Technology as a Tool: Argentum Translations on AI and Machine Translation

Interview with Cynthia Penovi, Founder and CEO of Argentum Translations, on artificial intelligence, machine translation, ethics and the need for human intervention.

Cynthia Penovi sitting behind a laptop.

Q: The language industry has undergone major changes with the rise of machine translation and AI. How does Argentum Translations view these developments, and what is your relationship with these new technologies?

There is no question that the industry has changed significantly. Machine translation (MT) and large language models (LLMs) are now part of the broader conversation around language services. Our position is measured and professional. We recognize that these systems are tools. When used correctly, in the right context, and under strict human control, they can enhance processes and support quality.

It is also important to be honest about history. Early machine translation outputs were often low quality, inconsistent, and unreliable for professional use. However, systems have improved with time. Advances such as terminology injections, retrieval-augmented generation, prompt engineering, and structured workflows have made outputs more stable and controllable.

As a company, we decided not to rush into implementation. Our team has invested in formal training to learn more about these technologies; for instance, we have all recently concluded a Master in AI and Innovation for Localization. This allowed us to understand not only the capabilities of these systems, but also their limitations, risks, and ethical implications. We continue to deepen our knowledge before shaping internal processes and conducting controlled testing. Feeling more confident in the fact that we could support a secure, high-quality process that could benefit our clients, we have just now incorporated machine-generated first draft into our regular workflows for certain content types and clients.

Q: What is the difference between using Neural Machine Translation systems and Large Language Models, and which do you use?

This distinction is important. Neural Machine Translation systems, or NMT systems, are built specifically to translate text from one language to another. Their function is focused and task-specific. They are designed to reproduce the meaning of the source text as accurately and directly as possible in the target language. When configured with TMs, curated terminology injections and client-specific glossaries, they tend to produce consistent and predictable outputs.

Large Language Models, by contrast, are general-purpose systems trained to generate language across many tasks. They can translate, but they can also paraphrase, summarize, expand, and creatively reformulate content. Because they are generative by design, they may occasionally introduce wording or information that is not explicitly present in the source text, a phenomenon often described as “hallucination.” For highly precise professional translation, especially in technical or institutional contexts, that broader generative capacity requires stricter controls.

NMT systems are optimized specifically for bilingual sentence mapping. They are not designed to invent or expand content, but to transfer meaning from one language to another. That narrower objective makes them more appropriate for structured translation workflows. We currently do not use general-purpose LLMs for translation drafting.

As with every component of our workflow, no machine-generated draft is ever delivered without human professional review. Our linguists verify accuracy, terminology, tone, and completeness in multiple stages. By limiting our implementation to a secure, task-specific NMT system and combining it with rigorous human oversight, we prioritize reliability, predictability, and client safety.

Q: What ultimately led you to formally integrate a machine-generated first draft into your workflow?

After research, training, and internal pilots, we determined that a carefully controlled first-draft layer could strengthen our quality assurance process. This draft always goes through structured human post-editing, typically in two or three stages depending on project complexity and other factors.

The objective will never be replacement, but reinforcement. A structured draft can reduce omissions, unintended additions, and mechanical inconsistencies. It provides a consistent terminological base, especially when we inject carefully curated client-specific glossaries into the system, and help define tone and style by feeding it with our internal style guides.

Q: Some people worry that using machine translation means paying lower rates. What distinguishes your approach?

Argentum Translations was created by professional translators for professional translators. That philosophy has not changed. From the beginning, we have required that our linguists hold bachelor’s and master’s degrees in translation studies or related disciplines. Many also hold national certifications, including ATA credentials, and have specialized experience in technical, medical, academic, and institutional fields. Professional translation is a highly trained discipline, not a mechanical task.

The first-draft layer we have integrated is not a shortcut and does not reduce the intellectual effort required. It is an additional structured step within a rigorous workflow. Every project still undergoes thorough human post-editing, and depending on the scope and subject matter, two or three review stages are implemented to ensure terminological precision, syntactic accuracy, stylistic consistency, and contextual appropriateness. The draft simply provides a structured base that can reduce mechanical omissions or inconsistencies, allowing the linguist to focus more deeply on meaning, nuance, and subject-matter validation.

Most importantly, we do not reduce per-word rates because of this workflow. Unlike industry models that treat machine-assisted translation as discounted work, we consider post-editing at this level to be advanced professional work. Our translators are not doing less; they are applying higher-level analytical skills in higher volumes, and ensuring that the final text meets professional and academic standards.

Ethical compensation is one of our core values. Sustainable quality depends on qualified professionals being fairly paid for their expertise. Lowering rates simply because a structured draft exists would undermine the profession and, ultimately, the quality delivered to clients. By maintaining fair pay, we protect professional standards, ensure accountability, and reinforce the integrity that has always distinguished Argentum Translations.

Q: How do you address confidentiality concerns?

Confidentiality is never an afterthought in our process. Before any text is processed through a drafting layer, there is a careful content selection stage. Only appropriate materials, or clearly defined portions of texts that do not contain sensitive or personally identifiable information, are eligible. We also define strict content categories. General marketing materials or non-sensitive technical documentation may qualify for this workflow. Highly sensitive legal, medical, financial, governmental, or unpublished intellectual property content is either excluded or subject to additional authorization and controls.

Our professional providers offer strict contractual commitments regarding data protection. These agreements explicitly state that content is not retained, not logged for training purposes, and not reused in any way. They operate under strong data protection standards and implement technical and organizational safeguards designed to prevent unauthorized access.

In addition, we take several measures, such as using secure API connections and applying a structured redaction process when necessary. Any personally identifiable information and sensitive elements are replaced with placeholders prior to processing and securely reinserted afterward. This adds an additional protective layer even when working with secure systems.

Q: Thank you for your time!