical exchange of words, but an act of interpretation and a bridge between cultures. No matter how sophisticated a generative algorithm may be, ethical responsibility, discernment, and critical sense will remain, irreducibly, the domain of the human mind.

---

Bachelor’s Degree in Linguistics from the University of Colima. Feel free to send questions or comments to lcastilloochoa@gmail.com

Sources Consulted:

• ISO 18587:2017: Translation services — Post-editing of machine translation output — Requirements.
• International Federation of Translators (FIT): FIT Position Paper on Machine Translation in the Age of AI.
• Lokalise & memoQ: Guides on AI Translation, Adaptive Generative Translation (AGT) and MTPE Workflows.
• Seprotec & Interpro Translation Solutions: Machine Translation Post-Editing and ISO Standards Implementation.